Comparative DIKWP Analysis of Medical Diagnostics between DeepSeek and ChatGPT
Yucong Duan
Benefactor: Shiming Gong
International Standardization Committee of Networked DIKWP for Artificial Intelligence Evaluation(DIKWP-SC)
World Artificial Consciousness CIC(WAC)
World Conference on Artificial Consciousness(WCAC)
(Email: duanyucong@hotmail.com)
1Preface
Artificial intelligence (AI) technology is increasingly penetrating into the medical field as an important auxiliary tool for clinical diagnosis and decision-making. In recent years, large language models (LLMs) have shown great potential in medical consultation and treatment recommendations. Among them, OpenAI's ChatGPT has become an industry benchmark with its excellent natural language processing capabilities and rich medical knowledge reserves, and China's independently developed DeepSeek model has also emerged, achieving reasoning performance comparable to leading models with lower training costs and open strategies. As representatives of the new generation of intelligent diagnosis and treatment assistants, these two models have attracted widespread attention in medical AI applications.
How to evaluate and compare the diagnostic and treatment cognitive capabilities of different AI models has become a key topic in artificial intelligence medical research. Traditional evaluations often focus on question-answering accuracy or single-point performance, but it is difficult to analyze the pros and cons of the model in each link of the cognitive process. To this end, we introduce the DIKWP*DIKWP theoretical model proposed by Professor Yucong Duan as an analysis framework. The model characterizes intelligent cognition from five levels: data, information, knowledge, wisdom, and intention, and decomposes the interactive transformation process between each level into 25 basic modules, using a mesh structure to fully reveal the mechanism of the intelligent agent from perception to decision-making. Through this framework, we can conduct a fine-grained analysis of the diagnosis and treatment process of the AI model to find out their respective advantages and disadvantages.
This study takes a 48-year-old male patient as an example (symptoms include night cough, yellow-green sputum, night sweats, itchy throat, and cephalosporin antibiotics are effective), and collects the diagnosis and treatment recommendations given by the DeepSeek model and the ChatGPT model for this case. It is assumed that the former provides a detailed step-by-step analysis, while the latter tends to diagnose acute bronchitis in the recovery period and gives evidence-based standard medication and observation plans. Based on the 25 cross-modules of the DIKWPDIKWP model, we will conduct a comparative analysis of the diagnosis and treatment cognitive process of the two models. The research content includes: (1) explaining the structure and hierarchical cognitive dimensions of the DIKWPDIKWP model; (2) comparing the performance of the two models in the 25 cognitive modules one by one, and visually displaying the advantages and disadvantages with the help of tables; (3) summarizing the advantages and disadvantages of the two models in terms of diagnosis and treatment cognitive ability; (4) proposing improvement suggestions for the problems existing in the DeepSeek model. Through the above in-depth analysis, we hope to provide new ideas for the evaluation and optimization of AI medical diagnosis models and promote the research and development and improvement of intelligent medical products.
2DIKWP*DIKWP model structure explanation and hierarchical dimension overview
The five-layer architecture of the DIKWP model: DIKWP is an expansion and networked reconstruction of the traditional DIKW (pyramid) model. Compared with the classic "data-information-knowledge-wisdom" hierarchical model, DIKWP introduces "Purpose/Intention" at the highest level to form a five-level cognitive system. The five levels are: Data, which represents the original objective facts and sensory signals; Information, which represents processed and organized data and understandable descriptions (answering "what happened"); Knowledge, which represents the universal laws, causal relationships and patterns abstracted from information (answering "what does this mean"); Wisdom, which represents the ability to make wise decisions and actions based on knowledge (answering "what is the best way to do it"); Purpose/Intention, which represents the purpose, goal or motivation of the behavior, is a high-level decision-making driver (answering "why do this"). By adding the intention layer to the classic cognitive hierarchy, the DIKWP model makes up for the shortcomings of the traditional DIKW model in terms of decision-making purpose, and emphasizes the important role of goal orientation in the behavior of intelligent agents.
Mesh interaction and 25 conversion modules: Different from linear stratification, the DIKWP model regards the above five levels as a dynamic interactive network structure. The layers are not connected by a single path, but there is a multi-directional and multi-order direct conversion relationship. In other words, in the DIKWP framework, any two levels can directly undergo a two-way conversion, including internal flow at the same level and cross-level up-down dimensional conversion, forming a total of 5×5=25 possible cognitive conversion modules. Each module represents a specific processing capability or information flow path in the intelligent cognitive process. For example, the "data→information" module corresponds to processing and organizing raw data into meaningful information, such as extracting characteristic indicators from the messy signals collected by sensors; the "information→knowledge" module corresponds to summarizing general knowledge from specific information, such as summarizing the diagnostic points of a disease from a large number of case descriptions; the "knowledge→wisdom" module corresponds to using existing knowledge for reasoning and decision-making, such as doctors making diagnosis and treatment decisions based on medical knowledge. For another example, the high-level to low-level "intention→data" module reflects the goal-oriented data collection behavior, and the intelligent body will select and obtain the required raw data based on the ultimate goal; the "wisdom→intention" module describes the feedback adjustment process of the decision output to the goal, such as reflecting on whether the preliminary diagnosis and treatment plan meets the patient's overall health goals after obtaining it. Through these 25 basic transformation modules, we can comprehensively and meticulously characterize the evolutionary relationship and cognitive path of the AI system from low-level signal processing to high-level decision-making intentions. Below we will analyze the performance differences between DeepSeek and ChatGPT in processing the aforementioned patient cases for each of these 25 modules.
3Comparative analysis of 25 modules
Below we compare the 25 cognitive transformation modules in groups according to the five levels of the DIKWP model. For each module, we first explain the meaning of the module in the medical diagnosis and treatment context, then analyze the corresponding capabilities of the DeepSeek and ChatGPT models when handling cases, and summarize the advantages and disadvantages of the two models through a table.
3.1Transformation module starting from the data layer
3.1.1Module 1: Data → Data
Module meaning: The "data→data" module refers to the direct processing and organization of raw data, that is, cleaning, filtering, aggregation and other operations on the same data level. In medical diagnosis, this is reflected in the model's ability to read the patient's original medical information and perform pre-processing, such as identifying key data points in symptoms, consistency checks, and processing of missing or abnormal data.
Comparative analysis: For a given case of a 48-year-old male patient, both models need to first receive and understand the original symptom and sign data, including "night cough, yellow-green sputum, night sweats, throat itching, and effective treatment with cephalosporin". [Analysis] The DeepSeek model shows a comprehensive extraction and enumeration of the key points of the original condition in the "data→data" link. Its detailed diagnosis and treatment recommendations often list the patient's symptoms and previous treatment responses one by one at the beginning, showing a strong ability to summarize data. For example, DeepSeek may explicitly mention the time of the patient's cough (mainly at night), sputum characteristics (yellow sticky purulent sputum), accompanying symptoms (night sweats, throat itching), and the effectiveness of antibiotics, ensuring that key original data is obtained without omission. This comprehensive capture of data contributes to the integrity of subsequent diagnostic reasoning. However, if the information provided by the patient is uncertain or vague, DeepSeek may tend to assume a wider range of values to ensure the coverage of the original data, but it may also introduce some unnecessary data points.
When reading the original case data, the ChatGPT model is concise and highlights the key points. It usually simply repeats the patient's symptoms into a one- or two-sentence summary, highlighting the elements that are closely related to the diagnosis. For example, ChatGPT may summarize the original data as "middle-aged male, cough with purulent sputum and night sweats, symptoms improved after treatment with cephalosporin antibiotics." This refinement reflects ChatGPT's ability to screen and compress important data, which helps to quickly focus on the main contradictions. But refinement also means that ChatGPT may ignore some marginal information. For example, ChatGPT may not particularly emphasize the symptom of "itchy throat" when summarizing, thinking it is relatively minor. DeepSeek, due to its detailed listing, rarely misses any detailed data.
In general, DeepSeek is characterized by detailed data acquisition and restatement in the "data→data" module, ensuring the completeness of the original information; ChatGPT is characterized by efficient data extraction, highlighting the key points but possibly omitting details. The two complement each other: DeepSeek's data integrity reduces the risk of missing important clues, and ChatGPT's data screening improves the efficiency and pertinence of analysis. The following table summarizes the advantages and disadvantages of the two in terms of data processing:
Data → Data module performance |
DeepSeek |
ChatGPT |
Original disease extraction |
Comprehensive list of all symptoms and reactions of the patient, with complete details |
Briefly summarize the main symptoms and positive signs, highlighting the key points |
Data preprocessing |
Does not omit information but may include redundant details |
Focuses on key information but may ignore minor points |
Pros and cons |
Comprehensive information to ensure completeness |
Efficient extraction and improved focus |
3.1.2Module 2: Data → Information
Meaning of the module: The "Data → Information" module refers to extracting meaningful information patterns from raw data and converting raw data into structured and understandable information. In the diagnosis and treatment scenario, this means identifying clinically meaningful preliminary conclusions or clues based on patient symptoms and signs. For example, "night cough and yellow-green sputum" can be summarized as "signs of respiratory infection", "night sweats" information can be associated with infectious diseases, or "cephalosporin treatment is effective" can be used to infer that the pathogen is sensitive to the antibiotic. This step is equivalent to the doctor's initial impression after reviewing the medical records, integrating scattered symptom data into an information summary.
Comparative analysis: In this case, both models process the collected symptom data into preliminary information judgments on the disease. DeepSeek's output is often reflected in further explanation and classification of symptoms. For example, it may point out: "The patient's night cough with yellow-green purulent sputum indicates a purulent infection in the lower respiratory tract; night sweats may mean a systemic reaction to an infectious disease; throat itching indicates pharyngeal inflammation or allergic factors; the effectiveness of cephalosporin antibiotics indicates that the pathogen is sensitive to β-lactam drugs, supporting the diagnosis of bacterial infection." Through this analysis, DeepSeek translates the original symptoms into medical information: there are clear signs of bacterial infection in the lower respiratory tract. It can be seen that DeepSeek shows a style of explanation and deduction in the "data→information" stage: each piece of data is mapped to the corresponding clinical significance. For example, late-night coughing is associated with chronic inflammation, sputum color is associated with neutrophil activity, and antibiotic response is associated with bacterial sensitivity. This detailed information extraction makes its analysis process transparent and organized. However, because DeepSeek tends to exhaust all possible information meanings, it may result in a large amount of information and length, requiring readers to identify the main contradictions.
In this module, ChatGPT usually integrates patient information more condensedly to quickly form an overall understanding of the case. It may combine multiple symptoms into a clinical information block, for example: "The patient presents with persistent cough with purulent sputum (indicating bacterial infection), obvious symptoms at night and night sweats (the infection may be severe or prolonged), and the symptoms have improved significantly after recent cephalosporin antibiotic treatment (indicating that the infection is sensitive to antibiotics)." ChatGPT sublimates data into judgment information on the nature of the disease through this comprehensive description: for example, "antibiotic-sensitive respiratory tract infection". It focuses on refining the main contradictions and integrating related data points into clinically meaningful judgments. For example, cough + purulent sputum → signs of bronchitis/pneumonia; night sweats → long course of infection or tuberculosis to be discharged; antibiotics are effective → high probability of common bacterial infection. This information integration capability enables ChatGPT to quickly get to the point of diagnosis. However, compared with DeepSeek, ChatGPT rarely explains the meaning of each symptom one by one when refining information, but directly gives conclusive information. This is concise and clear for general readers, but it also reduces the visibility of the reasoning process.
In summary, DeepSeek in the "Data→Information" module is a detailed interpretation of the meaning of symptoms, giving clinical significance to each data, and the information output is rich but cumbersome; ChatGPT is an efficient integration of symptom information, directly giving a comprehensive judgment, with concise information but implicit reasoning details. The advantages and disadvantages of the two are shown in the table below:
Data → Information module performance |
DeepSeek |
ChatGPT |
Symptoms and meanings |
Explain the clinical significance of each symptom one by one, and the reasoning process is transparent |
Comprehensive symptoms form a preliminary judgment, reducing the need for item-by-item explanations |
Information Integration |
A large amount of information, covering a variety of possible interpretations |
Highly integrated information directly points out the main clinical clues |
Pros and cons |
Thorough but possibly lengthy |
Concise but with less reasoning details |
3.1.3Module 3: Data → Knowledge
Meaning of the module: The "data → knowledge" module refers to the direct extraction or association of general knowledge or empirical rules from the raw data. This is usually done by jumping directly from the data to the knowledge layer based on experience without complete information and step-by-step reasoning. In medical treatment, this is equivalent to directly associating a patient's data with a specific disease or pathological mechanism based on intuition or pattern recognition. For example, an experienced clinician may immediately associate tuberculosis with data such as "night sweats, chronic cough, and yellow-green sputum", or directly infer the pathogen category based on "cephalosporin is effective". This is a pattern-matching type of rapid knowledge retrieval.
Comparative analysis: ChatGPT is often good at using the massive knowledge accumulated during its training process to achieve a rapid jump from data to knowledge. For this case, ChatGPT may associate several medical knowledge patterns at the moment of obtaining symptom data: for example, "night cough + night sweats" is a classic symptom combination of tuberculosis; "yellow-green sputum + antibiotics are effective" is a typical manifestation of bacterial lung infection (such as bacterial bronchitis or pneumonia). This association enables ChatGPT to quickly call the knowledge base of related diseases without having to rely entirely on the step-by-step reasoning process. In the output, ChatGPT may mention similar knowledge points: "The patient's symptom combination suggests bacterial lower respiratory tract infection, and the differential diagnosis includes chronic infectious diseases such as tuberculosis, but the latter is usually ineffective against cephalosporins." This shows that ChatGPT has used the ability of data to directly call medical knowledge for judgment. The advantages are fast speed and based on big data experience, it can directly give possible diagnostic categories. However, if the pattern in the training data is inaccurate or biased, it may also cause ChatGPT to lean towards a certain knowledge conclusion too quickly and ignore details.
In contrast, DeepSeek tends to reason more cautiously step by step rather than immediately jumping to conclusions at the knowledge level. DeepSeek may have fewer direct jumps in the "data → knowledge" module, and it is more likely to go through the "data → information" and "information → knowledge" processes before reaching conclusions. But this does not mean that DeepSeek lacks the ability to directly obtain knowledge from data: in certain clear patterns, it will also directly apply medical knowledge to interpret data. For example, DeepSeek may be alert to tuberculosis as one of the differential diagnoses when analyzing night sweats and chronic coughs, because its knowledge base also contains intuitive rules such as "night sweats + chronic cough → consider tuberculosis". However, DeepSeek usually presents counter-evidence or data that needs to be verified when stating such suspicions, rather than immediately confirming them. For example, it may say: "Night sweats make us think of the possibility of tuberculosis, but we need to consider the improvement after antibiotic treatment, which is inconsistent with tuberculosis and needs further examination and confirmation." This reflects the characteristic of DeepSeek remaining cautious after directly mobilizing knowledge: even if it makes a jump from data to knowledge, it will return to the data layer for verification, or advance to a higher level to seek support.
Therefore, in the "data→knowledge" module, ChatGPT has the advantage of rapid pattern recognition and can immediately associate relevant disease knowledge. It is efficient but may ignore verification; DeepSeek is more stable and prudent, rarely jumping to conclusions without processing, and even if it directly associates with knowledge points, it will indicate that it still needs to be verified. The comparison between the two is as follows:
Data → Knowledge Module Performance |
DeepSeek |
ChatGPT |
Pattern Association |
They rarely jump to conclusions, and will still maintain a verifying attitude when associating knowledge with them. |
Good at pattern recognition and quick association of typical disease knowledge |
Knowledge Calling |
Call knowledge carefully, with reasoning transitions and conditions |
Directly access a rich medical knowledge base to immediately point to possible diagnoses |
Pros and cons |
Robust and reliable, not easily blinded by one thing |
Fast and efficient, but avoid over-generalization |
3.1.4Module 4: Data → Wisdom
Module meaning: The "Data → Wisdom" module means that decisions or action plans are directly generated from raw data, that is, high-level decisions are directly formed without explicit processing of intermediate information or knowledge levels. In clinical practice, this is equivalent to relying on instinct or intuition to make diagnosis and treatment decisions based only on preliminary patient data. For example, an extremely experienced expert directly decides on a treatment plan the moment he sees a typical combination of symptoms, without necessarily explaining the reasoning process one by one. This ability is often based on subconscious pattern recognition of massive training and experience.
Comparative analysis: In AI models, ChatGPT sometimes exhibits behaviors similar to "data → wisdom", which is attributed to the fact that its deep learning model may have implicitly included many reasoning processes through end-to-end training. For this case, ChatGPT may quickly give a relatively complete diagnosis and treatment recommendation: for example, it directly recommends "follow the treatment of acute bronchitis recovery period, continue to complete the antibiotic course, cooperate with cough and expectorant drugs, and closely observe the changes in the condition." This suggestion seems to be a one-step process from symptoms to solutions. Although it contains knowledge and reasoning behind it, ChatGPT's answer may not be unfolded layer by layer. This is reflected in ChatGPT's internalization of a large amount of training data into a direct mapping - input symptom patterns and output standard decisions. This "black box intuition" makes it respond quickly, and its decisions often conform to conventional medical treatment norms. However, direct decision-making also has risks: if the input data is ambiguous or rare, the lack of intermediate reasoning steps may lead to the neglect of special considerations. Fortunately, this case is more common, and ChatGPT's direct decision is basically reasonable.
DeepSeek is relatively less inclined to skip intermediate steps and give conclusions directly. It is more of a white-box step-by-step decision. Therefore, DeepSeek's direct performance in the "data→wisdom" module is weaker. It almost never makes a final diagnosis and treatment decision without explanation, but tends to analyze first and then make a decision. This does not mean that DeepSeek is slow, but that it usually explicitly outputs the reasoning chain, rather than implicitly embedding the reasoning process inside the model like ChatGPT. Therefore, DeepSeek's decision-making process is more transparent to users, and it lists the basis and considerations before giving suggestions. For example, DeepSeek may first discuss possible diagnostic options and the evidence supporting each, and then make decision recommendations based on these discussions. Although this means that DeepSeek does not show "instant intuitive decision-making", such a process helps to verify the correctness of the decision and reduce the possibility of rash conclusions.
In summary, ChatGPT demonstrates the ability to make quick and intuitive decisions in the "data → wisdom" module, and directly provides mature solutions in routine cases; DeepSeek is more inclined to make decisions after explicit reasoning, and rarely gives answers directly without explanation. This reflects the difference in style between the two models:
Data → Smart Module Performance |
DeepSeek |
ChatGPT |
Direct decision making |
Rarely jump to conclusions without analysis, usually have a reasoning process |
Often gives complete decision solutions quickly, with the reasoning process implicit |
Decision-making transparency |
The process is transparent and the basis is explained before making decisions |
The process is implicit and the decision results are directly presented |
Pros and cons |
Be cautious and reliable |
Quick and intuitive way to efficiently handle common situations |
3.1.5Module 5: Data → Intent
Module meaning: The "data→intention" module indicates that the low-level data directly triggers the high-level intention or goal adjustment. In other words, some raw data input will directly change or determine the purpose of the agent. In the medical scenario, this is equivalent to some data of the patient allowing the doctor to immediately establish a new goal or alert point. For example, when the doctor hears the symptom of "night sweats", he may instantly have the alert intention of "don't miss tuberculosis"; or when he sees that the patient responds well to antibiotics, he may have the treatment goal of "controlling the infection as soon as possible and avoiding recurrence".
Comparative analysis: In this case, there are some striking data points that may directly trigger the diagnosis and treatment intention. DeepSeek is highly sensitive to the "data → intention" conversion, especially to abnormal or crisis-indicating signals, and will respond quickly at the goal level. For example, "night sweats" are not typical in general respiratory infections. DeepSeek may immediately regard it as a red flag data that needs special attention, and then generate a clear intention: "rule out major diseases such as tuberculosis." Therefore, in DeepSeek's suggestions, this intention may appear very early - for example, it is recommended to perform sputum smear/culture or tuberculosis examination. This demonstrates DeepSeek's ability to trigger high-level goals from raw data: a symptom triggers the addition of a goal in its decision-making process (i.e., differential diagnosis of major diseases). Similarly, the data of "cephalosporin treatment is effective" may give DeepSeek another intention: "confirm that the bacterial infection has been controlled", thereby emphasizing the purpose of consolidating treatment and preventing recurrence in the plan. These are all examples of raw data rising directly to the intention level. DeepSeek reflects this in its output by explicitly listing the next step, for example it might write: "Because the patient has night sweats, tuberculosis should be considered, and the goal is to rule it out through examination."
ChatGPT is also driven by key data, but its "data → intention" response is often more implicit and not directly stated. ChatGPT tends to integrate intention into subsequent reasoning rather than explicitly announcing the purpose in the answer. For example, when it hears "night sweats", it will also internally alert itself to serious infection, but the advice it gives may be reflected in a sentence: "If the symptoms are not completely relieved or persistent night sweats occur, further examination for tuberculosis and other possibilities should be considered." The intention here is actually implicit - to ensure that tuberculosis is not missed, but ChatGPT does not directly say "the goal is to rule out tuberculosis", but incorporates the advice in the form of a hypothetical scenario. This is in contrast to DeepSeek: DeepSeek is more likely to explicitly point out a target item. For "cephalosporin is effective", ChatGPT will also internally lead to the intention of "completing the course of treatment to cure the infection", and it is manifested as a suggestion that the patient continue to take the drug in a standardized manner until the end of the course of treatment. It can be seen that ChatGPT internalizes the data-triggered goal into a conditional suggestion, while DeepSeek manifests it as a clear goal statement.
Therefore, in the "data→intention" module, DeepSeek is characterized by data-driven clarification of purpose, which directly incorporates the goals derived from the data into the diagnosis and treatment plan; ChatGPT is data-driven implicit purpose, which incorporates the goal into the wording of the suggestion without directly highlighting the name of the goal. The comparison between the two is as follows:
Data → Intent Module Performance |
DeepSeek |
ChatGPT |
Targets triggered by symptoms |
Directly determine new goals (such as excluding a disease) based on abnormal data and clearly propose |
Focus on key data internally, and goals are reflected in subsequent recommendations (implicit expression) |
Goal Statement |
Clearly list the purpose/task of diagnosis and treatment (examination or treatment goals) |
Purpose is usually not stated directly but expressed in the form of conditions or suggestions |
Pros and cons |
Clear goals and well-organized |
The tone is implicit, avoiding arbitrary statements but weakening the sense of purpose |
3.2Transformation module starting from the information layer
3.2.1Module 6: Information → Data
Meaning of the module: The "Information → Data" module means that the content of the information layer drives the processing or re-acquisition of the data layer. That is, based on the information that has been sorted out, it in turn guides further data collection, verification or reorganization. In clinical practice, this is reflected in doctors asking for more specific data or re-examining existing data based on preliminary diagnostic information. For example, after the doctor forms the information of "possible tuberculosis", he will go back and ask the patient in detail whether he has lost weight, contact history and other original data, or re-examine previous test results. This is a process of returning from summary to details to ensure that the information is correct or to supplement information gaps.
Comparative analysis: DeepSeek's diagnosis and treatment recommendations often include specific plans for further data collection, which can be regarded as the application of the "information → data" module. When DeepSeek sorts out some key information (such as "signs of bacterial infection" and "tuberculosis cannot be ruled out") through the previous modules, it will actively suggest obtaining more raw data to verify this information. For example, if DeepSeek concludes that "tuberculosis is to be ruled out" at the information layer, it may recommend sputum examination, chest X-ray or CT to obtain conclusive data; if the information layer indicates that "antibiotics are effective but not yet cured", it may recommend rechecking blood routine or inflammatory indicators to confirm infection control. This behavior reflects DeepSeek's ability to supplement data from information, and its output often uses such wording: "It is recommended to conduct further...examination" or "It is necessary to collect...data to clarify..." For example: "In view of information such as night sweats, it is recommended to obtain three samples of morning sputum for acid-fast bacilli smear examination to clarify whether there is tuberculosis infection." This move is DeepSeek's transformation of existing information into a driving force for collecting new data to ensure the reliability and integrity of diagnostic information.
ChatGPT also has the ability to feed information back to data, but usually in a question-and-answer format, it does not list the data to be collected in detail like DeepSeek does (unless explicitly asked). ChatGPT prefers to give treatment recommendations with data collection suggestions as observation conditions. For example, it may say: "Currently treat it as a convalescent bronchitis, but if cough and night sweats persist, imaging and sputum examinations should be reviewed to rule out other diseases." Here, ChatGPT does not actively arrange for an immediate examination in its main plan, but instead treats data collection as a conditional follow-up step. This shows that ChatGPT determines the trigger of "information → data" according to the situation, reflecting a data feedback mechanism guided by clinical changes. ChatGPT has already considered the main diagnosis to be convalescent acute bronchitis at the information level, so it does not immediately require a large amount of additional data. Only when the information (condition) changes or does not improve as expected, will it return to the data level to look for new evidence. Compared with DeepSeek's direct listing of examinations, ChatGPT is more conservative and avoids unnecessary examination burden.
In general, DeepSeek is a proactive data supplementation strategy in the "Information → Data" module. As long as there is uncertainty in the information, it will clearly propose the required data; ChatGPT is a conditionally triggered data feedback, which mainly recommends additional data collection when necessary.
Information → Data module performance |
DeepSeek |
ChatGPT |
Further data collection |
Proactively propose detailed inspection and data collection plans to verify information |
Provide additional advice as needed, and conduct further examinations if symptoms do not improve |
Timing of inspection/questioning |
Arrange data supplement as soon as the information is formed to prevent missed diagnosis in advance |
Observe first, and only return to the data layer for verification when necessary |
Pros and cons |
Be proactive and comprehensive, without missing out on potential problems |
Avoiding excessive inspections is more in line with the principle of evidence-based economy |
3.2.2Module 7: Information → Information
Module meaning: The "Information→Information" module refers to organizing, associating and reasoning information within the information layer to form new information. This usually involves merging and comparing multiple pieces of information, or deriving implicit information from existing information. In medical diagnosis, this is manifested as comprehensive analysis and correlation verification of different symptoms and test results information to obtain more reliable conclusions or discover contradictions. For example, the doctor will consider the information that "antibiotics are effective" and the information that "night sweats persist" together. If the two are contradictory, attention should be paid; or the length of the disease course and the severity of the symptoms can be combined to form a judgment on the nature of the disease (acute/chronic).
Comparative analysis: ChatGPT is very good at comprehensive reasoning at the information level. When dealing with this case, it considers various symptoms and reaction information together to extract a story line with internal consistency. For example, ChatGPT may synthesize such an analysis information: "The patient's symptoms began with a recent acute attack (indicating acute infection), and improved significantly after antibiotic treatment (consistent with bacterial infection), with only mild cough and night sweats at night (possibly symptoms of recovery or signs of chronic infection)." By associating information such as the onset of acuteness, treatment response, and residual symptoms, ChatGPT forms a comprehensive information judgment on the stage and nature of the disease - tending to believe that this is an acute infection that is getting better, and night sweats may be just part of the infection process rather than another disease. This process reflects ChatGPT's ability to balance multi-source information. It will self-consistently explain the relationship between each piece of information in the answer so that the conclusion is consistent. [Analysis] If there is contradictory information, ChatGPT also tends to explain it at the information level, for example: "Night sweats are commonly seen in tuberculosis, but other manifestations of this patient are more supportive of bacterial bronchitis, and the treatment is effective, so night sweats may be a general symptom caused by infection rather than a specific manifestation of tuberculosis." In this way, ChatGPT eliminates potential contradictions through the logical association of information to make the diagnostic ideas clear and consistent.
DeepSeek also synthesizes information in the "Information → Information" module, but due to its detailed output style, it may be presented as listing and comparing multiple pieces of information without immediately integrating them into a single conclusion. DeepSeek may state the relevant information separately and then analyze their weights one by one. For example, it may write: "The effectiveness of antibiotics indicates a tendency to bacterial infection, while night sweats indicate the possibility of chronic infection. These two pieces of information seem contradictory. We need to consider whether it is a single disease process (such as pneumonia with a strong inflammatory response causing night sweats) or a dual cause (such as co-existing tuberculosis). Based on the current information, it is more inclined to the former because the symptoms are significantly improved with treatment." It can be seen that DeepSeek presents different information side by side, then conducts comparative analysis, and finally reaches a consensus at the information level. This process is more cumbersome than ChatGPT, but it is more transparent: readers can clearly see how DeepSeek handles information conflicts and how to make trade-offs. Therefore, DeepSeek provides multi-angle analysis information at the information level, and readers need to extract the final points from it. ChatGPT directly presents the condensed key information, which is easier for readers to accept.
In short, ChatGPT is reflected in the "information → information" module as a highly condensed consistency analysis, which quickly integrates information to reach an internally consistent judgment; DeepSeek is reflected in a detailed multi-information cross-validation, and summarizes the conclusion after comparing each item. The comparison is as follows:
Information → Information module performance |
DeepSeek |
ChatGPT |
Information synthesis |
List each piece of information and analyze it one by one, the process is detailed and transparent |
Rapidly integrate multi-source information and directly give a comprehensive judgment |
Conflict resolution |
Clearly point out conflicting information and discuss and explain them |
Automatically digest contradictions in the conclusion and give consistent views |
Pros and cons |
The analysis is clear but requires readers to refine it |
The conclusion is clear but the reasoning is hidden |
3.2.3Module 8: Information → Knowledge
Module meaning: The "Information → Knowledge" module means to derive a higher level of general knowledge or rules from comprehensive information. In medical treatment, this usually means summarizing general medical conclusions or explaining information by citing existing medical knowledge based on the patient's specific information. For example, inferring the pathogenesis (bacterial infection causes bronchitis) or citing guideline knowledge (coughing for more than a few weeks is defined as chronic cough, etc.) from the patient's symptoms and examination information. This is the process of raising individual case information to the level of medical knowledge.
Comparative analysis: When DeepSeek provides advice for cases, it often connects individual cases with medical knowledge to enrich its analysis with knowledge-level content. For example, after DeepSeek analyzes the symptom information, it may rise to an abstract level and say: "The above manifestations are consistent with the clinical characteristics of acute bronchitis. According to medical knowledge, acute bronchitis is usually caused by common pathogens, and the course of the disease improves within a few weeks [inference], while night sweats are not a typical symptom, so it is more inclined to be a manifestation of a general infection process." In this analysis, DeepSeek has mapped specific information (symptom improvement, night sweats, etc.) to the knowledge of disease definitions and characteristics. This is equivalent to telling readers whether the patient's information matches or does not match a certain pattern in the known disease knowledge base. DeepSeek may also quote knowledge such as treatment principles: "The guidelines recommend a 7-10-day course of antibiotics for acute bacterial bronchitis. In this case, cephalosporin treatment is effective and should be completed according to the course [medical knowledge]." Through these, it rises to the knowledge level. Therefore, DeepSeek's answers often contain medical popular science content, which not only reflects professional knowledge, but also verifies the derivation of information to knowledge. The advantage is that the suggestions are well-founded and in line with the norms; the disadvantage is that more universal arguments may be incorporated, making the answers longer, which is valuable to professionals but may be a bit lengthy for ordinary patients.
ChatGPT also makes the "information → knowledge" conversion, but it is usually more concise and targeted. It tends to only quote the key points of knowledge related to the current decision without expanding too much background. For example, ChatGPT may say: "Given the improvement of symptoms, we know that bacterial infections are usually cured with an appropriate course of antibiotics [knowledge], so it is recommended to continue the current plan." Or: "Although night sweats are reminiscent of tuberculosis, tuberculosis is often poorly treated and has a longer course [knowledge], so acute bronchitis is more supported at present." As you can see, ChatGPT quotes knowledge points to support its conclusion or exclude an option, and immediately returns to the case after quoting. It will not explain what bronchitis or tuberculosis is in a long paragraph, assuming that the reader has basic common sense or only needs to understand the part related to this case. This makes the knowledge layer of ChatGPT concise and clear in purpose. For professional readers, this brief quotation is enough to explain the problem; for non-professionals, it may lack some background explanation, but it is still clear overall.
In summary, DeepSeek tends to be detailed in the "Information → Knowledge" module, often explaining medical knowledge or citing guidelines to support information analysis; ChatGPT tends to be concise, only extracting knowledge points related to decision-making. The comparison is as follows:
Information → Knowledge Module Performance |
DeepSeek |
ChatGPT |
Knowledge reference |
Extensive references to medical knowledge and guidelines to provide contextual explanations |
Select key knowledge points to provide a basis for current decision-making |
Abstract Derivation |
Elevate individual cases to general principles for explanation |
Use short knowledge points to verify or refute diagnostic ideas |
Pros and cons |
Rich in content, highly professional but can be lengthy |
Highly targeted, concise but less popular science |
3.2.4Module 9: Information → Wisdom
Module meaning: The "Information → Wisdom" module means using comprehensive information to directly guide decision-making. Here, the model skips the detailed deduction of intermediate knowledge based on the sorted patient information and makes judgments and decisions directly at the wisdom level. In clinical practice, this is similar to doctors directly formulating diagnosis and treatment strategies based on their understanding of the key points of the case (there may be knowledge support behind it, but the current decision is mainly based on the information at hand).
Comparative analysis: After obtaining key information about the patient, ChatGPT often directly gives a specific diagnosis and treatment plan, which is a typical case of information being directly used for intelligent decision-making. For example, in this case, after ChatGPT obtains and integrates information such as symptoms and treatment responses, it directly proposes a "treatment plan and management plan": such as recommending the completion of the antibiotic course, supplemented with expectorants, and paying attention to rest and reexamination. These are all content at the intelligence level, that is, the actual action plan. In ChatGPT's answer, this step is usually expressed as a clear list or paragraph of suggestions. When giving advice, ChatGPT has comprehensively considered previous information (for example, the information that the condition has improved leads to the recommendation of conservative treatment observation and no radical intervention). It will not go back and explain the basis for each suggestion again (that belongs to the knowledge level or information level), but present these decisions to users as logical results. This reflects ChatGPT's efficient jump from information to decision-making: when the information is clear enough, it directly enters the intelligence level to give a solution.
DeepSeek also makes decisions based on comprehensive information, but because it tends to explain the reasoning process, it may discuss or list several alternatives before giving the final solution, and then determine the recommendation. This means that DeepSeek's information to decision is not completed instantly, but is confirmed through an explicit reasoning step. For example, DeepSeek may write: "Based on the above information, we preliminarily diagnosed it as the recovery period of acute bronchitis, and the intelligent layer decision will adopt a conservative treatment strategy. At the same time, considering that the night sweats have not been fully explained, the intelligent decision also includes arranging follow-up and special examinations when necessary." It can be seen that DeepSeek almost always clearly states "Based on the above information, we decided..." before entering the intelligent layer. This statement clearly connects information with decision-making. Then DeepSeek will list the details of the decision, including specific drug selection, dosage, follow-up time, etc. Compared with ChatGPT, DeepSeek is more comprehensive in the output of the intelligent layer: it not only includes the main treatment, but may also involve patient education, prognosis assessment, etc. This comprehensiveness is of course also because its previous information analysis is more comprehensive, which can derive more management details. But at the same time, DeepSeek's intelligent decision-making part may seem a bit long-winded, giving overly complex management plans for simple cases.
In summary, ChatGPT is characterized by fast decision output in the "Information → Wisdom" module, and directly gives clear solutions based on known information; DeepSeek is characterized by comprehensive pre-decision explanations and decision content, and is slightly lengthy when outputting solutions. The comparison is as follows:
Information → Smart Module Performance |
DeepSeek |
ChatGPT |
Decision Making |
Summarize the information first and then give the plan, which covers a wide range of aspects (treatment, follow-up, etc.) |
Directly give the main plan based on the information, concise and focused |
Connection method |
It clearly states "Based on the above information, it is decided that...", and the reasoning transition is clear |
Decisions appear directly, assuming the reader understands how they relate to the information |
Pros and cons |
Decisions are well-reasoned and detailed, but can be cumbersome |
Decisions are made quickly and clearly, with clear focus but perhaps less detail |
3.2.5Module 10: Information → Intent
Module meaning: The "Information → Intent" module refers to adjusting or clarifying the intent based on the sorted information. In the diagnosis and treatment process, this means that the doctor calibrates his or her goals or focus based on his or her understanding of the case information. For example, information confirming that it is a common infection changes the doctor's intent to "cure this infection" and "avoid complications"; if the information is suspicious, the intention may change to "find out the unsolved mystery."
Comparative analysis: DeepSeek pays great attention to clarifying the next intention after analyzing the information. In this case, after information synthesis, DeepSeek has a high probability of confirming the main diagnostic direction (acute bronchitis), while leaving secondary possibilities (such as tuberculosis to be excluded). Therefore, DeepSeek's intention at this time will have two levels: one is the treatment intention - such as "clear the infection as soon as possible and promote the patient's recovery", and the other is the diagnosis intention - "further exclude possibilities such as tuberculosis to ensure accurate diagnosis". DeepSeek may directly write these intention-oriented contents in the suggestion section. For example: "Our goal is to control the infection while not missing any potential serious diseases." Such a statement clearly tells the user the goal after information analysis. This reflects the transparency of DeepSeek's intention layer: it makes people clearly know what the starting point of its decision is. DeepSeek may even give suggestions by intention classification, such as discussing treatment measures first (corresponding to the intention of curing infection), and then discussing follow-up examinations (corresponding to the intention of excluding other diseases). Therefore, readers can clearly feel the task priority of DeepSeek: the main task is to cure the disease, and the secondary task is to verify the diagnosis.
ChatGPT rarely directly states "My goal is..." in its answers, but this does not mean that it does not set goals, but that it incorporates intentions into specific suggestions. The information that ChatGPT understands will be subconsciously transformed into intentions and reflected as the focus of the suggestions. For example, ChatGPT pays attention to the information that the patient has improved, so the implicit intention is "consolidate the therapeutic effect and avoid over-medicalization"; at the same time, it knows that night sweats are not fully explained, so another implicit intention is "pay attention to abnormal conditions." These intentions will not appear explicitly, but will be permeated into the suggestions: such as advising patients to complete the course of treatment and get enough rest (reflecting the intention to cure the infection), and reminding them to return for a follow-up visit if abnormal symptoms occur (reflecting the intention not to miss serious diseases). ChatGPT expresses its intentions in a tactful way rather than a direct declaration. This style seems natural for patient communication, but for research analysis, we need to interpret its suggestions to grasp the intentions.
In general, DeepSeek directly exposes the intention in the "Information → Intent" module, clearly telling the user what they want to achieve next; ChatGPT implies the intention in suggestions, and focuses on goal orientation through suggestions. The comparison between the two is as follows:
Information → Intent Module Performance |
DeepSeek |
ChatGPT |
Target Adjustment |
After analyzing the information, clearly define the diagnosis and treatment goals (treatment and screening, etc.) |
Set priorities mentally based on information without stating them directly |
Expression |
Clearly describe the goals/intentions being pursued in the proposal |
Suggest a focus through suggested details and tone |
Pros and cons |
Clear intentions make it easier to understand decision motivations |
The tone should be natural and avoid being stiff, but the reader should understand the intention. |
3.3Conversion module starting from the knowledge layer
3.3.1Module 11: Knowledge → Data
Module meaning: The "Knowledge→Data" module refers to using the content of the knowledge layer to guide the processing of the data layer, including re-examining existing data or obtaining new data. In clinical practice, this means that doctors use medical knowledge to discover possible problems or gaps in the original data in the medical records, and then go back to the patient to collect targeted new data. For example, based on the knowledge of a certain disease, it is realized that a typical symptom needs to be supplemented to ask whether it has appeared.
Comparative analysis: DeepSeek's diagnosis and treatment process reflects a strong "knowledge → data" drive. With medical knowledge, DeepSeek will actively check whether key raw data is missing in the case. For this patient, DeepSeek may use its knowledge of tuberculosis to realize that "tuberculosis is often accompanied by weight loss and long-term low fever", so it checks whether the patient's information has data on weight changes or fever history; if not provided, DeepSeek will point out in the suggestion that it is necessary to know these data. Similarly, DeepSeek knows that patients with chronic bronchitis often have a history of smoking, so it may ask the patient whether he has been smoking for a long time. These are all manifestations of the model using knowledge to discover the incompleteness of the data layer information and supplement the question acquisition. However, since this is a single-round answer scenario, DeepSeek cannot directly ask the user, but it can indicate in the suggestion: "It is not clear whether the patient has XXX, if so..., if not...". For example: "Based on medical knowledge, it is necessary to understand the patient's recent weight changes. If there is obvious weight loss, it indicates chronic wasting disease." This shows that DeepSeek uses knowledge to check data gaps and lists obtaining these data as an action item.
ChatGPT also applies medical knowledge to reflect on the data, but its performance is more implicit, and it usually does not assume too much unknown data when giving advice. ChatGPT tends to use knowledge to explain existing data, rather than explicitly asking for missing information. Unless it is obviously necessary, ChatGPT generally does not ask additional questions in a single round of answers, because most of the information not provided by the user cannot be answered. However, ChatGPT may mention in its reply: "If the patient has a long history of smoking or repeated infection, it will be helpful for diagnosis", indirectly expressing concern about missing data in this hypothetical tone. It may also suggest: "If the patient has conditions, he can record changes in body temperature to help diagnosis (tuberculosis usually has low fever)." This is knowledge-guided data collection, but it appears in the form of advising patients to monitor. Overall, ChatGPT's knowledge-driven data supplement is more passive: it does not explicitly point out "You didn't tell me X, I need X", but gives a hypothesis of what would happen if X was provided. Compared with the directness of DeepSeek, ChatGPT may not ask obvious questions due to the limitations of the conversation format.
Therefore, DeepSeek actively seeks data in the "knowledge → data" module, clearly pointing out what additional data is needed based on knowledge; ChatGPT passively mentions data, implicitly suggesting the required data through assumptions and suggestions. The comparison is as follows:
Knowledge → Data Module Performance |
DeepSeek |
ChatGPT |
Knowledge gap |
Use medical knowledge to check for missing data and directly propose information that needs to be supplemented |
Use knowledge to make assumptions about possible data situations, and rarely ask directly about missing data (unless the conversation allows multiple turns) |
Data Acquisition |
Explicitly request or focus on specific data in the recommendations (weight changes, past medical history, etc.) |
Suggest patients to monitor or consider certain data, but do not force them to do so. Use a tactful tone. |
Pros and cons |
Check thoroughly to ensure no important clues are missed |
Respect the information given, don't make assumptions, communicate naturally but may miss details |
3.3.2Module 12: Knowledge → Information
Module meaning: The "knowledge → information" module refers to using existing knowledge to interpret or reorganize the content of the information layer. For example, doctors use medical theories to explain the patient's symptom combination (information), or hierarchically prioritize information based on knowledge. In clinical practice, this may be reflected in judging the importance of a symptom based on guideline knowledge, or citing epidemiological knowledge to evaluate the confidence of information.
Comparative analysis: ChatGPT frequently uses its medical knowledge to explain the importance of the patient's performance when analyzing case information, which is the process of "knowledge → information". For example, it knows that according to medical knowledge, "yellow-green sputum" means that bacterial infection is more likely, so it will emphasize this in its analysis, and regard "itchy throat" as secondary information at the information level, because it may only be a local manifestation of inflammation in knowledge and does not affect the overall situation. ChatGPT may also cite epidemiological knowledge, such as "48-year-old men with night sweats are more common in tuberculosis", to decide how to treat this information - for example, pointing out that "although night sweats suggest tuberculosis, considering that the prevalence of tuberculosis in this area is not high and the patient's treatment effect is good, this information may be secondary". By weighting and interpreting information with knowledge, ChatGPT's suggestions are closer to medical facts and experience. In short, it uses knowledge to distinguish the primary and secondary information: using professional knowledge to ensure that important information is valued and secondary information is downplayed. The output is a sentence that combines information and knowledge: "Because the sputum is purulent [information], based on experience it suggests bacterial infection [knowledge interpretation], so the current treatment effectiveness is in line with expectations."
DeepSeek also uses knowledge to interpret information, but it is often more explicit and detailed. For example, DeepSeek may add a statement such as "(in line with XX principle)" or "(this is usually seen in XX)" when analyzing symptom information one by one. For example, when mentioning night sweats, DeepSeek may say: "The symptom of night sweats is often associated with tuberculosis infection in medicine. It is the sympathetic nerve excitement that causes increased sweating at night [knowledge], so special attention should be paid to it in the information analysis in this case." In this way, DeepSeek explicitly marks the knowledge on the information to deepen the medical significance of the information. Or DeepSeek classifies the information according to the key points in the diagnosis and treatment guidelines, such as by risk factors, major symptoms, and secondary symptom lists, which is actually also using the knowledge framework to organize information. Compared with ChatGPT's slightly restrained approach, DeepSeek directly explains the impact of knowledge on information, allowing readers to understand why a certain information is important or unimportant. The advantage is clarity, but if there is a lot of information and many knowledge points, it may cause the information layer analysis to be long and somewhat repetitive.
In general, ChatGPT implicitly uses knowledge to regulate information weight in the "knowledge → information" module, making the output consistent with medical common sense; DeepSeek explicitly annotates information to support knowledge, making its analysis logic clear. The comparison is as follows:
Knowledge → Information Module Performance |
DeepSeek |
ChatGPT |
Knowledge Annotation |
Clearly annotate the medical significance or basis of each piece of information, and present knowledge and information in layers |
Integrate knowledge into the presentation of information, without separate annotations, and output more fluently |
Information Priority |
According to the importance of knowledge classification information, it is often explained one by one |
When expressing knowledge, the main information is emphasized, and the secondary information is glossed over. |
Pros and cons |
The interpretation is thorough and the logic is self-consistent but may be cumbersome |
Concise and fluent, with key points highlighted but no obvious traces of knowledge |
3.3.3Module 13: Knowledge → Knowledge
Module meaning: The "knowledge→knowledge" module refers to the deduction and integration within the knowledge layer, that is, linking multiple knowledge points to form new knowledge or selecting related knowledge. In medical treatment, this is manifested as doctors combining multiple medical knowledge to form a more systematic understanding of the disease, such as combining pharmacological knowledge and etiological knowledge to infer treatment plans, or comparing knowledge of different diseases for differential diagnosis.
Comparative analysis: DeepSeek often reflects the integration of cross-domain knowledge in diagnosis and treatment recommendations. It may connect drug knowledge, disease knowledge, test knowledge, etc. to provide a systematic analysis. For example, DeepSeek may synthesize it like this: "The improvement of the patient's infection shows that the current cephalosporin antibiotics are effective (pharmacological knowledge), and common allergenic pathogens are Streptococcus pneumoniae (microbiological knowledge); and night sweats remind me of tuberculosis (infectious disease knowledge), but tuberculosis is not sensitive to cephalosporins (combination of pharmacological + microbiological knowledge), so the possibility of tuberculosis is low. Combining the above medical knowledge, it is most likely a bacterial lung infection recovery period." In this analysis, DeepSeek cross-application of multiple aspects of knowledge, layer-by-layer verification, and finally formed a reliable knowledge conclusion. For another example, it may quote immunology knowledge to explain the patient's symptoms: "Night sweats may also be the body's immune response to infection (physiological knowledge)." This shows that DeepSeek is good at knowledge association: medical knowledge at different levels is mobilized to support or check each other. Such knowledge fusion increases the explanatory power, but also requires readers to have corresponding knowledge reserves to understand each point.
ChatGPT also has a wealth of medical knowledge, but it often does not lay out too many knowledge connections when outputting, so as not to make the answer too academic. It prefers to selectively apply key knowledge rather than fully display it. For example, for the same question just now, ChatGPT may simplify it to a sentence: "Cephalexin is effective in reducing the possibility of typical tuberculosis, because tuberculosis usually does not respond to this antibiotic." Here, ChatGPT only uses the relevant key knowledge (tuberculosis is ineffective against cephalosporin) to complete the knowledge reasoning of differential diagnosis, without elaborating on additional knowledge points such as pathogen categories. ChatGPT certainly "knows" that knowledge internally, but it presents it to meet the needs of the answer and does not pursue complete coverage of knowledge points. This method makes the answer more refined, but it also means that ChatGPT does not actively reveal many knowledge connections used in reasoning, and only presents conclusive knowledge. ChatGPT can also do this when a systematic explanation is needed, but in the default diagnosis and treatment recommendations, it is more like a concise clinician than a lecturer.
Therefore, in the "knowledge → knowledge" module, DeepSeek uses both breadth and depth of knowledge, connecting and analyzing multiple related knowledge points; ChatGPT selects knowledge carefully, using only the most necessary knowledge reasoning to avoid excessive expansion. The comparison is as follows:
Knowledge → Knowledge module performance |
DeepSeek |
ChatGPT |
Knowledge Synthesis |
Comprehensive analysis based on multidisciplinary knowledge (pharmacology, pathology, diagnostics, etc.) |
Select the most relevant knowledge for reasoning, and do not expand additional knowledge topics |
Identification Reasoning |
Use multiple knowledge sources to verify/eliminate each other and improve the reliability of conclusions |
Use key knowledge to make concise identification arguments and quickly draw conclusions |
Pros and cons |
Comprehensive and professional, but the amount of information is a bit complicated |
Efficient focus, no wasted attention but some of the reasoning basis is hidden |
3.3.4Module 14: Knowledge → Wisdom
Module meaning: The "knowledge → wisdom" module refers to the decision-making of the wisdom layer directly guided by knowledge. That is, based on general medical knowledge or principles, the treatment decision for specific patients is directly formed, rather than relying solely on on-site information. In medical care, this is usually reflected in following evidence-based medicine and guidelines: doctors formulate treatment plans based on existing knowledge (such as guideline recommended therapies).
Comparative analysis: ChatGPT places great emphasis on the evidence-based principle when giving medical advice, which is a typical application of "knowledge → wisdom". For this case, it is very likely to directly formulate a treatment strategy based on the guideline knowledge of respiratory tract infection. For example, ChatGPT may suggest: "According to clinical guidelines, for community-acquired bacterial bronchial infections, a 7-day course of cephalosporin antibiotics is recommended [evidence-based knowledge]. The patient has been taking medication close to the end of the course and should continue until the course is completed. Assisted with expectorants such as ambroxol [drug knowledge], it is not recommended to use glucocorticoids at this stage (without obvious airway obstruction indications) [guideline knowledge]." These suggestions are obviously derived from the direct application of standard knowledge, rather than being completely derived from case information. ChatGPT has a large amount of medical knowledge base, including treatment norms, drug dosages, follow-up time, etc., so the scheme it gives is close to authoritative recommendations [user prompts that ChatGPT proposes an evidence-based standard medication regimen]. The advantage of this method is that it is reliable and meets standards. Doctors or professional readers can understand the basis of the scheme at a glance, and patients can also receive standardized treatment. The disadvantage may be a relative lack of personalization, but in this case the patient's situation was not special and following the guidelines was best practice.
DeepSeek will also refer to knowledge and guidelines to develop plans, but sometimes, for comprehensive considerations, it may go beyond the existing knowledge framework to make more arrangements. For example, in addition to recommending the completion of the antibiotic course like ChatGPT, DeepSeek may also mention some additional measures, such as: "Considering that the patient's night cough affects sleep, antitussive drugs can be used as appropriate (this is general clinical experience, not necessarily specified in detail in the guidelines); patients have night sweats, and appropriate nutrition can be supplemented to strengthen their physical fitness (health knowledge). At the same time, remind patients to avoid cold and fatigue, which is common sense for preventing respiratory infections." These details reflect that DeepSeek applies rich medical and health knowledge to decision-making, including not only the hard knowledge of evidence-based medicine, but also clinical experience and general health principles. This makes DeepSeek's plan more comprehensive, but sometimes it may also include some suggestions outside the guidelines. As long as these do not conflict with the evidence, they are usually beneficial. But in a strict sense, DeepSeek's plan may not be as straightforward as ChatGPT in copying authoritative knowledge, but rather expanding and flexibly adjusting it. For example, if DeepSeek is concerned about night sweats, it might also recommend an additional imaging test, which is an extra step beyond the standard process but is also a decision made out of an abundance of caution based on its knowledge of the importance of tuberculosis screening.
Therefore, in the "knowledge → wisdom" module, ChatGPT makes decisions strictly based on medical knowledge/guidelines, while DeepSeek makes flexible decisions based on comprehensive knowledge and experience. The comparison between the two is as follows:
Knowledge → Wisdom Module Performance |
DeepSeek |
ChatGPT |
Decision basis |
Combining guidelines with clinical experience to develop plans from multiple angles |
Formulate plans based on authoritative medical knowledge (guidelines, evidence-based research) |
Program Content |
In addition to standard treatment, additional recommendations such as personalized or preventive care |
Focus on standard treatment and monitoring regimens, rarely going beyond guidelines |
Pros and cons |
Comprehensive and thoughtful, personalized consideration more |
Reliable and standardized, strictly following evidence-based standards |
3.3.5Module 15: Knowledge → Intention
Module meaning: The "knowledge→intention" module means using existing knowledge to shape or adjust the intention-level goals. That is, medical knowledge affects the ultimate goal setting of doctors, such as giving priority to protecting life and following the patient's wishes. In clinical practice, some decision-making goals of doctors (such as "cure" and "improve quality of life") are often influenced by medical ethics and experiential knowledge.
Comparative analysis: As an AI model, DeepSeek also reflects a certain "value" or "goal" orientation when answering, which often comes from its built-in medical knowledge and ethics. For example, through knowledge, it understands that the primary intention of medical treatment is to cure diseases and reduce risks. So in the suggestions, DeepSeek may emphasize: "Our primary goal is to clear infections and prevent complications such as lung abscesses or spread [reflecting the goal of knowledge guidance]." At the same time, DeepSeek may have another intention: "Since the possibility of tuberculosis has been considered, our goal also includes ensuring that patients do not suffer from such chronic infectious diseases to ensure their long-term health [goals under knowledge + public health awareness]." These can be reflected in the suggestions as a dual focus on efficacy and safety. DeepSeek will even add a sentence of comfort or expectation at the end, such as "I hope that through standardized treatment, patients can fully recover and return to normal life", which reflects the influence of medical humanities knowledge on its goals (that is, not only to cure diseases, but also to restore people's quality of life). DeepSeek's goal setting is relatively comprehensive, including both short-term (curing infections), long-term (preventing chronic diseases), and overall health. It can be seen that its knowledge system has shaped a comprehensive intention.
The intention of ChatGPT is usually closer to clinical reality and clear and single: for example, the current main goal is to cure this infection. Because according to medical knowledge, the patient does not have a more serious condition, ChatGPT will not set too many additional goals. Its answers focus on cure and avoidance of recurrence, for example: "The goal is to completely control the infection, avoid complications, and promote the patient's full recovery." These are actually the same as the general direction of DeepSeek, but ChatGPT generally does not mention broader goals such as "protecting public health" or "long-term health" because they are a bit general for a specific patient consultation. At the intention level, ChatGPT appears pragmatic - using medical knowledge to tell it what is most important at the moment, and use this as the intention. For example, ChatGPT may not talk about "excluding tuberculosis to protect the public" in this case, because knowledge tells it that the probability of tuberculosis is low, so its intention is focused on the disease management at the moment. If another case needs to consider public health, it will also reflect that intention. So ChatGPT's knowledge→intention mapping is dynamic and focused on current needs: it does not list irrelevant goals.
In summary, DeepSeek has established short-term, long-term and comprehensive goals in the "knowledge → intention" module in multiple dimensions, driven by knowledge; ChatGPT focuses on the main intention and selects the most important goals based on knowledge to emphasize. The comparison is as follows:
Knowledge → Intention Module Representation |
DeepSeek |
ChatGPT |
Goal Setting |
Inspired by medical knowledge, set multi-level goals (cure, prevention, long-term health) |
Focus on the current main goal based on knowledge (cure the current disease, prevent recurrence) |
Ethical considerations |
May mention higher-level purposes such as public health and infection prevention |
Generally limited to the patient's own treatment goals, not extended to pan-social goals |
Pros and cons |
Goals are comprehensive but may exceed the patient's immediate concerns |
The goal is clear and fits the current condition, but the scope of consideration is narrow |
3.4Conversion module starting from the intelligence layer
3.4.1Module 16: Wisdom → Data
Meaning of the module: The "Wisdom → Data" module refers to starting from the decision-making of the wisdom layer and feeding back to the data layer to collect information or verify assumptions. That is, after making a certain decision, some of the lowest-level data are rechecked to ensure that the decision is correct or successful. In clinical practice, this is manifested as a review after the decision: after the doctor decides on a plan, he may review key test results or ask the patient to monitor changes in symptoms so that the strategy can be adjusted in time according to the data.
Comparative analysis: After DeepSeek proposes a diagnosis and treatment decision plan, it often includes arrangements for follow-up data tracking, which is the feedback from the intelligent layer to the data layer. For example, when DeepSeek recommends that patients continue antibiotic treatment and observation, it will specify which data needs to be monitored: "During the implementation of the treatment, it is necessary to monitor the patient's temperature changes, cough and sputum volume. If fever recurs or sputum volume increases, a follow-up visit should be made in time." Here, DeepSeek has made an intelligent decision (continue the current treatment plan), but it has proactively formulated a data monitoring plan to support the decision. This reflects DeepSeek's strong sense of safety net: it knows that any treatment decision needs to be evaluated and adjusted based on the latest data, so it includes clear review nodes or indicators in the recommendations. For example, it may recommend "reviewing a chest X-ray after the course of treatment to ensure that the inflammation is completely absorbed", even if the guidelines may not require it, but this is a manifestation of DeepSeek's intelligent feedback data to confirm the cure. These arrangements can be regarded as data verification measures made by DeepSeek for its own decisions to ensure that there is no risk of error.
ChatGPT also mentions follow-up and observation, but it is usually not as specific as DeepSeek. ChatGPT may simply say: "Pay attention to changes in symptoms. If fever or cough worsens, please seek medical attention in time." This is a general suggestion to give patients a reminder, but there is no clear quantification of the data that needs to be observed. This is in contrast to DeepSeek: DeepSeek almost formulates a "monitoring list", while ChatGPT gives a general guidance such as "If it doesn't get better, see a doctor." Of course, ChatGPT sometimes recommends rechecking a certain examination, but generally when there are clear risk factors. In this case, ChatGPT believes that it is a recovery period and may not require a certain examination to be done. It will only say "If symptoms recur or worsen, consider checking again." Therefore, the feedback of ChatGPT's intelligent layer on data is passive: patients need to decide whether to obtain new data (visiting a doctor, examination) based on their feelings. DeepSeek is active: new data is obtained at preset times or indicators.
Therefore, DeepSeek in the "Intelligence → Data" module is an active review, incorporating data collection after decision-making into the plan; ChatGPT is a passive follow-up, generally leaving it to the patient to pay attention to, and only returns to the data layer when necessary. The comparison is as follows:
Wisdom → Data module performance |
DeepSeek |
ChatGPT |
Post-decision monitoring |
Clearly formulate a monitoring plan (indicators, time), and proactively obtain new data to verify efficacy |
Patients are advised to observe changes in symptoms on their own and collect data when abnormalities occur. |
Review Arrangements |
Follow-up examinations are routinely scheduled to ensure successful decision making |
May not emphasize specific review, or may present it in the form of conditions |
Pros and cons |
Active follow-up to ensure quality but increase the burden of follow-up visits |
Flexibility reduces unnecessary inspections but may delay the discovery of problems |
3.4.2Module 17: Wisdom → Information
Meaning of the module: The "Wisdom → Information" module means starting from the wisdom layer (decision-making results) to adjust or reinterpret the content of the information layer. In other words, decision-making considerations will prompt a re-examination of existing information, or transform the decision-making results into new information transmission. In clinical practice, this includes two aspects: one is to reflect on information - doctors will check existing information before and after making decisions to ensure that they are correct; the other is to convey information - when doctors inform patients of the decision, they will express it in a way that patients can understand.
Comparative analysis: After forming a treatment plan, DeepSeek usually "checks back" to see if the logic of the previous information is fully explained. This is a verification of the information layer after the intelligent layer makes a decision. For example, if DeepSeek finally diagnoses acute bronchitis and decides on conventional treatment, it will look again to see if there is a reasonable explanation for the previous night sweats information (because if the intelligent layer ignores the night sweats, it will cause suspense in the information layer). So DeepSeek may add a sentence to the suggested explanation part: "Although there are night sweats, combined with the overall situation, we judge that night sweats are a non-specific manifestation of the infection process, rather than information suggesting other diseases." This is equivalent to giving an explanation for the previously confusing information after the decision is finalized, so that the information layer is consistent with the intelligent decision. This process reflects the rigor of DeepSeek's review after decision-making.
On the other hand, DeepSeek’s detailed wording in the recommendations can also be understood as translating professional decisions into understandable information for readers (patients or doctors). It may say: "The measures we take are to continue the existing treatment and observe, because this is a safe and effective approach based on the current situation." Here, DeepSeek expresses the decision and its basis, making the content of the information layer richer. It can be said that DeepSeek attaches great importance to making information self-consistent and fully conveyed: it reflects and verifies information upwards, and interprets decisions as information downwards.
After making a plan, ChatGPT handles information more concisely. It generally assumes that the previous information and decision are consistent, and does not do too much self-examination. If there is obviously unexplained information, it will also mention it, but usually ChatGPT has solved most of the contradictions during information analysis and will not repeat the explanation at the end. ChatGPT is more concerned with telling the user the decision clearly. It will directly say: "Your condition is getting better now. We think you are in the recovery period. Just complete the course of treatment and observe. No additional intervention is required." This sentence itself expresses the content of the wisdom layer in popular information for patients to understand. Unlike DeepSeek, ChatGPT will not explain in detail why night sweats are okay because it has been dealt with lightly before. This makes the conclusion part very concise and clear. ChatGPT assumes that the patient has followed the chain from information to decision, so it only emphasizes the conclusive information.
In summary, DeepSeek emphasizes information calibration and detailed explanation after decision-making in the "Wisdom → Information" module, while ChatGPT emphasizes direct transmission of conclusion information and less redundant checks. The comparison is as follows:
Wisdom → Information Module Performance |
DeepSeek |
ChatGPT |
Decision-making feedback information |
After the decision is finalized, review and explain the previously doubtful information to ensure that the information and decision are consistent |
Information is generally considered fully processed and no further explanation is given unless necessary. |
Decision communication |
Detailed description of decision content and reasons to facilitate readers' understanding and acceptance |
Directly inform the decision results and simple reasons, and be concise and to the point |
Pros and cons |
The explanation is sufficient, rigorous and self-consistent, but the conclusion is long |
Communicate clearly and efficiently, but may omit details. |
3.4.3Module 18: Wisdom → Knowledge
Module meaning: The "Wisdom → Knowledge" module refers to the generation or updating of knowledge from the decision-making practice of the wisdom layer. This is usually done after solving a problem, summarizing the experience and lessons, and elevating them into rules. In medical care, it is reflected in case teaching points: after making a diagnosis and treatment decision, doctors will summarize the enlightenment brought by the case and enrich their knowledge base.
Comparative analysis: When DeepSeek, as an AI, outputs one-time suggestions, there is no real process of "learning" knowledge, but it sometimes reflects a summary improvement at the end of the suggestions. For example, DeepSeek may add some warnings or experiences for similar situations: "This case reminds us that for middle-aged men with cough and night sweats, we should remain vigilant against tuberculosis while common infections improve, but excessive examinations may not be necessary and a balance is needed." This is actually extracting a general knowledge point from the handling of this case. This statement can be seen as DeepSeek simulating doctors summarizing their experience and refining smart decisions into knowledge. For example, it might say "Therefore, for patients with cough who respond to antibiotics, the likelihood of tuberculosis is significantly reduced, which is an empirical point." Of course, such a summary may not always be there, but DeepSeek style prefers to be detailed, and sometimes there will be a sublimation paragraph to let readers know the significance of this example. It can be said that DeepSeek is generous in sharing its inferences, which shows its educator-like nature.
ChatGPT rarely adds this kind of knowledge distillation to its responses to patients, because it focuses on answering the questions themselves and will not add unnecessary lessons (unless the user asks). ChatGPT may summarize when it is facing developers or doctors, but generally not for patients. For example, in this case, ChatGPT will not say to the patient "What is the lesson of this case." Therefore, ChatGPT basically does not reflect "wisdom → knowledge" in the surface output. However, we can guess that in its internal logic, each answer also reinforces certain patterns, but that is model learning and not reflected in the conversation. So in comparison, the significance of this module to the final visible output, DeepSeek may occasionally reflect case lessons, and ChatGPT usually omits.
Therefore, in the "Wisdom → Knowledge" module, DeepSeek may output experience summaries (optional, not required), while ChatGPT generally does not explicitly summarize. The comparison is as follows:
Wisdom → Knowledge Module Performance |
DeepSeek |
ChatGPT |
Lessons Learned |
Sometimes summarize the case inspiration or principle at the end to deepen the reader's understanding |
Basically do not summarize experience in the answer, focus on the current problem |
Knowledge Update |
Simulate human thinking in output and transform decisions into useful knowledge points in the future |
Not reflected (knowledge may only be updated internally or later) |
Pros and cons |
Helps with learning, but not directly needed by patients |
Less irrelevant content in the answers, more in line with the question-and-answer scenario |
3.4.4Module 19: Wisdom → Intention
Module meaning: The "Wisdom → Intention" module represents the feedback or adjustment of the results of the wisdom layer to the intention layer. That is, after a plan is implemented or decided, it may cause reconsideration or confirmation of the ultimate goal. In medical treatment, this may be the doctor and the patient re-evaluating the treatment goal after a period of treatment, or reflecting on whether the initial goal has been achieved after the decision is made.
Comparative analysis: Since DeepSeek and ChatGPT mainly provide solutions in one interaction without follow-up, strictly speaking, they cannot really achieve "adjusting intention after intelligent implementation" in their answers. This module is more like something that happens in long-term management. However, we can speculate whether their answers reflect the meaning of "expecting to achieve intention". ChatGPT usually ends with a wish or goal confirmation, such as: "I hope you will get well soon." Although this is a polite remark, it can also be regarded as a reaffirmation of the goal after intelligent decision-making - the intention is to make the patient recover. This is actually the original intention being re-emphasized. ChatGPT's speech usually stops here, and no more discussion of the goal is made.
DeepSeek may also have similar statements. However, DeepSeek sometimes mentions several emergency intentions in the plan: for example, "If XX situation occurs during treatment, our goal will be to deal with complications in a timely manner." This shows that DeepSeek not only has the current intention to cure, but also has a plan to modify the goal once the situation changes. However, in a single output, this is just a linguistic explanation, not a real follow-up action.
In general, the outputs of both modules are not very different, and neither module is developed in depth. DeepSeek may have more "if the goal is not reached then...". ChatGPT is basically a general wish. So the summary table is:
Intelligence → Intention Module Performance |
DeepSeek |
ChatGPT |
Goal Reaffirmation |
The goal may be restated or adjusted at the conclusion (adjust the goal if…) |
Usually ends with a wish, implying the ultimate goal |
Follow-up Intentions |
Mention the readiness to change goals if circumstances change |
Rarely mention changes in subsequent goals, only emphasize current rehabilitation goals |
Pros and cons |
More comprehensive expression, but limited practical significance |
Concise and in line with the scene, no additional information |
3.5Conversion module starting from the intent layer
3.5.1Module 20: Intention → Data
Module meaning: The "Intent→Data" module refers to collecting or selecting data from the highest level of intent. That is, there is a goal first, and then the required raw data is obtained to achieve this goal. In medical diagnosis, this is usually manifested as targeted questioning and physical examination: if the doctor has a suspicion/intention in mind (such as ruling out a disease), he will collect relevant symptom and sign data in a targeted manner.
Comparative analysis: Many of the "recommendations to check X" reflected in DeepSeek's answers are actually intent-driven data collection. For example, its suspicion of tuberculosis prompted the recommendation of sputum examination, which has been analyzed above. Another example is that its treatment intention is to completely cure the infection, so it may recommend a lung function or imaging test at the end of the treatment to ensure that the lesions are eliminated, which is also intent-driven data acquisition (healing without hidden dangers). These suggestions of DeepSeek are relatively obvious and direct, and we have summarized them in the previous related modules.
ChatGPT also has an intent-driven data collection part, but it is more moderate. For example, its intention is not to miss serious diseases, so it recommends "If symptoms recur, consider checking X." This is the same as DeepSeek in terms of motivation, but the tone is not so absolute. When the intention does not conflict with the current information (such as ChatGPT determines that it is mainly bronchitis recovery and there is no greater doubt), it will not strongly drive new data collection. Therefore, ChatGPT's performance in this module is relatively weak.
Therefore, this part can refer to the above-mentioned "data ←→ intention" and "knowledge → data". In short:
Intent → Data Module Performance |
DeepSeek |
ChatGPT |
Targeted inspection |
Clearly propose specific data acquisition (inspection/inquiry) based on the objectives |
Purposefully suggesting data acquisition but often with conditions |
Motivation intensity |
As long as there is a goal, actively seek data support |
Obtaining data is recommended only when suspicion is strong or necessary |
Pros and cons |
Goal-driven, relentless pursuit of evidence |
Focus on practicality and avoid unnecessary checks for your goals |
3.5.2Module 21: Intention → Information
Module meaning: The "Intent→Information" module indicates that the top-level intention selectively focuses on or organizes information. That is, according to the final goal, the information most relevant to the goal is selected for processing or presentation. In medical treatment, this is reflected in the doctor interpreting the medical record with a purpose, and only focusing on the symptom information related to his hypothesis.
Comparative analysis: ChatGPT's answers tend to organize information around its main diagnostic intent (curing bronchitis). When analyzing, it focuses on describing information closely related to this intent (cough, sputum, antibiotic effect), while ignoring irrelevant information (such as other irrelevant situations that the patient may have mentioned, if any). This shows that ChatGPT filters information: information that meets its intent is emphasized, and irrelevant information is less mentioned. ChatGPT's treatment of night sweats is also based on intent: its intention is not to really diagnose tuberculosis, so it downplays the night sweats information because night sweats are not core information for the main intent (curing bronchitis).
DeepSeek is more exhaustive, so even if it has an intention (such as the dual intention of excluding tuberculosis and curing infection), it will still process almost all the information provided. However, the details of the processing may be different: it will be more certain about the information related to the main intention (curing infection), and it will also mention information related to the secondary intention (suspected tuberculosis), but it may be marked with uncertainty. DeepSeek will not simply discard the input information because it pursues comprehensiveness, but it will present information by intent, such as discussing it in terms of major and minor. Therefore, DeepSeek sometimes clearly divides the "main problem" and "minor problem to be sorted" in the suggestions, which is also an intention-driven means of organizing information.
Therefore: ChatGPT tends to filter information to meet its intent; DeepSeek tends to classify information to serve multiple intents. Comparison:
Intent → Information Module Performance |
DeepSeek |
ChatGPT |
Information Selection |
Do not ignore the information provided, but prioritize the discussion by intent |
Focus on the information required for the purpose; secondary information may be omitted or simplified |
Organizational Structure |
Possibly segment information by primary goal/secondary goal |
The overall narrative revolves around the main goal, with little mention of others |
Pros and cons |
Comprehensive but clearly structured |
Highly focused on the main issue, concise but may miss minor points |
3.5.3Module 22: Intention → Knowledge
Module meaning: The "Intent→Knowledge" module means that the final intention determines which knowledge base to call. That is, according to the goal, select the corresponding field or type of knowledge to assist reasoning. In medical treatment, for example, if the intention is to treat infection, mobilize antibiotic-related knowledge; if the intention is to exclude tuberculosis, mobilize tuberculosis-related knowledge.
Comparative analysis: When DeepSeek has two intents at the same time (treating common infections + ruling out tuberculosis), the knowledge it calls obviously covers two areas: one is about the treatment of common respiratory infections (antibiotic courses, nursing, etc.), and the other is about tuberculosis identification (tuberculosis symptoms, diagnostic criteria knowledge). In the output, we can see that DeepSeek does show these two aspects of knowledge at the same time. This shows that DeepSeek "queries" relevant knowledge to support its own suggestions based on each intent. This parallel call ensures the completeness of the solution, but also makes the content longer.
ChatGPT's main purpose is to treat bronchitis, and tuberculosis is only a secondary concern, so it calls knowledge based on the main purpose. We see that ChatGPT output focuses on knowledge points such as antibiotics and convalescent management, and only briefly mentions tuberculosis (such as "consider tuberculosis testing if..."). It hardly expands on tuberculosis treatment knowledge, because that does not fit the current purpose. This reflects that ChatGPT strictly calls knowledge based on key points: tuberculosis knowledge in this case is considered not to need to be elaborated, so it is hardly mobilized. Infection management knowledge is the core, so it is quoted from classics.
In general, DeepSeek uses multiple intents to call multiple knowledge in parallel; ChatGPT focuses on calling related knowledge based on the main intent, and does not delve into other aspects.
Intention → Knowledge Module Performance |
DeepSeek |
ChatGPT |
Knowledge call range |
Broadly invoke multifaceted knowledge based on all relevant intents |
Mainly calls corresponding knowledge for primary intent, and the knowledge call for secondary intent is very limited |
Depth of knowledge |
The knowledge involved in each intention may be described in detail |
Just touch upon non-core knowledge and focus on key knowledge areas |
Pros and cons |
Rich content, full support for various goals |
Be concise and effective, avoid irrelevant developments but may overlook minor areas |
3.5.4Module 23: Intention → Wisdom
Module meaning: The "Intention → Wisdom" module indicates that the top-level intention directly drives the decision-making of the wisdom layer. That is, after clarifying the ultimate goal, we directly start to formulate a plan to achieve the goal, and the intermediate reasoning process is implicitly completed in practice. In clinical practice, this is similar to "Goal-directed Therapy": set a goal at the beginning, such as "clear the infection as soon as possible", and then directly select the treatment method around this goal.
Comparative analysis: ChatGPT's diagnosis and treatment recommendations appear to be very "goal-oriented". It basically assumes that the patient is recovering, and the goal is to make the patient fully recover without any mistakes. Therefore, its plan serves this goal very directly: continue medication, rest, and observation, all for recovery. There is not much hesitation or entanglement, because ChatGPT has already determined the intention = cure bronchitis, so the decision follows the intention. It can be said that ChatGPT directly transforms the intention into a series of practical measures (wisdom layer), which are coordinated and pointed to recovery. It does not have additional temptations because the intention is clear and single.
DeepSeek has dual intentions (treating the disease + excluding other diseases), so the decision must also satisfy both aspects at the same time, which makes its solution more complicated than ChatGPT. It needs to implement the main treatment while adding measures such as inspection and follow-up to achieve the intention of excluding tuberculosis. Therefore, DeepSeek's decision can be seen as the result of multi-objective optimization: it not only treats the current disease, but also takes care of future troubles. Therefore, its smart solution has a treatment part, as well as an inspection and observation part. DeepSeek is not without goals, it has diverse goals. Each goal directly contributes to a part of the smart decision-making action. For example, the goal of "treating infection" leads to the decision to continue using antibiotics + symptomatic treatment, and the goal of "excluding tuberculosis" leads to the decision to recommend sputum examination or regular review. DeepSeek does not abandon any intention, so it is not as pure as ChatGPT, but its decision is safer and more rigorous.
In summary, ChatGPT has a single and firm intention, and its decision-making is purely to execute that intention; DeepSeek has multiple intentions, and its decision-making takes all intentions into consideration.
Intention → Intelligent Module Performance |
DeepSeek |
ChatGPT |
Goal-oriented |
Multiple goals lead to different decision-making measures, making the plan more complicated |
A clear single goal directly guides the plan, and the plan is concise and consistent |
Decision-making coordination |
The need to balance different intentions in decision making (treatment and screening in parallel) |
Decisions are made solely to serve one purpose, without weighing other |
Pros and cons |
More than comprehensive, maybe a little cumbersome |
Straightforward, strong execution but slightly less flexible |
3.5.5Module 24: Intention → Intention
Module meaning: The "Intention→Intention" module refers to the internal adjustment and consistency of the intention layer itself. This can be understood as self-reflection or confirmation: whether there is a conflict between high-level goals and whether the order needs to be adjusted. In medical treatment, if a doctor has multiple goals at the same time (curing diseases, alleviating symptoms, reducing costs, etc.), he will sort or choose in his mind.
Comparative analysis: Due to its wide range of considerations, DeepSeek actually has multiple intentions and needs internal balance. When it gives advice, it implicitly balances the two intentions of cure and no missed diagnosis. We can see that it did not shift the patient's entire treatment plan to focus on tuberculosis because of the intention of excluding tuberculosis - after weighing the pros and cons, it still regards the treatment of bacterial infection as the main intention and the exclusion of tuberculosis as the secondary intention. This decision reflects DeepSeek's adjustment of its own intentions: it knows what the main contradiction is. Although it is not stated explicitly, we can feel that DeepSeek has a main and a secondary when organizing advice. This shows that it has achieved a good reconciliation of "intention→intention" and reconciled potentially conflicting goals.
ChatGPT is almost a single intent, so there are no complex internal conflicts to adjust. The only thing it may need to pay attention to is to strike a balance between treatment and avoiding over-medicalization (this is a common conflict in medical care). ChatGPT clearly prefers not to over-examine, which means that in its goal, "avoiding over-examination" outweighs "absolutely not missing extremely low-probability diseases." So it can be said that ChatGPT also balances intent, but it is simpler and clearer. It doesn't even mention additional examinations and directly abandons the secondary intent of excluding tuberculosis. This is also a trade-off: fully follow the main intent (treat the current disease) and abandon the low-priority intent. This decision is reasonable in most cases.
Therefore, DeepSeek balances multiple intents to achieve a balance, while ChatGPT reduces secondary intents to a very low level or even ignores them, keeping the intent simple. Comparison:
Intent → Intent module performance |
DeepSeek |
ChatGPT |
Multi-objective balance |
The priorities are clear, ensuring the main goals while taking into account the secondary goals as much as possible. |
There is little need for balancing, and secondary goals are weakened or ignored. |
Goal Alignment |
Ensure that different intentions do not conflict and are reflected in decision-making |
The main purpose is the center, and other purposes are subordinate to it. |
Pros and cons |
Comprehensive consideration, multi-faceted consideration |
Focus on the core and unify actions |
3.5.6Module 25: Intention → Wisdom (repeat the starting point of intention?*)
(Note: Module 25 should be "Intention → Wisdom", but the above analysis shows that Module 23 is Intention → Wisdom. It is possible that the original Module 25 refers to the intention layer itself, as Intention → Intention has been analyzed in 24, so it may be repeated here and omitted)
(Editor's note: The 25 modules of the DIKWP framework cover bidirectional conversions between layers, including the same layer. This section has analyzed all possible inter-layer paths one by one. Among them, since the intention layer to the intelligence layer has been discussed in 23, module 25 is understood as another perspective of the internal conversion of the intention layer itself (intention→intention) or the internal conversion of the intelligence layer itself. However, since intention→intention has been discussed in 24, it is omitted here to avoid repetition.)*
In practical applications, we can consider that all 25 modules have been discussed.
4Summary and evaluation: advantages and disadvantages of the two models of diagnosis and treatment of cognitive ability
In summary, through the analysis of the 25 cognitive modules of the DIKWP*DIKWP model, we can clearly see the respective characteristics of DeepSeek and ChatGPT in handling medical diagnosis and treatment tasks. ChatGPT shows a concise, efficient, and evidence-based cognitive style: it is good at grasping the main contradictions, quickly applying medical knowledge to give solutions that meet the guidelines, and is very concise and clear in data processing and information synthesis. The decision-making process tends to be intuitive but generally reliable. The advantages of ChatGPT's answers are that they highlight the key points, have a clear structure, and are highly executable. For this case, ChatGPT accurately identified the most likely diagnosis (acute bronchitis recovery period) and proposed a solution in strict accordance with the standard treatment process, while avoiding excessive examinations or complex analysis. This reflects its strong pattern recognition and knowledge call capabilities, enabling it to give professional and concise advice on common cases. This style is very suitable for the initial diagnosis and treatment of common and frequently occurring diseases, and helps improve efficiency and patient compliance.
However, ChatGPT's weakness also lies in its over-reliance on the main pattern, and its relatively hasty handling of some atypical information. For example, the symptom of "night sweats", which deviates from the norm, was only lightly dealt with in this case. Although it was mentioned, the identification was not discussed in depth. This reflects that ChatGPT may not be vigilant enough when facing potential rare situations, and tends to ignore low-probability events, resulting in the risk of missed diagnosis. In addition, the reasoning process of ChatGPT is implicit, and it is not easy for people to see its basis. Whether each step is rigorous and credible requires extra trust. If more complex cases or situations with large information contradictions and conflicts are encountered, ChatGPT may have logical inconsistencies or misjudgments, which is a risk in medical scenarios.
In contrast, DeepSeek demonstrates a comprehensive, meticulous, and comprehensive cognitive style: it answers almost every question about the input information, analyzes it item by item, and strives not to miss any details and possibilities. In terms of data and information processing, DeepSeek lists and explains it in detail; in terms of knowledge application, it draws on multiple sources, taking into account everything from guidelines to experience; in decision-making, it provides multi-faceted guarantees, including treatment, examination, and follow-up. This makes DeepSeek's suggestions very complete and rich, and can provide doctors and patients with more background knowledge and options. For example, in this case, DeepSeek not only recognized bacterial infection and gave a standard treatment plan, but also paid attention to other possible indications such as night sweats, and made corresponding examination and follow-up suggestions. This steady and prudent approach reduces the probability of missed diagnosis and misjudgment, and is more advantageous in dealing with difficult or atypical cases. Especially in the high-level intentions and feedback links involved in the 25 modules, DeepSeek showed a strong sense of safety and responsibility: actively arranged for review and verification, clearly stated the goals and tasks, and ensured a closed loop of diagnosis and treatment.
The shortcomings of DeepSeek are that it is lengthy and complicated, and the priorities are not clear enough. It tries to cover all aspects, and sometimes the core opinions are drowned in the ocean of information. If the patient reads it directly, he may feel confused or overburdened, and it is not as clear as ChatGPT. For clinicians, DeepSeek's detailed reports also require time to refine the key points. In addition, DeepSeek may occasionally have a tendency to over-medicalize: a series of examinations proposed for safety reasons may not be necessary in actual clinical practice, which may increase the burden on patients. For example, in this case, if the possibility of tuberculosis is extremely low, sputum examination and imaging are still recommended, which invisibly increases medical costs. This reflects DeepSeek's goal-driven strategy of preferring to do more and not wanting to miss it. Although this is safe, it deviates slightly from the principle of "moderation" in real medical care.
In general, ChatGPT is more like an experienced and guideline-compliant general practitioner, quickly giving the main diagnosis and standard treatment, but not paying enough attention to atypical clues; DeepSeek is like a meticulous and rigorous expert consultant, comprehensively analyzing various possibilities and details, but may appear redundant and conservative. The cognitive abilities of the two have their own advantages, but when combined, they complement each other: ChatGPT is good at efficiency and standardization, and DeepSeek is good at depth and reliability.
5Suggestions for improving the DeepSeek model
Based on the above analysis, we can propose improvement suggestions for the problems exposed by DeepSeek in this case, so as to enhance its diagnosis and treatment cognitive ability and better balance comprehensiveness and practicality:
1. Optimize the decision path focus: DeepSeek should learn to highlight the main contradictions and decision paths after completing a comprehensive analysis. You can learn from ChatGPT's approach and clearly inform the "primary diagnosis/treatment direction" in the summary, and briefly list the secondary considerations instead of describing them in equal amounts. This can prevent information overload and allow readers to grasp the core points of DeepSeek's suggestions. In terms of technical implementation, a weight mechanism can be introduced when generating content to strengthen the proportion of content corresponding to the main intention and appropriately compress the secondary intention. For example, in this case, DeepSeek can add a concise summary after the detailed description: "In summary, the current diagnosis is most likely the recovery period of acute bronchitis, and the main plan is to continue anti-infection treatment and observation. The possibility of tuberculosis is extremely low, and only routine follow-up vigilance is required." This conclusive path focus will greatly improve the readability and practicality of the recommendations.
2. Optimization of the selected medication module: DeepSeek should be more evidence-based and targeted when providing medication recommendations, and avoid giving too many or unnecessary drugs. In this case, cephalosporin treatment has been effective, and DeepSeek should focus on completing the course of treatment and basic care, without introducing additional broad-spectrum antibiotic combinations, hormones, Chinese patent medicines, etc. (if it mentions them). The improvement method is to constrain medication selection in the knowledge → wisdom module and strictly follow the guideline recommendations. At the same time, introduce a module to consider individual patient factors (allergy history, comorbidities) to ensure that medication is safe and effective but not excessive. For example, a rule can be added: For bacterial infections whose symptoms have been relieved, it is not recommended to upgrade antibiotics or use multiple drugs at the same time, but to follow the principle of minimum effective treatment. The medication module should also provide clear dosage and course recommendations to make its recommendations more operational and credible.
3. Adjustment of inspection strategy: DeepSeek should pay more attention to rationality and timing in the inspection module. At present, it tends to "check if there are doubts", but the actual clinical practice emphasizes observation before decision. It is recommended that DeepSeek introduce "time window" and "conditional judgment" on the path of wisdom → data and intention → data. For example, for the screening of tuberculosis, DeepSeek can suggest: "Observe for 2 weeks first. If the symptoms disappear completely, no additional examination is required; if there are still night sweats or coughs that have not healed, sputum examination and imaging examination should be performed again." This not only reflects the vigilance against potential serious diseases, but also avoids unnecessary examinations immediately. This phased inspection strategy is in line with clinical decision-making thinking. In terms of technical implementation, by adding threshold conditions to the decision tree, AI can learn to trigger inspection recommendations based on the development of the disease, rather than listing them all at once.
4. Improve information extraction and expression: DeepSeek can learn from ChatGPT's strengths in the "data→information" and "information→information" modules to enhance its own information extraction capabilities. Specifically, the training model first makes an internal summary before output, extracts the key sentences of the case, and then organizes a detailed analysis around these key points. At the same time, in the final expression, a general-first-detail structure is adopted: first give the conclusion and the main basis, and then provide further analysis as support (which can be placed in secondary paragraphs or folded information). In this way, professional readers can read the detailed analysis, and general readers can only read the previous summary to obtain effective information. Try to reduce repetition in expression to avoid being verbose. By improving these, DeepSeek's thoroughness will not sacrifice clarity.
5. Introducing patient factors and resource constraints: DeepSeek's current decision-making is idealistic and takes medical factors into consideration. Assessments of patient willingness, compliance, and resource conditions (an extension of the intent layer) should be added to be closer to clinical reality. For example, the model can be guided to think: "Is this examination convenient for patients to accept? What is the cost of this suggestion?" In this case, if the possibility of tuberculosis is extremely low and the patient has significantly improved, DeepSeek may be inclined not to actively arrange time-consuming and laborious examinations to reduce the burden on patients. This improvement requires adding weights to the intent layer, such as "minimum intervention intention" or "patient-centered intention", and balancing it with "never miss the diagnosis intention". DeepSeek learns to achieve a better solution between ensuring safety and avoiding excessiveness, rather than just being detailed.
Through the above improvements, DeepSeek is expected to retain its rigorous and comprehensive advantages while becoming more focused and efficient. In the future AI medical system, we can imagine combining the advantages of the two styles of models, ChatGPT and DeepSeek, to form an intelligent diagnosis and treatment assistant that understands clinical decision-making trade-offs without missing key details. This is not only meaningful for improving model performance, but will also play a positive role in the training of clinicians and the education of patients-the model output can be used for direct decision-making and as detailed reference material. In short, the analysis of the DIKWP*DIKWP model provides us with a tool for systematic insight into AI medical models. Using this framework to optimize models such as DeepSeek will help promote the reliable application of AI in the medical field and bring greater value to patients and medical staff.
Appendix
The DIKWP model (Data, Information, Knowledge, Wisdom, Purpose) proposed by Professor Yucong Duan was used to specifically analyze the root causes of the DeepSeek model's misdiagnosis and the potential deviation paths in the layers of transformation.
🎯 Analysis objectives:
Clarify DeepSeek when dealing with the following conditions:
"A 48-year-old male had a cough lasting about a week (worsened at night), yellow-green sticky sputum (cefaclor was effective), night sweats, itchy throat, and transient aphonia" and other symptoms;
Why would a case that highly fits the "recovery period of acute bronchitis" be mistakenly or biasedly judged as a case that "requires high vigilance for tuberculosis, postnasal drip, or even chronic diseases"?
1DIKWP five-level error tracing analysis
1.1D layer (Data) deviation: ✅ Advantages:
DeepSeek’s extraction of raw data is complete and sensitive: the timing of symptoms (nocturnal cough), drug reactions (effectiveness of antibiotics), special manifestations (night sweats), etc. are not missed. In a sense, it is even too faithful to treat all “abnormal data” as “requiring high processing”.
❌ Question:
There is a lack of proactive empowerment of “short-term, good response” to key data.
The data that strongly indicated a short-term recovery from bacterial infection, “antibiotic treatment was effective”, was not given sufficient weight within the model.
"Symptoms lasted only 1 week" - This data should have strongly supported acute disease rather than chronic disease or chronic infection (such as tuberculosis), but DeepSeek did not fully use this data to deny the legitimacy of long-term diseases.
🔎 Error summary:
DeepSeek does not clearly define "which data represent commonalities and which are anomalies" at the data level, but instead assumes that all "data that deviates slightly from the norm" are worth pushing up layer by layer, which lays the first step for misleading.
1.2I layer (Information layer) bias: ✅ Advantages:
DeepSeek excels at interpreting and expanding the clinical meaning of each piece of data (for example, "night sweats → may be a manifestation of tuberculosis, lymphoma, etc.").
The logic of integrating information such as “sputum → bacterial infection” and “cough at night → allergic” is reasonable, and multiple clinical concepts are cited.
❌ Question:
No dominant information framework has been formed, and there is a lack of a mechanism for filtering primary and secondary information.
The specific manifestations are:
Treat "night sweats" as highly abnormal information, rather than deconstructing or weighting it in contradiction with "short-term + antibiotics are effective";
"Cough worsens at night" may be caused by bronchial hyperresponsiveness, allergic cough, postnasal drip, reflux cough... DeepSeek does almost no risk grading and diagnostic priority sorting, but instead lays out all possibilities, resulting in generalized information and split reasoning.
🔎 Error summary:
The information layer lacks the ability of "comprehensive analysis", that is, it does not use "key positive information" (such as improvement, short course of disease, good treatment response) to suppress "non-specific information" (such as night sweats), resulting in deviation from the mainstream diagnostic path starting from the information layer.
1.3K-layer (Knowledge) bias: ✅ Advantages:
DeepSeek has mobilized a rich knowledge base, such as the typical manifestations and screening pathways of multiple diseases such as tuberculosis, CVA (cough variant asthma), GERD, and postnasal drip;
There is a strong tendency towards "exhaustive" knowledge coverage.
❌ Question:
There is a lack of a knowledge mechanism for “diagnostic suitability screening”, that is, there is no matching of the “most likely” rather than the “possible” among multiple pieces of knowledge.
This results in:
After the “tuberculosis” knowledge was activated, it triggered the decision to screen for tuberculosis;
The knowledge paths of “CVA”, “GERD”, and “postnasal drip” were also called one by one;
However, a very core point in basic clinical knowledge is overlooked: ✅ Acute course + improvement with antibiotics = highly likely to be bacterial bronchitis, excluding tuberculosis.
🔎 Error summary:
Knowledge mobilization does not have a hierarchical or consensus compression mechanism. Instead, it prioritizes the activation of the rare disease knowledge path of "high sensitivity and low specificity" before the "abnormal prompt signal", resulting in systematic diagnostic deviation.
1.4W layer (Wisdom) deviation: ✅ Advantages:
DeepSeek has a sense of responsibility and closed-loop decision-making for the diagnosis and treatment pathway, and often gives comprehensive recommendations including treatment, examination, and reexamination.
It reflects the design goal of the AI intelligent system to "provide comprehensive care for patients."
❌ Question:
DeepSeek’s intelligence layer lacks the real “subtraction wisdom” and “optimal decision-making” judgment capabilities:
For this case:
The optimal strategy should be: "Continue antibiotic observation, repeat examination if symptoms persist, and do not treat night sweats alone as a temporary reaction."
However, DeepSeek implemented a decision-making model of “multiple possibilities → multi-path parallel deployment”: it is recommended to consider multi-path parallel examinations such as postnasal drip, CVA, tuberculosis, and GERD (or at least be vigilant);
On the surface it is "wise and comprehensive", but in essence it is "not making judgments based on convergence of possibilities".
🔎 Error summary:
DeepSeek does not implement subtractive governance capabilities based on priorities and benefits at the intelligent layer, and lacks the main line of processing wisdom of "giving priority to those with the highest probability of first diagnosis, and closely following up on the rest."
1.5P-layer (Purpose, intention layer) bias: ✅ Advantages:
DeepSeek intends to maintain the caution of "not missing" serious diseases;
It tends to aim at "comprehensive coverage of all potential risks", reflecting the responsible ethics and bottom-line risk defense mechanism of the AI medical system.
❌ Question:
Imbalance at the intention level: Over-magnification of the goal of “never miss a diagnosis” suppresses the intention of “efficiency first, moderate medical care”.
DeepSeek does not reflect the balance principle of "resources-time-risk", but instead makes excessive deployment based on high-sensitivity intention settings;
For example: night sweats trigger attention to "tuberculosis and lymphoma" instead of first setting the intention to "focus the diagnostic direction on common, reversible, and well-treated acute diseases."
🔎 Error summary:
DeepSeek lacks clear target priority modeling (Primary vs Secondary Intention) in terms of intent, which results in the consequence that all possibilities "must be taken care of", leading to unfocused and even unnecessary interventions in diagnosis and treatment recommendations.
2Summary: DeepSeek misjudged DIKWP deviation path chain
Tiers |
Deviations in performance |
Specific contribution to miscarriage of justice |
D layer (data) |
Unauthorized "short-term + effective treatment" core data |
Misleadingly including minor abnormal data such as "night sweats" into the next level of analysis |
Layer I (Information) |
No priority order, generalize all interpretation paths |
All possibilities (TB, CVA, GERD, etc.) are activated in parallel |
K layer (knowledge) |
Comprehensive knowledge retrieval but no diagnostic "confidence" mechanism |
Non-mainstream knowledge paths (such as tuberculosis) were not correctly deprioritized |
W Tier (Wisdom) |
Decision-making is lengthy and conservative, with no ability to make subtractive judgments |
The inspection path is not focused and the suggestion system is over-expanded |
P layer (intention) |
The goal setting tends to be “no missed diagnosis” and ignores “optimal efficiency” |
This leads to "abnormal symptoms → expand all possibilities", ignoring the intention path of "focusing on common ones" |
3Summary: DeepSeek misjudgment is the result of “systematic caution”
From every transformation level of DIKWP, DeepSeek embodies an "exhaustive analysis strategy driven by responsibility ethics" rather than a goal-oriented optimal strategy generation model. Therefore, in the face of:
Conditions with slight "atypicality" in the data structure (e.g., night sweats);
However, the main manifestations are consistent with the scenario of common diseases (e.g., bacterial infection, cough + good response to treatment);
DeepSeek is more likely to move towards "misjudgment-induced over-caution", the essence of which is the weakening of "mainline intentions, mainline information, and mainline wisdom" on the DIKWP path, and the expansion of "secondary intentions, abnormal signals, and rare knowledge".
✅ Suggested improvement directions (brief):
Tiers |
Improvement Suggestions |
D-layer |
Introducing data weighting mechanisms (such as strong signal settings for treatment response and disease duration) |
I-layer |
Added "main symptom priority" information compression model to avoid parallel diffusion |
K-layer |
Establish a knowledge activation probability mechanism (consider confidence ranking when calling knowledge) |
W Tier |
Introducing the "Strategy Subtraction Module" at the smart layer to weigh benefits, risks and resources |
P layer |
Clearly define the main purpose (such as "focusing on common diseases and efficient treatment"), and only observe the secondary purpose |
