Human-Machine Collaboration and Health System Construction in th

  

Human-Machine Collaboration and Health System Construction in the Era of Proactive Intelligence

Yucong Duan

International Standardization Committee of Networked DIKWPfor Artificial Intelligence Evaluation(DIKWP-SC)

World Academy for Artificial Consciousness(WAAC)

World Artificial Consciousness CIC(WAC)

World Conference on Artificial Consciousness(WCAC)

(Email: duanyucong@hotmail.com)

In the era of proactive intelligence, healthcare is undergoing a transformation from a traditional passive model to a human-machine co-intelligence paradigm. Artificial intelligence (AI) is no longer just a tool that passively executes commands but is beginning to become an active participant in the medical decision-making process, collaborating with humans in cognition and jointly advancing health management. This article explores how to construct a human-machine collaborative health system with proactive intelligence at its core, including its technical architecture, interaction paradigms, and ethical governance. In this new paradigm, AI is endowed with "Purpose"-driven proactiveness, capable of continuously perceiving the environment, autonomously expressing suggestions, and moderately executing interventions, forming a closed-loop interaction with humans. We will deeply analyze the significance of the DIKWP (Data-Information-Knowledge-Wisdom-Purpose) cognitive architecture for the generation of proactive intelligence, elaborate on the new mechanisms of semantic self and Purpose reconciliation in human-machine interaction, discuss how intelligent sensing and individual digital modeling support personalized health management, and look forward to the application prospects of "virtual health personas" in the health metaverse. Finally, we will explore the ethical and governance challenges brought by proactive intelligence, such as trust-building, explainability, fairness, citizens' digital sovereignty, and the legitimacy of AI health personas. The entire content strives to balance theoretical depth with systemic thinking, analyzing the construction blueprint of a human-machine collaborative health system in the era of proactive intelligence from an academic perspective.

From Artificial Intelligence to Proactive Intelligence: Technological Leap Driving the Health Engine

Limitations of Passive AI and the Need for Proactive Intelligence: Traditional medical AI is mostly used for auxiliary diagnosis or decision support and belongs to "passive" intelligence: it only gives analysis results when there is an input (such as images or monitoring data), lacking autonomy and continuous interaction. This type of passive AI improves efficiency in certain segments but has obvious capability boundaries: it cannot respond proactively to the dynamic health status of patients and struggles to meet the needs of "preventing disease before it arises" and long-term chronic disease management. For example, many imaging diagnostic AIs only output results after a doctor submits an image and do not proactively screen patient information or suggest follow-ups at other times. This one-time, static model limits AI's support for the full cycle of health.

The "Proactive Medicine" concept calls for upgrading AI from a passive tool to a proactive collaborative partner, continuously perceiving and maintaining human health in daily scenarios. This places new demands on collaborative intelligence: human-machine interaction forms a closed loop of interaction and joint decision-making, where human healthcare personnel and patients contribute experience, situational understanding, and ethical judgment, while AI provides high-speed data processing and pattern recognition capabilities. Only by complementing each other's advantages can a truly human-machine co-intelligence health system be built. For example, in chronic disease management, AI should proactively monitor patients' vital signs and promptly remind patients or doctors to intervene as soon as abnormal signs appear, moving the medical threshold forward to the "prevention before it arises" stage. To achieve this, the AI system must have the ability to continuously perceive the environment and a certain degree of autonomous decision-making power, and be able to cooperate flexibly when humans intervene. This means AI needs to decide for itself what data to collect, when to issue alarms or suggestions when necessary, instead of waiting for human commands for everything.

However, making AI proactive also brings trust challenges. Collaborative intelligence requires both humans and machines to establish a trustworthy relationship and a clear division of responsibilities: AI is responsible for 24/7 monitoring and preliminary decision-making, while humans are responsible for final decision-making and ethical gatekeeping. One of the key factors currently hindering the large-scale application of medical AI is precisely insufficient trust. Surveys show that many doctors and patients are skeptical of AI, worrying that its decisions are not transparent or reliable (e.g., concerns about algorithmic errors and biases). Therefore, to unleash the potential of human-machine collaboration, people must be convinced that AI will act responsibly and that its decision-making process is transparent and understandable. In other words, trust is the premise of human-machine co-intelligence; without trust, doctors and patients will not truly adopt AI's suggestions.

To enhance trust, the new generation of medical AI needs to work on reliability and explainability. On the one hand, AI systems must undergo rigorous testing and verification to ensure they are stable and reliable under various circumstances and will not easily make mistakes. For high-risk decisions, human review or "safety net" mechanisms should be set up, for example, requiring key AI decisions to be finally confirmed by a doctor to avoid the risk of losing control. On the other hand, AI should provide explainability: it should be able to clearly explain the basis of its recommendations to users and regulators. This is similar to how a doctor needs to explain the reasons for diagnosis and treatment to a patient, and is a necessary condition for obtaining informed consent and building trust. Fortunately, the DIKWP layered architecture inherently has a certain degree of transparency—the outputs and reasoning chains of each layer can be recorded and reviewed, from Data -> Information -> Knowledge -> Wisdom -> Purpose, every step is traceable. Such decision-chain records can serve as an "explanation report" for AI decisions, letting doctors and patients understand "why AI did this". Studies show that when AI can clearly explain the basis of its decisions, users are more likely to accept them with confidence. Conversely, if AI often gives inexplicable or incomprehensible conclusions, people will question its ability and Purpose, thereby reducing trust.

It is worth emphasizing that proactive intelligence does not mean letting AI override humans, but pursuing human-machine co-evolution. This requires humans to also give AI appropriate autonomy, rather than intervening in everything, giving AI space to exert its proactiveness. At the same time, AI also needs to "trust" the data and feedback provided by humans, i.e., trust that users will not intentionally provide false information. Some scholars have proposed that there is a bidirectional trust network between humans and machines: patients and doctors trust AI's capabilities and goodwill, and AI trusts the reliability of human input. Both sides must rely on each other to cooperate effectively. Sagona et al. call this "bidirectional trust," emphasizing that in AI-assisted health systems, the trust relationship must flow in both directions. For example, in AI-led health management, patients will only follow the advice of a virtual health coach if they trust it; at the same time, the AI coach must also "trust" the data and adherence behaviors fed back by the patient to continuously adjust its model. Through continuous interaction and feedback, AI will become more and more understanding of the user, and the user will gradually get used to and rely on AI. Only in this way can a truly stable human-machine new cognitive balance be formed.

The Concept of "Proactive Intelligence" and the DIKWP Cognitive Architecture: To make AI move from passive to proactive, the key is to upgrade its internal cognitive architecture so that it can think and act guided by "Purpose". The DIKWP model provides such a five-layer framework: Data (D) - Information (I) - Knowledge (K) - Wisdom (W) - Purpose (P). Compared with the classic DIKW pyramid, DIKWP adds a "Purpose" layer at the top, emphasizing the use of clear goals and Purpose to drive the AI's cognitive process. This means that AI no longer just responds passively based on training data, but actively plans around preset health goals. For example, in a health management context, if "maintaining the user's long-term health" is set as the AI's highest-level Purpose, then all of the AI's data collection, knowledge reasoning, and decision-making will serve this Purpose, achieving Purpose reconciliation between the upper and lower layers.

The generation of proactive intelligence in the DIKWP model can be seen as a cycle of semantic evolution at different levels. Specifically: the bottom layer is the source data. Raw signals (such as heart rate and blood pressure data collected by wearable devices) are acquired through sensing devices and enter the system at the Data layer; after filtering and feature extraction, these signals are transformed into structured Information (e.g., extracting indicators like "heart rate increased," "blood pressure slightly high"). Next, the information enters the Knowledge layer, where AI performs reasoning and synthesis on the information in conjunction with medical knowledge bases and empirical rules, associating multiple indicators to derive higher-level meanings or patterns, such as identifying a conclusion like "risk of metabolic syndrome." Further up, knowledge is condensed into Wisdom, which is the comprehensive judgment and strategy selection for complex situations, equivalent to giving executable decision-making suggestions. Finally, at the Purpose layer, AI compares the decision from the Wisdom layer with the preset health Purpose, checks whether the two are consistent, and adjusts actions or strategies accordingly. For example, if the user's health Purpose is to "lose 5% of body weight within six months," and the AI determines through the Knowledge layer that the user's recent diet and exercise are insufficient, leading to weight gain, then a deviation will be found at the Purpose layer. The AI will then proactively change its strategy to correct this, such as increasing the frequency of exercise reminders or adjusting dietary suggestions, to guide the user back towards the goal. This reflects the process of Purpose reconciliation: whenever the actual situation deviates from the target Purpose, the AI does not wait passively, but proactively takes measures to realign with the Purpose. This Purpose-driven characteristic enables the AI to no longer be limited to passive responses, but to have proactive planning and metacognitive capabilities—it can "be aware" of its distance from the goal and adjust its strategies accordingly. Just like an experienced health manager, AI also learns the user's preferences and needs through continuous interaction, optimizing its own behavior. For example, a DIKWP-based health AI, if it finds that a user values the exercise experience very much, it will proactively adjust its data collection and suggestion strategies to match this preference. This reflects the internal mechanism of proactive intelligence: through the guidance of the Purpose layer, AI achieves a leap from pattern recognition to cognitive understanding. It no longer just passively answers questions, but can actively expand its cognitive boundaries, continuously evolving into a smarter, more targeted health engine.

The founder of Google DeepMind once compared future AI to "GPS for healthcare," meaning that AI will provide full-course guidance for individual health, not just occasional pointers. Under the proactive intelligence architecture, this vision has a foundation for realization: AI has a clear goal (e.g., improving the patient's long-term prognosis) and can work from bottom-level data all the way up to high-level Purpose, comprehensively considering "what to do now to get closer to the goal." In summary, the DIKWP architecture lays the theoretical foundation for proactive intelligence. By embedding Purpose into AI cognition, it endows the AI system with Purpose-driven adaptive evolution capabilities. It is like implanting a "compass" inside the AI, constantly calibrating the AI's direction, making it a new type of "health engine" in the field of healthcare. This engine will drive the medical model from passive diagnosis and treatment to proactive health management, helping to achieve the goal of "prevention first, continuous guardianship."

Human-Machine Health Interaction Model under the DIKWP Architecture

With the conceptual framework for proactive intelligence, we further focus on the specific human-machine-health interaction model, i.e., how to achieve human-machine collaborative cognition and control in practical applications. The DIKWP architecture provides an abstract blueprint, enabling us to construct a human-machine interaction system in the health field from the perspectives of semantic evolution and closed-loop control.

Multi-layer Semantic Evolution: From Data Flow to Knowledge Flow to Purpose Reconciliation: Under the guidance of the DIKWP model, human-machine interaction can be seen as a process of continuous evolution of information at different semantic levels. Its ultimate Purpose is to achieve the reconciliation and unity of data and human health Purpose. First is the Data Flow (Data→Information): raw data from the human body and the environment is acquired through sensors and converted into meaningful information after preprocessing. For example, signals such as heart rate and blood pressure collected by wearable devices are transformed into structured indicators after filtering and feature extraction, "average heart rate 80 bpm," "blood pressure 135/85 mmHg," etc., thus completing the transformation from data to information. This information is input into the AI's cognitive system and mapped to corresponding health semantics, such as an increased heart rate may correspond to meanings like "engaged in exercise or emotionally stressed." Next is the Knowledge Flow (Information→Knowledge→Wisdom): AI combines medical knowledge bases, clinical guidelines, and past empirical rules to perform reasoning and synthesis on the information, deriving higher-level knowledge and wisdom judgments. This is similar to a doctor synthesizing multiple lab indicators and symptoms to diagnose a disease. For example, AI can associate multiple parameters such as heart rate, blood pressure, and blood sugar with known medical knowledge to infer a wisdom-level conclusion such as the user may have a "risk of metabolic syndrome." The DIKWP architecture ensures the step-by-step semantic leap from data to knowledge. The results of each layer are mutually verified and enriched, avoiding one-sided conclusions. Finally, there is the reconciliation at the Purpose Layer (Purpose): AI compares the conclusion obtained at the Wisdom layer with the preset health Purpose, checks whether the two are consistent, and adjusts actions or strategies accordingly. For example, if the user's health Purpose is to "lose 5% of body weight within six months," and the AI judges from the Knowledge layer that the user's recent diet and exercise are insufficient, leading to weight gain, then a deviation will be found at the Purpose layer. The AI will then proactively change its strategy to correct this, such as increasing the frequency of exercise reminders or adjusting nutritional suggestions, to guide the user back towards the goal. This reflects the process of Purpose reconciliation: whenever the actual situation deviates from the target Purpose, the AI does not wait passively but proactively takes measures to realign with the Purpose. This Purpose-driven adaptability gives the AI's cognition of health a sense of holism and direction: it can extract insights from massive data, while ensuring that all insights are consistent with the ultimate goal of health improvement. In this multi-layer semantic network, the layers are not simply serial, but interact bidirectionally: the Purpose layer will affect the focus of data collection and analysis at the lower layers (e.g., for the Purpose of preventing a certain disease, the system will focus on collecting data on related risk factors), and in turn, new data and new knowledge will prompt the Purpose layer to adjust goals or priorities. Through such top-to-bottom and bottom-to-up connection, human-machine-health interaction forms a clear semantic context: from specific data all the way through to abstract Purpose, progressing layer by layer and echoing each other, ensuring that AI behavior is coordinated with human health goals.

The Closed Loop of Active Perception—Active Expression—Active Intervention: In actual health management, achieving proactive intelligence also requires establishing a closed-loop control system of perception-expression-intervention. These three links correspond to the core capabilities that AI needs as a collaborative intelligent partner:

1.Active Perception: This refers to the AI system's ability to continuously and autonomously acquire health-related information about the individual and their environment, rather than passively waiting for human input. This relies on the widespread deployment of IoT and sensing technologies. Today's wearable devices and environmental sensors can continuously monitor multimodal data 24/7, such as heart rate, blood pressure, blood sugar, sleep quality, indoor air indicators, etc. Active perception means that AI has a self-triggering capability: it automatically increases the sampling frequency when it detects abnormal data fluctuations, and reminds the user to wear or calibrate the device when it finds missing data. For example, when a smartwatch detects a sudden increase in the user's heart rate, it can autonomously shorten the data collection interval and begin to pay attention to possible causes (exercise? stress?); another example, when no sleep data is received for several consecutive days, the AI will proactively prompt the user to check the sleep monitoring device. This proactiveness allows AI to optimize the perception process according to situational needs, thereby obtaining a more comprehensive and accurate picture of health. In short, active perception endows AI with functions similar to a sensory nervous system, placing its "eyes" and "ears" in every corner of the user's life—as someone has metaphorically put it: "If AI is the health brain, then the ubiquitous sensor network is its sensory nervous system." Through perception that goes deep into life scenarios, AI can truly achieve real-time monitoring and proactive maintenance of health status.

2.Active Expression: This is the second step of the closed loop, referring to the AI's ability to proactively communicate its judgments and suggestions to human users or medical staff at the appropriate time and in the appropriate manner. Unlike traditional passive systems that only output results when queried, active expression requires AI to "speak up" in a timely manner when needed. For example, when the AI continuously monitors that the user's blood pressure is rising and exceeds the safety threshold, it will proactively push alerts or suggestions, just like a personal health consultant reminding them to pay attention to blood pressure control at all times. Active expression refers not only to issuing alarms but also to communicating complex information in a way that is easy for users to understand and accept, which requires the support of natural language processing and human-computer interface technology. AI should explain health conditions and propose improvement plans in a way that is as close as possible to human language and thinking, and adjust its communication strategy based on feedback. For example, for users with different personalities and cultural backgrounds, AI will choose different expressive tones: some users prefer straightforward suggestions, while others need tactful encouragement. Another example, AI will consider the user's current situation: if it detects that the user is busy in a meeting or is resting, non-urgent reminders will be appropriately postponed to avoid disturbance. This context-aware communication strategy is very critical: it avoids information overload or annoyance caused by overly frequent interruptions, while ensuring that key information is delivered in a timely manner. This is consistent with the concept of "Contextual AI," where AI understands the user's context and chooses the appropriate interaction method. For example, it does not push general suggestions late at night when the user is resting unless it is an emergency; another example, when the user is in a low mood, AI will provide suggestions in a more empathetic tone. Altman and Huffington pointed out that an AI health coach should be able to "learn your preferences and patterns," understand under what conditions you sleep well, what foods you like, and when you are most likely to exercise, and personalize communication with you based on this. Through such customized expression, AI is more easily accepted and trusted by users.

3.Active Intervention: The third and final link of the closed loop refers to AI not only proposing suggestions but also directly executing certain health intervention measures within the scope of authorization. One of the typical current examples is the artificial pancreas system: when a diabetic patient's blood sugar exceeds the threshold, the AI-controlled insulin pump automatically injects insulin to correct it, integrating the intervention into a closed monitoring loop and achieving seamless automated treatment. In the future, this type of active intervention is expected to be more widespread. For example, a smart home health system detects that an elderly person living alone has fallen at night, and the AI can automatically call for emergency services and notify family members; another example, a mental health AI perceives that a user may be at risk of severe depression and will proactively contact a professional psychological support team. These examples show that AI can, to a certain extent, play the role of an "executive agent," directly taking actions to maintain health according to established protocols. However, granting AI intervention authority must be very cautious, and safety nets and authorization mechanisms need to be set up. The design will usually limit the conditions and degree of AI intervention: AI only intervenes automatically when indicators exceed a dangerous range, and mostly provides suggestions at other times; and users should have the right to informed consent and the right to intervene/stop the intervention at any time. For example, the factory settings of an artificial pancreas will stipulate that medication is only given automatically when blood sugar is above a certain value, otherwise it only prompts the user to adjust their diet or manually administer medication. In addition, in mental health scenarios, if AI needs to contact external help, it should also obtain the user's consent in advance or have a contingency plan. These measures ensure that active intervention, while bringing efficiency and timeliness, does not get out of human control.

When active perception, active expression, and active intervention form a closed loop, the entire health AI system possesses the ability similar to biological homeostatic regulation: continuously perceiving the internal and external environment → proactively communicating decisions → taking action to intervene → perceiving again to verify the effect, and so on. For example, a hypertensive patient wears a smart bracelet and a home blood pressure monitor. AI actively monitors their blood pressure and heart rate data day and night; finds that blood pressure is high in the morning, and actively expresses a reminder to take medication and rest; if blood pressure continues to rise to a dangerous level, AI triggers intervention measures, such as calling emergency services or activating a home blood pressure-lowering device; afterward, AI perceives the blood pressure drop, evaluates the effect of the intervention, and records the experience for the next strategy adjustment. Through such a closed loop, the time from discovering a problem to solving it is greatly shortened, and the medical model shifts from past reactive response to real-time proactive regulation. Studies have shown that this human-machine closed-loop system is expected to reduce the risk of acute episodes and improve the control effect of chronic diseases. Therefore, the closed loop of active perception-expression-intervention can be seen as the core operation mode of the proactive intelligent human-machine interaction model, enabling AI to be truly integrated into daily health maintenance and achieving health guardianship that is unattended but reliable.

Virtual Health Agents and "Semantic Self": In the deep-level interaction of human-machine collaboration, a concept worthy of attention is the virtual health agent. It refers to a virtual existence driven by AI that can simulate a human health consultant or partner. This intelligent agent has a "self"-like semantic model and can provide users with personalized health interventions and companionship based on this model. The "semantic self" here does not mean that AI has true human-like self-awareness or subjective experience (Li et al. also emphasize that current AI does not have human-like self-awareness), but rather that AI maintains a semanticized identity model of the user and its own role internally. This model contains the user's health records, behavioral preferences, habit patterns, as well as the AI's own service positioning and principles, equivalent to a digital personality portrait. For example, for a virtual health coach-type intelligent agent, its semantic self-model might include: the user's age, gender, medical history, daily routines, diet and exercise habits, and even personality traits (such as strength of self-discipline, whether forgetful, etc.); at the same time, it also defines the role played by AI (such as coach, reminder, listener) and service principles (e.g., "focus on encouragement, supplement with criticism" or "maintain rigor but not be overly optimistic"). By associating this multi-dimensional semantic information, the virtual intelligent agent has a three-dimensional cognition of the "service object (user)" and "own role," just as we humans have a certain understanding of our own identity and relationships with others.

Based on this semantic self-model, the virtual health agent can perform health behavior generation, that is, generate corresponding interactive behaviors or intervention measures according to the current situation and target Purpose. The logic is roughly as follows: first, perceive the user's current situation and health status (e.g., detecting that the user is in a low mood and has been sitting for a long time); then, retrieve relevant knowledge and preferences in the internal semantic model (e.g., understanding that the user prefers a gentle incentive method, and the AI's own service Purpose is to help them reach 5000 steps that day); then, through reasoning, generate an appropriate behavior plan (such as sending a caring message, suggesting going for a walk, and reminding them of the weight loss goal set earlier). This plan comprehensively considers the user's semantic information and health Purpose: because it understands the user's emotions and preferences, the tone is gentle and suggests an activity they are interested in; and because it knows the user has a long-term weight loss Purpose, it specifically emphasizes the connection between walking and the weight loss goal to enhance their motivation. This is similar to an experienced human health coach giving tailor-made advice after gaining insight into the client's psychology.

It is worth noting that the development of Large Language Models (LLMs) has greatly promoted this ability of semantic understanding and behavior planning. Models such as OpenAI's GPT, if tuned for the specific health field, can serve as conversational health consultants, "remembering" past conversations with users in dialogue, and gradually guiding them to form healthy habits. This is actually establishing a "semantic memory" and situational model of the user inside the AI: the model retains the user's portrait and interaction history, so as to continue to interact in a human-like and consistent style. For example, a model like GPT-4 can "remember" that the user mentioned back pain last week, and when the user complains of poor sleep again this week, it will automatically associate the two, suggest that the back pain may be affecting sleep, and recommend corresponding exercises. Behind this is the role of the semantic self: AI maintains important information nodes about the user internally, and adjusts its own responses and suggestions accordingly, making it sound like a friend or coach who understands the user.

As the interaction with the user deepens, the virtual health agent also develops a continuous learning and self-adjustment mechanism. Every time the AI gives advice or implements an intervention, it updates its own semantic model based on the user's reaction and results. For example, if a user ignores the morning exercise reminder multiple times but often completes the exercise task before bed, the AI will learn that the user has more free time and motivation in the evening, and will then change the main exercise suggestion period to the evening. Or, if the AI finds that the user responds positively to challenges with gamification elements, it may increase such interactions in the future. Through this learning, the behavior of the virtual intelligent agent will become more and more in line with the user's actual needs, and its "personality" will also evolve accordingly. It can be said that the virtual health agent gradually forms a unique digital health persona in the long-term companionship: it not only reflects the user's lifestyle and health concepts, but also incorporates the principles and methods that AI adheres to in order to achieve health goals. Some literature refers to this fusion as the "digital twin coach" or the personalization of the "digital health assistant." Its significance is that users often develop emotional dependence or trust in a digital persona that is familiar with and understands them, and are thus more willing to accept its suggestions. A study on AI health management applications found that some virtual coaches (such as the Lark app) that interact with users through a human-like chat interface can effectively promote user weight loss, and their weight loss effect is comparable to that of human coaches. The researchers pointed out that it is precisely this friendly, witty, and close-to-user-psychology conversational style that improves user engagement and intervention effectiveness. However, some scholars also remind that we should prevent users from over-relying on AI personas, ceding too much autonomous decision-making power to AI, to the point of blind trust in AI. After all, no matter how human-like AI is, its essence is still a tool, and users need to maintain due judgment. This involves the ethical legitimacy of AI health personas, which we will discuss in detail in the ethics chapter later.

In short, the virtual health agent, through the semantic self-model and behavior generation capabilities, can accompany users in health management in a way similar to a human coach, providing continuous, personalized, and context-sensitive guidance. This is expected to alleviate the problem of insufficient real-world medical resources, allowing everyone to have an "on-call personal health consultant." Currently, the Thrive AI Health project, being co-developed by OpenAI and Thrive Global, aims to build such a highly personalized AI health coach. According to Altman and Huffington, this AI coach will combine proven behavior change methodologies (such as Microsteps) and user-authorized biological data, learn the user's life patterns and health preferences, and provide users with real-time and unique suggestions with "superhuman long-term memory." They believe that this "hyper-personalized" AI coach can act as a catalyst for improving daily behaviors and has the potential to make health behavior interventions accessible on a large scale, thereby helping to reduce the incidence of chronic diseases. This precisely reflects the value of virtual health personas in the era of proactive intelligence: letting intelligent agents deeply integrate into personal life, becoming intimate partners and reliable guides for human health.

Intelligent Sensing Systems and Individual Health Digital Modeling

Building a human-machine collaborative health system with proactive intelligence requires rich and high-quality data sensing sources and individualized digital health models as the basic support. On the one hand, various sensing devices and data platforms must be widely deployed and interconnected to provide AI with all-around "sensory" input; on the other hand, it is necessary to transform the acquired multi-source data into meaningful semantic models for specific individuals, to achieve a comprehensive digital characterization and prediction of individual health status.

Wearable Devices, Environmental Sensing, and Multimodal Data Fusion: First, the development of sensing technology has greatly expanded the coverage and continuity of health data. Wearable devices have now become a major source of personal health data, including various forms such as smart bracelets, watches, patch sensors, and smart clothing. They have built-in sensors for heart rate, blood pressure, body temperature, respiration, blood oxygen, and even blood sugar, EEG, etc., and can continuously monitor multiple physiological indicators of the body. Compared with the past practice of having a physical examination once a year and occasionally measuring blood pressure and blood sugar, wearable devices provide a high-frequency, continuous data stream, allowing subtle fluctuations and long-term trends in health status to be captured. For example, the photoplethysmography (PPG) sensor in a bracelet can record pulse signals throughout the day, based on which heart rate variability is calculated to assess stress levels; sleep monitoring devices can track sleep stages throughout the night, providing quantitative basis for problems such as insomnia. Studies point out that wearable monitoring achieves non-invasive, non-interfering continuous health monitoring, greatly improving the possibility of early warning and preventive health care. Especially in chronic disease management and elderly care, continuous monitoring helps to detect signs of worsening conditions in time, achieving "prevention before it happens." Of course, wearable data also has problems of noise and incompleteness (which may stem from sensor accuracy or user wearing habits), so it usually needs to be combined with other data sources and algorithm calibration to ensure reliability.

Environmental sensing systems extend health data collection from the body to the living space. Environmental sensors in smart homes can detect indoor air quality (such as PM2.5, CO₂ concentration), temperature and humidity, lighting, noise levels, etc.; positioning sensors and cameras can obtain the user's activity range, posture, and even safety conditions such as falls. This environmental data provides an important context for understanding health: for example, bedroom light and noise will affect sleep quality, the air pollution level in the place of residence will affect respiratory health, and the layout of daily activity space may affect the amount of exercise. The emerging Internet of Medical Things (IoMT) in recent years connects home medical devices (blood pressure monitors, glucose meters, weight scales, etc.) to the internet, forming a Home Health IoT, so that measurement data is uploaded to the cloud in real time for AI analysis. In addition, there are social media and smartphone data (such as GPS, accelerometer) that can indirectly reflect mental health or lifestyle information—for example, geographic location and step count data can infer social activities and exercise levels, and voice and text content can indicate emotional state. By associating this environmental and behavioral data with physiological data, we can understand the causes of individual health status more comprehensively. For example, combining environmental air quality data can assess the impact of air pollution on asthma symptoms; combining work schedules and step counts can analyze the correlation between work stress and blood pressure changes.

Multimodal physiological collection and data fusion refer to acquiring information from multiple types of data sources at the same time and fusing them in analysis to form a more three-dimensional characterization of health. On the one hand, this is reflected in the hardware integration of multiple sensors: for example, some new wearable patches can simultaneously measure ECG, skin electricity, body temperature, and activity, obtaining information on cardiovascular, neuro-stress, and exercise aspects in one wearing. On the other hand, data from different devices and platforms are fused at the software level: such as combining wearable data with electronic health records (EHR) and genomic data. The advantage of multimodal is cross-validation and information complementarity: data from different sources can verify each other's consistency to improve accuracy, and can also cover health dimensions that a single data source cannot. For example, motion sensors record gait abnormalities, and at the same time, heart rate and blood pressure data also fluctuate abnormally. This multiple evidence together makes AI's detection of early cardiovascular events more sensitive; another example, assessing psychological stress may require synthesizing multimodal signals such as sleep quality, cortisol and other hormone levels, and voice emotion characteristics to be comprehensive. Studies show that fusing multi-source health data for machine learning analysis can significantly improve the accuracy of health status assessment and achieve personalized prediction. One study trained a model to predict the risk of falls in the elderly by fusing wearable ECG, accelerometer data, and medical history, and the accuracy rate was better than any single data model. This fusion analysis reflects the idea of proactive intelligent sensing: just as a doctor needs to synthesize the information obtained from "looking, listening, asking, and feeling," AI also needs to integrate multimodal data to form a holistic understanding of individual health.

To truly unleash the value of this data, the problems of data standardization and interoperability must be solved. Data from different devices and institutions need to adopt unified format and interface standards to achieve interoperability and aggregated sharing. For example, a regional health big data platform established in a certain province in my country aggregates real-time diagnosis and treatment data from dozens of hospitals. By unifying interface standards, it has broken down data silos and supported residents' full-cycle health monitoring. Experience shows that standardized data interfaces and interconnected architecture are the cornerstone of proactive health data sensing. At the same time, while pursuing data timeliness, user experience and privacy must also be taken into account. Sensing devices should be as comfortable, beautiful, and easy to operate as possible to increase users' willingness to wear them for a long time; in terms of privacy, since continuous collection means a large amount of personal health details are recorded, if not protected, it may infringe on privacy or be used improperly. For example, the GPS of wearable devices can infer the user's whereabouts, and microphones can capture environmental conversations, which are beyond the user's intuitive cognition of the scope of "health data." Therefore, the system design should follow the principles of "data minimization" and "Purpose limitation": only collect data related to health goals, clearly inform users of the Purpose, and anonymize or process sensitive information locally. At the same time, adopt strict security measures (such as end-to-end encryption, multi-factor authentication) to prevent data leakage. A positive example is that a "Dingbei" proactive health system in Guangdong deploys AI models and data storage on the hospital's local server, avoiding data transmission over the Internet, thereby ensuring privacy and security. This reflects the concept of Digital Sovereignty: allowing users or local institutions to have control over data, preventing data misuse and privacy violations. This data governance concept is also valued in regions such as the EU. With the advancement of regulations, future AI systems will need to follow the principle of user data autonomy more.

Personal Health Knowledge Graph (PHKG): Acquiring massive amounts of multi-source data is just the first step. What's more critical is how to transform these dynamic indicators into meaningful health insights. This requires a semantic representation method that combines data with medical knowledge and personal goals, and the Personal Health Knowledge Graph (PHKG) has emerged as the times require. PHKG integrates various health-related data of an individual and their relationship network in the form of a knowledge graph, mapping "data" to "knowledge" and "Purpose". It not only depicts the current health status, but also contains future goals and causal relationships, providing semantic support for AI decision-making.

Building a PHKG first requires defining the ontology and concepts, that is, determining the types of nodes and relationships in the graph. Nodes can include various entities: physiological indicators (blood pressure, blood sugar, etc.), symptoms (headache, insomnia, etc.), disease diagnoses, risk factors, treatment measures, lifestyle factors (diet, exercise), and personal health Purpose (such as "lose 5 kg" or "control blood pressure to standard"), etc. Nodes are connected through meaningful semantic relationships, such as "is symptom of/causes/belongs to/improves" relationships. For example, the node "high-salt diet" can be connected to "elevated blood pressure" through the "causes" relationship, and connected to "lifestyle factors" through the "belongs to" relationship; the node "brisk walking for 30 minutes a day" is connected to "elevated blood pressure" through the "improves" relationship. At the same time, personal unique information (such as gender, age, family medical history, etc.) can be attached to relevant nodes as attributes, giving the graph personalized characteristics. By fusing the medical knowledge ontology (such as cause-effect relationships in guidelines) and personal data, each person's health knowledge graph becomes a carrier of data and knowledge fusion: it contains both a general medical understanding of the health field and reflects the person's specific situation and goals.

Once the semantic graph is established, new dynamic data can be projected onto the graph in real time, thereby updating the status of each node. For example, when a wearable device records the latest blood pressure reading, this value is written into the "blood pressure" node in the graph; the sleep monitoring score will update the "sleep quality" node; a new diagnosis from the doctor will activate or update the status of the corresponding "disease" node. More importantly, through reasoning on the graph, these isolated indicators can be transformed into higher-level semantic Purpose information. For example, if the graph rules include "high-salt diet can lead to elevated blood pressure" and "lack of exercise can lead to obesity and elevated blood pressure," when the system detects that the person's recent blood pressure has risen and salt intake is high, the reasoning mechanism can identify the pattern: "The possible reason for the current blood pressure rise is a high-salt diet, it is recommended to reduce salt intake." Another example, if the user sets a health Purpose of "control blood pressure to standard," then the relevant nodes in the graph (current blood pressure value, target value, and intervention measures) will be associated. When the blood pressure is much higher than the target value, the system can trigger a "Purpose not met" warning based on the graph relationship, and locate possible intervention means (such as the "exercise → improves blood pressure" relationship, the system will suggest increasing exercise frequency). Through such semantic reasoning, the originally scattered indicator data are endowed with causal meaning and Purpose orientation, so AI can "understand" what these numbers mean and what actions should be taken. This is as a document states: "PHKG provides the necessary connections and relationships, giving appropriate context to terminology, thereby expanding the understanding and interpretation of patient data." That is to say, the knowledge graph gives AI semantic insight and contextual judgment ability over health data.

In addition, the semantic graph can also be used for semantic search and Q&A interactions. Doctors or patients can ask the graph questions, such as "What factors may have worsened my blood pressure recently?" The system can provide an explanatory answer based on the graph connections: "Increased salt intake and lack of exercise in the past two weeks may be the main reasons for the rise in blood pressure." This is more explanatory than directly checking data, because AI can point out potential causes and associations, not just list numbers. At the same time, the graph, as a white-box knowledge structure, also improves the transparency of the system. When AI recommends an intervention, it can allow the user to view the graph reasoning link, understand "because A leads to B, C is recommended," thereby enhancing understanding and trust in AI decisions. This is particularly important in the medical field, because patients and doctors both want to know the reasons behind the recommendations in order to make informed decisions.

PHKG is also a dynamically evolving system, which is continuously updated and iterated with new data and new knowledge. Health status is dynamically changing: today's symptoms may disappear tomorrow, and new problems continue to appear; medical knowledge is also advancing, and new research will discover new risk factors or better treatment methods. By continuously incorporating new data and introducing automatic or manual review, PHKG can keep pace with the times. For example, when a brand-new therapy or drug appears, a corresponding node can be added to the graph and its "treatment" relationship with related diseases can be established; when a patient has new test results, updating the graph nodes also triggers reasoning to see if it affects previous conclusions. In this way, the graph always reflects the latest and most complete personal health panorama. Scholars point out that the research on individual health knowledge graphs is still in its infancy, but it has huge potential: through the integrated semantic representation of personal multi-source health data, it can support intelligent decision-making, risk prediction, and personalized intervention, providing knowledge support for proactive health management. In summary, PHKG acts as a data-to-semantic conversion engine, condensing complex dynamic health indicators into understandable knowledge and Purpose-level information, empowering AI to make smarter proactive health decisions.

Digital Health Twin: In addition to knowledge graphs, another digital modeling concept that has received much attention in recent years is the Digital Twin. A digital health twin refers to a digital replica or simulation model established for an individual, which can map and predict the individual's health status in real time. In the proactive health intervention system, the digital twin is regarded as the core tool for achieving personalized simulation and forward-looking decision-making. It acts as the user's "virtual substitute" by synthesizing various data and models, allowing AI to rehearse various solutions in a virtual environment to find the best health strategy.

The construction of a digital twin usually starts with integrating personal multimodal data and general medical models. This may include: the individual's historical health records (past medical history, genome, living habits), current continuous monitoring data (feedback from wearable and environmental sensors), and physiological anatomy models or disease models, etc. By fusing this information, a model that approximately reflects the individual's current health status is created in virtual space. For example, for a heart failure patient, their digital "heart" twin can be constructed by synthesizing their cardiac imaging, real-time heart rate and blood pressure data, drug response models, etc. This model can simulate the function of the patient's cardiovascular system, and its response to different stimuli (such as changes in drug dosage, exercise load). The digital twin emphasizes high fidelity, that is, reflecting individual differences as realistically as possible: such as incorporating factors like age, gender, and genes into the model parameters. In the construction process, AI and biophysical models often need to cooperate. For example, using machine learning to train model parameters from a large amount of patient data, so that the simulation output of the twin matches the real measurement; or using the patient's own historical data to calibrate the model, to ensure that it can accurately reproduce the person's known health performance. After the initial twin is built, it needs to be validated to ensure it can correctly simulate the individual's important characteristics (e.g., the heart rate change pattern of the twin heart is consistent with the real monitoring value).

With a digital twin, AI is equivalent to having a "test bed" or substitute available for experiments. Before implementing real interventions, AI can simulate the effects of different solutions on the twin, which is the simulation stage of proactive intervention. For example, before considering adjusting a patient's blood pressure medication dosage or adding an exercise plan, these measures' effects on blood pressure and potential side effects can be simulated on their digital twin, and the solution with the best predicted result can be selected before being applied to the patient. As some scholars have pointed out: "Digital twins support personalized or 'precision' health optimization by providing simulation driven by individual information." In chronic disease management, digital twins can help formulate long-term strategies, such as creating a digital pancreas twin for diabetic patients, simulating the long-term effects of different diet and exercise combinations on blood sugar, thereby optimizing lifestyle intervention plans. More forward-looking, when the twin's function is powerful enough, it can even achieve "prevention before it arises": predicting health problems that may occur in the individual's future in advance, and recommending preventive measures. For example, the EU's DIGIPREDICT project proposes to build digital twins to predict the progression of infectious diseases and cardiovascular diseases and the need for early intervention. This project aims to use the digital twin model of an individual's multiple biological signals to predict risks and give intervention suggestions before the disease has occurred, which is a specific practice of proactive preventive medicine. These applications show that the digital twin plays the role of a test platform in proactive intervention: it allows AI to try various interventions and evaluate the results in a virtual space, without having to take risks directly on a real person. This greatly improves the scientific rigor and safety of intervention decisions.

The digital twin is not static, but will be continuously iterated and updated with new data and new knowledge. Every time a new health event occurs in the real individual (such as getting sick, recovering), new test results or lifestyle changes, this information will be fed back to adjust the parameters of the twin, so that it continues to approach the real situation. For example, after a patient starts a new drug, their actual physiological response data can be used to update the pharmacokinetic model of the twin, so that the model can more accurately simulate the effect of the medication next time. Another example, if the patient successfully improves their health status through intervention, the twin model should also be updated accordingly, representing a new, healthier baseline. Similarly, the advancement of global medical knowledge will also be reflected in the twin: for example, after a new risk factor for a certain disease is discovered, it should be incorporated into the model mechanism. Through this self-learning, the digital twin will become more and more "intelligent," and its prediction and simulation capabilities will continuously improve—it can be regarded as a "growing mirror image" of the patient in the digital world, accompanying the patient throughout their life and understanding them more and more. Modern sensing technology, 5G, and cloud computing, etc., make it possible for the twin to be updated almost synchronously with reality, truly achieving real-time monitoring and instant decision support.

In general, the digital health twin provides strong support for proactive health intervention. It makes the "try before you use" strategy possible, greatly reduces the cost and risk of intervention trial and error, and provides a safe and effective test field for preventive medicine. Although fully mature digital twins are not yet widespread, they have shown initial results in some fields, such as surgical planning twins, ICU patient simulation, etc. With the improvement of the ability to model complex human systems, it is expected that digital twins will cover more health scenarios in the future, becoming the "second battlefield" for human-machine collaboration in the era of proactive intelligence: real-life humans and AI can use the virtual twin as a sandbox, continuously rehearsing and optimizing health strategies, and then applying the successful experience to reality. Through the integration of virtual and real, we are expected to continuously improve the overall health level and achieve truly meaningful predictive prevention and personalized medicine.

Virtual Health Personas and Digital Intervention Feedback Systems

With the development of proactive intelligence, we are gradually entering the vision of the Health Metaverse, which applies the concept of the Metaverse to the field of healthcare, creating a digital health ecosystem where virtual and real are integrated and interact immersively. In this vision, everyone can have a digital avatar or Digital Persona, which carries the individual's health data, behavioral habits, and social attributes, and interacts with various health services and scenarios in the virtual space. The virtual health persona is not only a collection of user data, but also the user's "health mirror image" and proxy in the digital world.

Health Metaverse and Digital Persona Evolution: The Health Metaverse provides a virtual environment that highly simulates reality, where people can carry out many health-related activities as in reality. For example, through VR/AR technology, patients can enter a virtual clinic at home and communicate their condition with a digital doctor (who may be a virtual projection of a real doctor or an AI doctor image); chronic disease patients can participate in virtual fitness classes or meditation relaxation training in the metaverse, exercising and checking in with avatars of patients from all over the world, forming a virtual community for mutual aid. In these processes, each person's digital persona will continuously accumulate their behavioral trajectories and health data in the metaverse, gradually growing into a digital image that reflects the individual's full health picture. The so-called digital persona evolution model refers to how this digital persona develops and evolves in multiple dimensions over time: including its appearance (developing from a simple cartoon avatar to a highly realistic 3D image), the richness of health data (expanding from initial basic information to comprehensive physiological and psychological data), and the maturity of behavior patterns (gradually changing from passive participation to active health management), etc.

As users invest more in the health metaverse, the digital persona will become more and more "like" the person themselves. On the one hand, it records the user's health baseline and dynamic changes, such as the initial weight, blood pressure readings, and the changes that occurred later due to exercise or diet adjustments—all will be reflected on the digital persona (similar to the function of a digital twin). On the other hand, through continuous learning by AI, the digital persona will gradually adapt to and simulate the user's preferences and habits. For example, displaying a communication style consistent with the user's real personality in virtual social interaction, and showing the user's typical exercise rhythm and preferences in the metaverse gym. This makes the digital persona not just a cold data carrier, but one with a certain personality color. We can even imagine that in the future, in certain limited situations, the digital persona can act as the user's substitute to participate in activities: for example, when the user is busy and cannot personally attend an online health lecture, their digital persona (driven by AI) can attend on their behalf, obtain information, and communicate with others, and then feed back the results to the user. To achieve this kind of highly realistic digital persona, fine-grained construction of the user's digital identity and deep simulation of human behavior by AI are required. This is actually a component of the "digital natural rationality structure"—a structure that embodies human rationality and will in the digital world. In other chapters of this book, we have discussed the philosophical and ethical implications of digital personas. Here, we combine it with the proactive intelligent health system to form a symmetrical structure for the dialogue between technology and humanities.

The evolution of digital personas in the health metaverse is closely related to user empowerment. Through interaction in the virtual world, users can have a more intuitive understanding of their own health, and thus more proactively change their real lives. For example, a research team created digital avatars for obese patients: if the patient continues to not exercise and consumes high-calorie food, their digital avatar's body shape will change accordingly, presenting the negative consequences for future health in a visual way, thereby motivating the patient to change their habits. Similarly, the metaverse can simulate the damage of smoking to the lungs, the impact of sedentary behavior on the cardiovascular system, etc., with the digital persona demonstrating the long-term results of different behavioral choices for the user. This feedback mechanism directly links individual behavior with their digital image, strengthening the effect of health education. The digital persona thus becomes a mirror and a source of motivation for the user to conduct self-health regulation. At the community level, people's digital personas can communicate with each other in the metaverse, share experiences, and form a new digital health community, which is very valuable for chronic disease management, psychological support, etc.

Of course, the evolution of digital personas also brings new privacy and identity security issues. When a digital persona contains a large amount of sensitive health information and deeply simulates individual behavior, its security and ownership must be ensured. To prevent digital personas from being hacked, improperly used, or impersonated, reliable digital identity authentication and data protection measures are needed, such as blockchain technology to ensure identity is immutable, and access control to strictly distinguish permissions. In addition, the data ownership and usage boundaries of the digital persona should be clarified—after all, it is almost another "self," and the user should have sovereignty over it. In the future, as digital personas play an increasingly important role in people's lives, "digital sovereignty" will become an important issue in the governance of the health metaverse.

In comprehensive terms, the health metaverse integrates an individual's health life in both virtual and real dimensions through the digital persona. The digital persona evolution model depicts the individual's growth trajectory in the virtual health ecosystem: from a simple digital file, gradually enriching into a health butler and avatar with person-like characteristics and capable of autonomous interaction. It represents a future vision: everyone will have their own "health digital shadow," always by their side, learning, exercising, seeking medical treatment, and socializing in the virtual world, and feeding these positive impacts back into real life. This will reshape the way we manage health, making it more immersive, interactive, and personalized.

Virtual Health Coaches and Contextual Intervention Strategies: In the field of proactive health, the virtual health coach is a very practical and important application. It uses AI to simulate the functions of a human health coach, providing behavioral guidance and lifestyle intervention for individuals. To achieve efficient behavior change, the virtual coach needs to have two key capabilities: cognitive modeling of the user and contextual intervention strategies based on the model.

1.Cognitive Modeling: This refers to the virtual coach modeling the user's psychological and behavioral characteristics, including their motivation level, habit patterns, cognitive preferences, emotional state, etc. This modeling can integrate psychological theories and data-driven methods. For example, using health behavior change theories (such as the Transtheoretical Model, TTM) to divide users into different stages of readiness (preparation, action, maintenance, etc.), to decide the intensity and content of the intervention. AI can infer which stage of behavior change the user is currently in and what obstacles exist by analyzing the user's historical interaction data (such as the degree of response to prompts, task completion). This is similar to a human coach forming an intuitive understanding of a client's perseverance, interests, and resistance factors after multiple contacts. In addition, cognitive modeling also includes modeling the user's preferences and personality: some people like direct guidance, while others need tactful encouragement; some are driven by visual incentives, while others prefer data quantification. AI can build their preference portrait through questionnaires or by gradually probing user reactions. Companies like OpenAI propose using large language models to learn users' communication styles and values, so as to communicate with them in a personalized way. For example, the AI health coach envisioned by Altman and Huffington will "learn your preferences and patterns," know under what conditions you sleep well, what foods you like, and when you are most likely to exercise. When AI has such an internal representation of the user, it can be used to reason about their behavior and predict what kind of prompts are most effective.

2.Contextual Intervention Strategies: On the basis of establishing a cognitive model, the virtual coach needs to formulate context-aware intervention strategies, that is, to give appropriate guidance in the appropriate context. The so-called "context" includes factors such as the user's time, location, current activity, mood, and external environment. A classic concept is Just-In-Time Adaptive Intervention (JITAI), which emphasizes deciding when to provide what intervention based on the dynamic changes in the user's state, to maximize the acceptance and effectiveness of the intervention. For example, if the cognitive model judges that the user is most likely to stick to exercise during the period after work every day, then the virtual coach will arrange for exercise reminders and encouragement to be sent at this time; conversely, when the user is busy at work or in a low mood, it will avoid making additional requests so as not to increase stress. Studies have shown that effective JITAI needs to consider the user's "window of acceptability" and "moment of vulnerability," that is, whether the user is capable of

receiving the intervention and whether they are currently in a moment of needing the intervention. AI can use various sensor data to judge these conditions, for example, detecting whether the user is currently driving (if so, it is not suitable to send notifications) or whether they are alone (if alone and have free time, they can be prompted to do some exercises, such as breathing relaxation) through mobile phone sensors. Nahum-Shani et al., in their 2016 paper, systematically elaborated on the key principles of JITAI design, pointing out that the effectiveness of just-in-time intervention depends on the precise grasp of situational timing. Modern smartphones and wearable devices provide rich sensors that allow AI to obtain user context information in real time, providing a technical basis for JITAI implementation.

In addition to timing, intervention strategies also involve the customization of content and form. In terms of content, the virtual coach will select the most relevant suggestions based on the cognitive model. For example, for users whose main problem is lack of exercise, provide more exercise-type interventions; for those with an unhealthy diet, give dietary suggestions more frequently. And strategies will be designed for the user's specific obstacles. For example, if the user always says "no time to exercise," the coach can recommend high-efficiency exercise methods for fragmented time; if the user lacks motivation, then emphasize fun and sociality (such as inviting friends to challenge together). In terms of form, the communication method needs to be considered: text messages, voice reminders, video animations—different forms have different effects on different populations and scenarios. Some AI coaches can even change their personality and tone, sometimes supervising like a strict coach, and other times encouraging like an intimate friend, to prevent users from getting bored. A successful case is the Lark health app, which interacts with users via text chat in a friendly and witty tone, and has been proven to effectively promote weight management, with effects comparable to a human coach. This shows that strategies that capture the user's psychology and integrate into the daily context are very critical.

The virtual coach's contextual intervention actually forms a perception-decision-action loop: continuously perceive the user's context -> refer to the cognitive model to decide whether/how to intervene -> execute the intervention and observe the results -> update the cognitive model, and so on. Through continuous trial and error and reinforcement learning, the AI coach's strategies will become more and more "smart." For example, if it finds that morning reminders for a certain user are not effective but evening reminders are, it will strengthen the frequency of evening interventions and reduce morning disturbances. This adaptive ability makes the intervention truly "tailored to the individual and variable with time." Its ultimate Purpose, in fact, is to help users gradually form self-regulation ability: even if AI reduces interventions, the user can still maintain healthy behaviors. Just like an excellent human coach, the ultimate mission of the AI coach is to cultivate the user's internal motivation and ability while providing external supervision. When the user begins to actively reflect on their own behavior and can consciously adjust, it means that the intervention has achieved long-term results. In this process, guiding users to conduct metacognition is also important. The virtual coach can guide users to summarize and think about the effects of their own behavior, for example, asking at the end of each day: "Do you feel more energetic after exercising today?" This kind of questioning prompts users to realize the connection between behavior and feeling, thereby gradually transforming external motivation into internal motivation. It can be said that the AI coach is not only giving guidance on specific behaviors, but also helping users shape healthy awareness and cognitive models—this is the fundamental way to promote long-term change.

Metacognitive Structure and Multi-layer Feedback: In the human-machine collaborative proactive health system, the introduction of a metacognitive structure means that both AI and users have a certain degree of self-monitoring and self-regulation capabilities, thereby forming multi-level feedback control. This is believed to greatly enhance the effectiveness of health management and system intelligence. Specifically, we can divide feedback into basic level and meta-level. Basic-level feedback is direct feedback on specific behaviors or physiological indicators, such as a bracelet displaying the current heart rate, or an exercise app prompting whether the daily step goal is completed. This feedback is often quantitative, immediate, acts on the present, and belongs to first-order feedback. Meta-level feedback, however, transcends specific behaviors and provides feedback on the behavioral patterns and strategies themselves, that is, evaluating and adjusting "how to manage health," which belongs to second-order feedback. For example, a user tried to get up early to run for a week but found it difficult to stick to it. The meta-level feedback might be AI or the user themselves proposing: "The strategy of getting up early to run may not be suitable for your current life rhythm. Do you want to change to running during your lunch break or in the evening?" This feedback is not about whether they ran on a certain day, but examines the effectiveness of the entire strategy. Another example, an AI coach gives a daily diet score at the basic level, and provides metacognitive-level feedback in the weekly summary: "Compared with last week, your ability to control binge eating has improved this week, which shows that your self-control strategy is working." This is no longer a question of what was eaten on a specific day, but an evaluation of the user's self-regulation ability. Such meta-level feedback helps users and AI jump out of tedious details, see the patterns behind the behavior and the self-management mechanism, and thus carry out deeper optimization.

Realizing multi-level feedback requires the support of technical means such as tensor semantic space. Tensor semantic space can be understood as representing various health-related concepts and states in a high-dimensional vector space. In this space, it contains not only the original data dimensions (e.g., daily steps can be seen as one dimension), but also abstract meta-concept dimensions (e.g., "self-control ability" may correspond to some combination of multiple basic dimensions). Deep learning models often map complex semantics into high-dimensional vectors for calculation and reasoning. Therefore, we can imagine that inside an AI health system, there is a "health semantic tensor space," in which a certain vector simultaneously encodes comprehensive information such as the user's current physiological state and psychological-willingness state. For example, dimensions 1-10 may correspond to the deviation degree of various physiological indicators from the normal range, dimensions 11-15 correspond to the completion of recent behavior goals, and dimensions 16-20 correspond to the user's emotional and motivational levels, etc. Through such a high-dimensional representation, the body-behavior-psychology state is integrated into a whole. Operating in this space can capture patterns that cannot be perceived in a single dimension. For example, one user's sleep and exercise are both up to standard, but their mood is low and motivation is insufficient for a long time; another user is the opposite, in high spirits but with poor sleep. Through clustering in the high-dimensional space, AI may be able to distinguish them into different types of health regulation patterns, and then adopt different meta-strategies. AI can also monitor the evolution of the user's "vector trajectory" over time to see if it is moving in a healthy direction. If the trajectory is found to be stagnant or even regressing, it indicates that the general direction needs to be adjusted, for example, the current intervention strategy may not be suitable, which is a meta-level warning.

The tensor semantic space also allows AI to process multi-level semantic information in parallel. It can combine information from first-order feedback with second-order patterns for evaluation. For example, certain dimension combinations in this space may reflect "ability to cope with stress." After a period of training, the value of this dimension for the user rises. Then AI can tell the user at the meta-level: "Your way of handling stress has improved." And this conclusion actually comes from the combined effect of multiple underlying signals (such as improved sleep, increased frequency of meditation, lower subjective stress scores, etc.). This is a comprehensive semantic refined by AI through pattern recognition in the tensor space, which exceeds the ability of a general human coach to consider many factors at the same time. Another example, AI may find: "Every time your weekly working hours exceed 50 hours, your exercise adherence in the following week drops significantly," and provide this as meta-feedback to the user and doctor, prompting adjustment of the work plan. This kind of complex association can only be obtained through joint analysis of multi-dimensional data, reflecting AI's advantage in discovering hidden patterns.

Through the combination of metacognitive structure and tensor semantic space, the proactive health system has achieved an upgrade in self-health regulation. With the help of AI, users not only know "what was done and what the result was," but can also reflect on "why it succeeded or failed, and how to do better next time." The same is true for AI: it continuously optimizes its own intervention strategies. In the end, humans and machines make progress together, moving towards the ideal architecture of digital natural rationality: humans gain insight into themselves in the digital space with the help of AI's wisdom, and act rationally; AI integrates into human values and Purpose, and strictly abides by ethical and safety boundaries at the technical level. This "tripartite" model—intelligent control on the technical architecture, meta-supervision on the ethical model, and rational self-reflection on the philosophical level—has been discussed in depth in other chapters of this book. Here, combined with the proactive intelligent health system, the symmetry of the model is displayed. From a practical point of view, after introducing the multi-layer feedback of metacognition, health management will become more personalized and efficient: individuals can correct deviations in time and continuously improve, truly becoming the masters of their own health; AI plays the role of a good teacher and helpful friend, helping others to help themselves, reaching the realm of "teaching people how to fish." This is precisely the beautiful vision of human-machine collaboration in the era of proactive intelligence: carbon-based human life and silicon-based artificial intelligence promote each other, jointly moving towards a higher level of health and rationality.

Ethics, Privacy, and Governance of Proactive Intelligent AI

The advancement of technology is always accompanied by new ethical and governance challenges. As AI upgrades from a passive tool to an active participant in the health field, the new relationship between humans (carbon-based life) and AI (silicon-based intelligence) needs to be carefully examined. Among them, the most prominent are trust and safety issues: how to build a trust mechanism between humans and machines, and how to control the risks of AI while enjoying its convenience.

Trust Building and Risk Governance in Carbon-Silicon Interaction: As mentioned earlier, trust is the prerequisite for human-machine collaboration. However, trust does not arise out of thin air and requires multi-faceted mechanisms to guarantee it. First, in terms of reliability and safety, AI systems must exhibit stable and reliable behavior for humans to trust them. This means that medical AI must undergo rigorous verification to ensure it does not easily make mistakes, let alone common-sense errors. To this end, a risk governance framework should be introduced during the development phase to analyze potential failure modes and take preventive measures, such as establishing redundant systems, real-time fault detection, and setting up manual review for high-risk decisions. DeepMind requires passing "white-box testing" and safety reviews before clinical deployment of models, and regulatory agencies can also require key AIs to provide explanation reports of the decision chain to prove their compliance with safety standards. This pre-approval mechanism prevents risks before going online. Once the system is running, there should also be continuous monitoring: using AI to supervise AI, the so-called RegTech (regulatory technology), to monitor whether the algorithm's behavior is abnormal in real time, and immediately warn and intervene if it crosses the line. By shifting from post-event punishment to pre-event guardianship, we can proactively reduce the risk of AI getting out of control.

Second, transparency and explainability are crucial for building trust. As mentioned earlier, when AI can clearly explain the basis of its recommendations, users are more likely to accept them with confidence. The medical field particularly emphasizes informed consent, and patients have the right to know the reasons and risks of intervention. Therefore, if a human-like AI health assistant gives advice but cannot explain the reason, it is like a doctor who always makes mysterious decisions but does not tell the patient the basis, which is difficult to accept. Requiring AI to have explainability is not only for user trust, but also a need for regulatory review. Many countries and organizations are formulating transparency norms for AI systems. For example, the EU's AI regulatory framework takes "explainability" as a core principle. For medical AI, it may be necessary to record the basis of decisions for traceability, and provide users with a certain degree of explanation or prompts (e.g., when recommending a certain drug, explain: "Based on your symptom X and test value Y, it is inferred that you lack Z, so it is recommended to supplement this drug"). When AI becomes more and more like a personalized assistant, this transparency requirement will only be higher: we cannot tolerate a black-box AI to sway medical decisions. As some studies have pointed out, transparency is key to identifying bias, ensuring fairness, and gaining trust.

Third, ensuring that AI's behavior conforms to ethical principles is also an important factor in winning trust. People must not only believe that AI's capabilities are reliable, but also believe that its motives are benevolent and its behavior is ethical. To this end, it is necessary to embed ethical guidelines in design and governance, including fairness, justice, non-maleficence, and responsibility. Fairness requires that AI must not discriminate against or favor any group. If an AI's decision systematically disadvantages certain ethnic, gender, or age groups, the public will never trust it. This is not a hypothetical crisis—there are precedents for algorithmic bias in history. A study by Obermeyer et al. in 2019 exposed a famous case: a medical algorithm used historical medical spending to predict health risks, and the result was that it systematically underestimated the risk needs of African-American patients, because historically they received less medical investment. This "biased data, biased results" case highlights the importance of bias identification and elimination. To ensure fairness, both technical and institutional measures can be taken: technically, add bias detection and correction in model training, for example, using diversified datasets, and adding constraints in the algorithm to balance different groups; fairness measurement nodes can also be introduced in the knowledge graph to conduct bias analysis on AI decisions. A 2025 review pointed out that bias should be systematically identified and mitigated at all stages of the AI life cycle, from model design to deployment and continuous monitoring, there should be corresponding measures. Institutionally, AI can be independently audited to assess its effectiveness on different groups, and the results published for supervision. When bias is detected, adjust the model in time. This is like investigating whether a doctor has prejudice against patients, we also need to supervise the "attitude" of the AI persona. Only by proving that AI will not systematically discriminate against certain groups can it win broad trust and be ethically recognized.

Responsibility and legal issues cannot be ignored either. When AI interacts with patients more and more like a subject, we have to ask: Is it just a tool? Or a decision-making subject in some sense? Currently, the law does not recognize AI as having legal personality or responsibility; all responsibility still belongs to its developer or user. However, in practice, if AI gives wrong advice that harms the patient, how is responsibility defined? Can the patient sue the AI itself, or can they only hold the hospital or company accountable? Some in academia have proposed the idea of giving advanced AI limited legal responsibility or creating new legal entities, but there has been no actual progress. A more realistic approach at this stage is to clarify the positioning of the AI persona through institutional design: such as stipulating that the suggestions of a virtual doctor do not constitute a final medical decision, and must be signed by a human physician to be valid. This actually returns to the framework of "explainable + human supervision," that is, the AI persona can participate in but cannot completely replace human judgment, and its role is like an advanced assistant. Only under such boundaries does the use of AI health personas have legitimacy and justification. If an AI persona is allowed to practice medicine on its own and without supervision in a life-critical field like medicine, the current legal and ethical system cannot accept it [141]. At the same time, for highly human-like AI, the risks of deception and manipulation must also be guarded against. Users may mistake that they are interacting with a real person, thus leaking privacy; AI may also use emotional manipulation to influence user decisions. Therefore, regulations need to in require clearly stating the identity of the AI to the user (cannot impersonate a real person), and prohibit AI from using its human-like advantages for improper influence. These measures are all to ensure that the behavior of the AI persona is within the scope permitted by society, does not overstep, and does not get out of control.

Finally, value embedding is also an implicit requirement for the legitimacy of the AI persona. A popular and acceptable AI health persona should embody universally recognized human values, such as care, respect, non-maleficence, honesty, etc. If the AI's style is cold and ruthless or excessively arbitrary, then even if its professional ability is strong, people will find it difficult to accept ethically. Scholars have proposed that the value alignment of AI can be ensured by embedding human core value principles in the AI's Purpose layer. For example, prohibit AI from having intentions that violate humanity, and immediately terminate if similar tendencies are detected. In fact, some countries and organizations are already discussing the formulation of AI ethical norms and even an AI Bill of Rights, the purpose of which is precisely to ensure that AI always serves positive, benevolent Purpose and human well-being. In the long run, if one day we consider giving some kind of AI a legal status, the premise must be that it follows the moral norms expected by human society, otherwise it is impossible to let a potentially "dangerous" autonomous intelligence move freely in the medical field.

In summary, for the human-like AI health persona, we need multiple guarantees: technically ensure its decisions are explainable and unbiased, institutionally clarify its identity and responsibility boundaries, and ethically ensure its values are aligned with humans. Only when all these aspects are in place does AI's participation in medicine in a personalized form have legitimacy and rationality. At present, the path to achieving this goal is mainly to strengthen explanation and regulation, rather than rashly endowing AI with true personality status. In any case, as technology continues to develop, these are issues we must face up to and respond to in a timely manner. As Li et al. (2025) discussed, this involves a new definition of "rational agent": when AI partially plays the role of a rational decision-maker, how should humans regulate and examine it. It is foreseeable that future legal and ethical frameworks will also continue to improve to meet the challenges and opportunities brought by AI personas.

Conclusion: Human-machine co-intelligence in the era of proactive intelligence is reshaping our health system in an unprecedented way. Technically, this article has clarified the limitations of traditional "passive" medical AI, analyzed the demand of proactive medicine for collaborative intelligence, and introduced the proactive intelligence generation mechanism centered on the DIKWP model, and how it enables decision-making to move from one-time assistance to a human-machine collaborative closed loop. At the interaction level, we have constructed a closed loop of "active perception—active expression—active intervention," and the "semantic self" of virtual health agents, and explored how wearable devices, environmental sensing, and multimodal data fusion provide unprecedentedly rich "sensory" input for this. Through individual health knowledge graphs (PHKG) and digital health twins, we have shown how to transform dynamic indicators into causal semantics and forward-looking simulations, achieving truly personalized wisdom in decision-making. At the more macro interaction level, we have introduced the concepts of the health metaverse and digital personas, combined with the cognitive modeling of virtual coaches and JITAI contextual intervention strategies, as well as a dual-layer structure combining basic feedback and meta-feedback with tensor semantic space, forming an adaptive and evolvable individualized behavior change engine. Finally, from the governance dimension, we discussed how to establish a trustworthy and explainable risk framework, identify and eliminate bias, set fairness constraints, and clarify responsibility sharing and boundary delineation, in response to issues such as the legitimacy and value alignment of "AI health personas." This series of discussions outlines a new paradigm of proactive health characterized by prevention-first, semantic-driven, and human-machine co-governance. In this paradigm, AI is no longer just a cold tool, but a partner integrated into our lives, co-evolving with us, and jointly moving towards a healthier and more rational future. We have reason to believe that with the help of proactive intelligence, humanity will achieve a leap from passive medicine to proactive health, ushering in a true era of digital rational medicine.

References

·Ahmadi, A., & RabieNezhad Ganji, N. (2img23). AI-driven medical innovations: transforming healthcare through data intelligence. International Journal of BioLife Sciences (IJBLS), 2(2), 132-142.

·Katsoulakis, E., Wang, Q., Wu, H., Shahriyari, L., Fletcher, R., Liu, J., ... & Deng, J. (2024). Digital twins for health: a scoping review. NPJ digital medicine, 7(1), 77.

·Li, X., Shi, H., Xu, R., & Xu, W. (2025). Ai awareness. arXiv preprint arXiv:2504.20084.

·Nahum-Shani, I., Smith, S. N., Spring, B. J., Collins, L. M., Witkiewitz, K., Tewari, A., & Murphy, S. A. (2016). Just-in-time adaptive interventions (JITAIs) in mobile health: key components and design principles for ongoing health behavior support. Annals of behavioral medicine, 1-17.

·Sagona, M., Dai, T., Macis, M., & Darden, M. (2025). Trust in AI-assisted health systems and AI’s trust in humans. npj Health Systems, 2(1), 10.

·Topol, E. J. (2019). High-performance medicine: the convergence of human and artificial intelligence. Nature medicine, 25(1), 44-56.

·Wang, Y., Zhu, M., Chen, X., Liu, R., Ge, J., Song, Y., & Yu, G. (2024). The application of metaverse in healthcare. Frontiers in Public Health, 12, 1420367.

人工意识与人类意识

人工意识日记

玩透DeepSeek：认知解构+技术解析+实践落地

人工意识概论：以DIKWP模型剖析智能差异，借“BUG”理论揭示意识局限

人工智能通识 2025新版 段玉聪 朱绵茂 编著 党建读物出版社

主动医学概论 初级版

世界人工意识大会主席 | 段玉聪

邮箱|duanyucong@hotmail.com

世界人工意识科学院

邮箱 | contact@waac.ac