Context-aware adaptive health intervention method, apparatus, and electronic device

Abstract

The application provides a kind of adaptive health intervention method, device and electronic equipment based on context awareness, it is related to health intervention technical field, the method comprises: intervention context descriptor for representing current intervention context is constructed based on the multi-modal context data of target object;With intervention context descriptor as query vector, at least one historical intervention record similar to current intervention context is retrieved in experience library, historical intervention record includes historical context data, historical intervention scheme and historical feedback result corresponding to historical context data;The standard medical intervention path of target object is obtained by retrieving in medical knowledge base based on the physiological sign data of target object;According to the historical feedback result in each historical intervention record and emotional perception data, the standard medical intervention path is regulated, and the final health intervention scheme adapted to current intervention context is generated.The application realizes the high execution degree and dynamic adaptability of intervention decision scheme in real life scene.

Description

Technical Field

[0001] This invention relates to the field of health intervention technology, and in particular to a context-aware adaptive health intervention method, device, and electronic device. Background Technology

[0002] With the continuous development of IoT technology, artificial intelligence algorithms, wearable devices, and environmental sensing terminals, health intervention systems are gradually shifting from a passive, response-based disease consultation model to a health management model characterized by proactive perception and dynamic decision-making. In this process, Retrieval-Augmented Generation (RAG) technology, by introducing external authoritative medical knowledge resources, constrains the generation process of large language models, thus improving the scientific rigor and reliability of generated health intervention plans to a certain extent. Simultaneously, with the widespread deployment of wearable devices, environmental sensors, and smart terminals in health scenarios, an individual's physical environment and psychological and emotional characteristics are increasingly becoming important bases for health intervention decisions. Numerous studies have shown that environmental factors (such as temperature, humidity, noise, and light) and emotional states (such as anxiety, depression, and stress levels) significantly affect an individual's adherence to health behaviors and the effectiveness of interventions. However, existing intelligent health systems mostly remain at the level of question-and-answer or recommendation based on static medical knowledge, failing to fully incorporate real-time environmental and emotional perception information into the decision-making process, making it difficult to achieve dynamic health intervention generation tailored to real-world situations.

[0003] Current research on RAG technology in the healthcare field primarily focuses on the precise retrieval and reasoning capabilities of medical knowledge. For example, it integrates medical literature, clinical guidelines, case databases, or medical imaging data to provide supplementary support for disease diagnosis and treatment recommendations. Some studies have further explored multimodal co-modeling of text and medical images, such as computed tomography (CT) and X-ray, to improve the model's ability to identify pathological features. However, the core focus of this approach remains on the medical representation of the disease itself, with insufficient consideration for the individual's real-world context.

[0004] In real-world health intervention scenarios, an individual's health decisions are not only influenced by their disease state but also highly dependent on their physical environment and psychological / emotional state. For example, environmental factors such as weather changes, air quality, and indoor temperature and humidity directly affect the rationality of intervention recommendations, such as exercise, respiratory disease prevention and control, and daily routines; emotional states such as anxiety, depression, or stress levels significantly influence users' acceptance and adherence to intervention measures. Therefore, existing health intervention plans generated by medical RAG systems suffer from contextual mismatches in real-world applications.

[0005] Therefore, an effective technical solution is urgently needed to solve the above-mentioned technical problems. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a context-aware adaptive health intervention method, device, and electronic device, which achieves high execution and dynamic adaptability of intervention decision-making schemes in real-life scenarios.

[0007] In a first aspect, the present invention provides a context-aware adaptive health intervention method, which includes the following steps: Based on the multimodal contextual data of the target object, an intervention context descriptor is constructed to represent the current intervention context; the multimodal contextual data includes environmental perception data and emotion perception data; Using the intervention scenario descriptor as the query vector, at least one historical intervention record similar to the current intervention scenario is retrieved from the experience base; the experience base stores multiple historical intervention records in advance, and the historical intervention records include historical scenario data, historical intervention plans corresponding to the historical scenario data, and historical feedback results; Based on the physiological signs data of the target object, a search is performed in the medical knowledge base to obtain the standard medical intervention path for the target object; Based on the historical feedback results in the retrieved historical intervention records and the emotional perception data, the standard medical intervention path is adjusted to generate a final health intervention plan adapted to the current intervention situation.

[0008] According to a context-aware adaptive health intervention method provided by the present invention, the multimodal context data further includes historical intervention feedback data; the multimodal context data based on the target object is used to construct an intervention context descriptor for characterizing the current intervention context, including: The environmental perception data is feature-mapped and converted into at least one discrete environmental state label. The emotion perception data is quantified to generate an emotion state vector; The environmental state labels, the emotional state vector, and the feedback parameters extracted from the historical intervention feedback data are fused to generate the intervention situation descriptor.

[0009] According to a context-aware adaptive health intervention method provided by the present invention, the step of adjusting the standard medical intervention path based on historical feedback results in each retrieved historical intervention record and the emotion perception data to generate a final health intervention plan adapted to the current intervention context includes: Based on the environmental perception data, the feasibility of at least one candidate intervention action in the standard medical intervention path is determined, and candidate intervention actions that are not feasible in the current environment are eliminated to obtain the remaining candidate intervention actions. Based on the emotion perception data and the retrieved historical intervention records, the execution priority weights of the remaining candidate intervention actions are dynamically adjusted to obtain the adjusted candidate intervention actions. Based on the emotion perception data, the intensity of the execution parameters corresponding to each of the modified candidate intervention actions is dynamically adjusted to generate the final health intervention plan.

[0010] According to a context-aware adaptive health intervention method provided by the present invention, the feasibility determination of at least one candidate intervention action in the standard medical intervention path based on the environmental perception data includes: If the environmental conditions required to execute any of the candidate intervention actions conflict with the environmental perception data, the candidate intervention action is determined to be an unexecutable candidate intervention action under the current environment.

[0011] According to a context-aware adaptive health intervention method provided by the present invention, the step of dynamically adjusting the execution priority weights of the remaining candidate intervention actions based on the emotion perception data and the retrieved historical intervention records to obtain the adjusted candidate intervention actions includes: If the retrieved historical intervention records show that, in a situation similar to the current intervention situation, the historical feedback result of any candidate intervention action is lower than a preset threshold, the current execution priority weight of the candidate intervention action is reduced. When the emotion perception data indicates that the target object is in a preset negative emotional state, the execution priority weight of the candidate intervention action that requires high compliance is reduced to obtain each of the modified candidate intervention actions.

[0012] According to a context-aware adaptive health intervention method provided by the present invention, the step of dynamically adjusting the intensity of the execution parameters corresponding to each of the modified candidate intervention actions based on the emotion perception data to generate the final health intervention plan includes: Determine the degree of deviation between the emotional state represented by the emotional perception data and the preset baseline emotional state; Based on the degree of deviation, a scaling factor is determined to adjust the intensity of the intervention; The scaling factor is used to reduce the standard execution parameters of each of the modified candidate intervention actions to obtain execution parameters that are adapted to the current emotional state. Based on the execution parameters, the final health intervention plan is generated.

[0013] According to a context-aware adaptive health intervention method provided by the present invention, the method further includes: Output the final health intervention plan and obtain the execution feedback data after the target object implements the final health intervention plan; the final health intervention plan is structured data, and the final health intervention plan includes at least environmental adaptability explanation information for explaining the reasons for the plan adjustment, emotional comfort information for empathy, and specific execution instructions; the execution feedback data includes objective physiological indicator change data and / or user subjective satisfaction rating data; The current intervention scenario, the final health intervention plan, and the execution feedback data are treated as a new historical intervention record; The new historical intervention records are stored in the experience base to update the experience base.

[0014] According to a context-aware adaptive health intervention method provided by the present invention, updating the experience base includes: If the execution feedback data indicates that the final health intervention plan is successful, the association vector established based on the current intervention situation and the executed action will be moved by a preset step size towards the successful cluster center in the vector space corresponding to the experience base. If the execution feedback data indicates that the final health intervention plan has failed, the correlation vector in the vector space is moved by the preset step size in the opposite direction to the successful cluster center.

[0015] Secondly, the present invention also provides a context-aware adaptive health intervention device, which includes the following modules: An intervention context descriptor construction module is used to construct an intervention context descriptor to represent the current intervention context based on the multimodal context data of the target object; the multimodal context data includes environmental perception data and emotion perception data; The hybrid retrieval module is used to retrieve at least one historical intervention record similar to the current intervention situation from the experience base, using the intervention situation descriptor as the query vector. The experience base stores multiple historical intervention records in advance, and each historical intervention record includes historical situation data, historical intervention plans corresponding to the historical situation data, and historical feedback results. Based on the physiological characteristics data of the target object, a retrieval is performed in the medical knowledge base to obtain the standard medical intervention path for the target object. The strategy control module is used to control the standard medical intervention path based on the historical feedback results in each of the historical intervention records and the emotion perception data, and generate a final health intervention plan that is adapted to the current intervention situation.

[0016] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the context-aware adaptive health intervention method as described above.

[0017] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the context-aware adaptive health intervention method as described above.

[0018] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the context-aware adaptive health intervention method as described above.

[0019] The present invention provides a context-aware adaptive health intervention method, device, and electronic device. The multimodal context data includes environmental perception data and emotion perception data. First, an intervention context descriptor is constructed based on the multimodal context data of the target object to characterize the current intervention context. Then, using the intervention context descriptor as a query vector, at least one historical intervention record similar to the current intervention context is retrieved from an experience base. The experience base pre-stores multiple historical intervention records, each including historical context data, corresponding historical intervention plans, and historical feedback results. Further, a search is performed in a medical knowledge base based on the target object's physiological characteristics data to obtain a standard medical intervention path for the target object. Finally, based on the historical feedback results and emotion perception data from the retrieved historical intervention records, the standard medical intervention path is adjusted to generate a final health intervention plan adapted to the current intervention context.

[0020] This invention overcomes the limitations of existing medical RAG systems that rely solely on textual semantics and lack environmental and emotional dimensions. By constructing intervention scenario descriptors, the system maps environmental physical indicators (such as noise and light) and user psychological states (such as anxiety and fatigue) into "intervention scenario descriptors." Based on these descriptors, a hybrid retrieval is performed to obtain candidate historical intervention records. Then, based on historical feedback results from each historical intervention record and the aforementioned emotional perception data, the standard medical intervention path is adjusted to generate a final health intervention plan adapted to the current intervention scenario. This ensures that the generated final health intervention plan is not only theoretically correct but also "executable" in the actual physical environment and user psychological state, thereby achieving high executability and dynamic adaptability of the intervention decision-making plan in real-life scenarios. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is one of the flowcharts of the context-aware adaptive health intervention method provided by the present invention.

[0023] Figure 2 This is the second flowchart of the context-aware adaptive health intervention method provided by the present invention.

[0024] Figure 3 This is a schematic diagram of the structure of the context-aware adaptive health intervention device provided by the present invention.

[0025] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0027] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, not limited in number; for example, a first node can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0028] To more clearly understand the various embodiments provided by the present invention, the technical content involved in the present invention will first be described as follows: On the one hand, existing medical RAG systems generally do not incorporate environmental perception data and emotional characteristics as core decision-making elements into the retrieval and generation process, resulting in a situational mismatch between the generated results and real-world applications.

[0029] On the other hand, most existing generative health systems rely on static input information for reasoning and generation. Even with the introduction of mechanisms such as chain thinking to simulate a doctor's analytical logic, the reasoning process remains closed, lacking the ability to connect with real-time environmental changes and emotional fluctuations. This static decision-making model makes it difficult for generated health intervention plans to respond promptly to changes in individual states, and it also lacks necessary empathy and flexibility. Furthermore, while some existing technologies attempt to introduce affective computing or environmental awareness modules, they often treat relevant information as supplementary descriptive inputs, failing to deeply couple them with retrieval strategies, knowledge selection, and generative expression methods. This makes it difficult to achieve context-based adjustment of intervention intensity, intervention method selection, and tone strategy adjustment, resulting in overall intervention decisions still exhibiting template-like characteristics.

[0030] Despite the excellent performance of existing medical RAG systems in diagnostic accuracy and medical knowledge reasoning, when it comes to providing actionable lifestyle intervention decisions in the specific scenario of home-based health management for older adults in the community, existing technologies still reveal the following significant limitations and shortcomings: (a) Lack of perception and adaptation to physical environment and real-time emotions ("Understanding the disease but not the person and environment") The input of existing RAG systems mainly relies on static medical texts, electronic health records (EHRs), or medical images. The above static information focuses on analyzing "pathology" and "symptoms," but often ignores the most critical external physical environment constraints (such as noise, light, temperature and humidity) and the user's real-time psychological state (such as anxiety and fatigue) when implementing intervention plans. For example, it is reasonable to recommend "meditation" in medical logic, but if the user is currently in a high-decibel noise environment, the suggestion becomes unfeasible due to the lack of environmental perception; similarly, if the user is extremely tired or in a low mood, the standard intensity exercise suggestion may trigger resistance. Existing technology lacks a mechanism to transform unstructured environmental physical quantities and emotional psychological states into decision variables, resulting in generated plans that are "medically correct" but "difficult to implement" in real-life scenarios.

[0031] (b) Lack of a long-term adaptive closed-loop strategy based on execution feedback (“one-time recommendation” rather than “long-term companionship”). Existing medical RAG research focuses primarily on the accuracy of single question-and-answer sessions (e.g., diagnostic accuracy), typically employing a unidirectional linear process of “retrieval-generation-termination.” While some studies utilize reinforcement learning to optimize inference paths, their primary goal is to generate better textual responses, rather than to continuously track the effects of real-world interventions. In elderly health management, intervention is a long-term, dynamic process. Existing systems lack a complete closed-loop mechanism of “protocol execution → effect feedback → strategy weight update.” The system cannot remember users’ past preferences for a particular intervention in specific contexts (e.g., “refusing high-intensity exercise”) or objective physiological feedback (e.g., “heart rate not improved”), preventing the system from learning from historical interactions like a family doctor, thus hindering the dynamic evolution and precise convergence of personalized strategies.

[0032] In summary, existing technologies have mainly solved the problem of "generating reasonable suggestions based on the patient's condition and medical knowledge," but have not yet effectively solved the key technical challenge of "dynamically generating timely and executable health intervention decisions by combining real-time environmental conditions and individual emotional characteristics," making it difficult to meet the needs of refined and context-sensitive health interventions.

[0033] Therefore, constructing a retrieval enhancement generation system that can integrate multimodal environment and emotion perception, and has the ability to adapt long-term strategies based on feedback, to achieve the leap from simple "medical question answering" to "scenario-based intervention decision-making," is a key technological bottleneck that urgently needs to be overcome in the field of elderly health management, and it is also the core research direction of this application.

[0034] The following is combined Figures 1 to 4 The present invention describes a context-aware adaptive health intervention method, apparatus, and electronic device.

[0035] Figure 1 This is one of the flowcharts of the context-aware adaptive health intervention method provided by the present invention, such as... Figure 1 As shown, the method includes the following: Step 101: Based on the multimodal contextual data of the target object, construct an intervention context descriptor to represent the current intervention context; the multimodal contextual data includes environmental perception data and emotion perception data.

[0036] The subject executing the context-aware adaptive health intervention method provided by this invention can be an electronic device, or any other subject capable of implementing the context-aware adaptive health intervention method, such as other context-aware adaptive health intervention systems.

[0037] The method includes the following steps: Multimodal contextual data includes environmental perception data. and emotion perception data Environmental perception data This refers to data related to the environment of the target object (noise level, indoor light intensity, temperature, humidity, air quality, timestamps, etc.) acquired through IoT devices in homes or communities. Emotional perception data. To receive the output of the multimodal emotion computing module, which is built on a large model, the input data sources include, but are not limited to: visual modality (facial expression images / video frames captured by indoor cameras or terminal device cameras, from which facial action unit features are extracted), audio modality (voice tone, speech rate, and pause features collected through a microphone array), and text modality (textual semantic features of user-system dialogue). The multimodal emotion computing module supports single-modal or multimodal fusion input to ensure the feasibility of solutions under different privacy settings. The output emotion perception data... Specifically, it refers to the emotional state of a target object obtained through reasoning using a large model based on the input data source.

[0038] In practical applications, it is based on a pre-defined [environment / emotion feature mapping table]. Regarding the above environmental perception data Emotion perception data Feature mapping and vector synthesis are performed to obtain the corresponding environmental constraint labels and emotion feature vectors. First, the original data is denoised and normalized, and then the mapping function is applied... Convert continuous numerical values ​​into discrete feature labels. This represents the input multimodal data, such as environmental perception data or emotion perception data. The [Environment / Emotion Feature Mapping Table] is mentioned above. The system can dynamically and adaptively update based on correction operators generated from historical user feedback to avoid static manual verification rules. Subsequently, the environmental constraint labels, emotional feature vectors, and feedback parameters corresponding to historical intervention feedback data are concatenated into tensors to generate an Intervention Context Descriptor (ICD) that represents the current intervention situation. This ICD will serve as the core query (query vector) for the "experience base" retrieval in subsequent steps.

[0039] Step 102: Using the intervention scenario descriptor as the query vector, retrieve at least one historical intervention record similar to the current intervention scenario from the experience base; the experience base stores multiple historical intervention records in advance, including historical scenario data, historical intervention plans corresponding to the historical scenario data, and historical feedback results.

[0040] Specifically, based on the generated intervention context descriptor ICD, a multimodal hybrid retrieval based on knowledge graph constraints and experience backtracking is performed, including two paths: Path 1: Contextual Experience Feedback Flow (Dynamic Adaptive Path). The core of this path lies in building and utilizing a dynamically evolving log database (experience repository). The construction logic of this database is as follows: maintaining a time-series heterogeneous behavioral graph database as the basis for experience backtracking. Each log entry is not just a simple text record, but a historical intervention record. The historical intervention record includes historical context data, the corresponding historical intervention plan, and the historical feedback results. That is, each log entry represents a structured "Context-Intervention-Reward" triple (State-Action-Reward, SAR). State (S) includes the current timestamp, multi-dimensional environmental vectors (physical parameters such as noise and illumination), and emotional state vector; Action (A) is the specific intervention plan recommended by the system and its execution parameters (such as duration and intensity); Reward (R) includes the objective execution completion rate (0-100%) and the user's subjective satisfaction rating. Simultaneously, the database introduces a time decay factor, giving higher weight to recently generated interaction records than older records when calculating similarity, to adapt to the dynamic drift of user preferences.

[0041] In practical applications, the retrieval logic for path 1 is as follows: Using the intervention context descriptor ICD as the query vector, a search is performed in the aforementioned experience base. For example, a K-Nearest Neighbor (K-NN) search based on cosine similarity is executed to retrieve at least one historical intervention record similar to the current intervention context. In other words, the set of historical intervention records under similar contexts ("similar physical environment and psychological parameters") is recalled. .

[0042] It's important to note that path 1 retrieves not the history of "the same disease," but rather the history of interactions under "the same environment and emotions." This process simulates how human doctors use past experience to aid decision-making. For example, the current user's ICD is... This indicates a high-noise environment and an "anxious" emotional state. The invention will recall all historical interaction records from this "high-noise and anxious" context. For example, the system retrieves historical records and finds that recommending "listening to light music" in this context resulted in an average feedback score of only 2.0 (out of 5.0), with an execution rate below 30%. The system also finds that recommending "indoor diaphragmatic breathing" resulted in an average feedback score of 4.5. Finally, in the generated candidate set, the system will automatically reduce the recommendation priority of "listening to light music" (or filter it directly based on a threshold) and increase the weight of "indoor diaphragmatic breathing," thereby avoiding repeating past mistakes and directly outputting historically validated solutions.

[0043] Step 103: Based on the physiological characteristics data of the target object, search the medical knowledge base to obtain the standard medical intervention path for the target object.

[0044] Specifically, path 2 of the multimodal hybrid retrieval is the medical professional knowledge flow (rule-based hard constraints).

[0045] This path acts as the system's "bottom-line guardian," providing standard treatment principles that remain unchanged regardless of the environment. It utilizes a pre-constructed medical knowledge graph (KG) as the underlying tool to output standard intervention paths (i.e., standard medical intervention paths) tailored to the specific physiological state of the target individual. ).

[0046] The retrieval target of this path is the safety boundary and standard treatment principles retrieved based on the user's physiological characteristics data. The output of path 2 is not directly used as the final instruction, but rather as the input object to be processed by subsequent pruning logic.

[0047] Step 104: Based on the historical feedback results and emotional perception data retrieved from each historical intervention record, adjust the standard medical intervention path to generate a final health intervention plan that is adapted to the current intervention situation.

[0048] Furthermore, upon obtaining standard medical intervention pathways and the retrieved historical intervention record set Subsequently, further analysis was conducted based on historical intervention record sets. Historical feedback results and acquired emotion perception data are used to adjust the parameters of the standard medical intervention path, that is, to execute strategy adjustment and pruning based on a parameterized control model. Specifically, historical feedback is used to "re-rank" and "filter" the search results to generate a final health intervention plan (final execution strategy) adapted to the current intervention context. The final health intervention plan is a structured strategy instruction after parameter adjustment. This step reflects the "flexibility" of the system. This step does not negate medical advice, but rather protects the feasibility of medical advice.

[0049] In practical applications, this invention innovatively introduces a parameter intensity degradation mechanism based on emotional sensitivity, while utilizing a medical knowledge graph (KG) as the underlying security constraint. For example, when a user's emotional vector deviates from the baseline range (e.g., extreme fatigue), the system does not mechanically push standard medical advice. Instead, it automatically switches the intervention intensity (e.g., exercise duration and frequency) to a "low-burden mode" by calculating a scaling factor. This humanized dynamic adjustment strategy retains the minimum feasible health behaviors while avoiding user resistance caused by high-intensity requirements, significantly improving intervention compliance among the elderly in negative physical and mental states.

[0050] The method provided in this invention uses multimodal contextual data, including environmental perception data and emotion perception data. First, an intervention context descriptor is constructed based on the multimodal contextual data of the target object to characterize the current intervention context. Then, using the intervention context descriptor as a query vector, at least one historical intervention record similar to the current intervention context is retrieved from the experience base. The experience base pre-stores multiple historical intervention records, which include historical contextual data, historical intervention plans corresponding to the historical contextual data, and historical feedback results. Further, a search is performed in a medical knowledge base based on the target object's physiological characteristic data to obtain the target object's standard medical intervention path. Then, based on the historical feedback results and emotion perception data in each historical intervention record, the standard medical intervention path is adjusted to generate a final health intervention plan adapted to the current intervention context.

[0051] This invention overcomes the limitations of existing medical RAG systems, which rely solely on textual semantics and lack environmental and emotional dimensions. By constructing intervention context descriptors, the system maps environmental physical indicators (such as noise and light) and user psychological states (such as anxiety and fatigue) into "intervention context descriptors (ICDs)." Based on the intervention context descriptors (ICDs), a hybrid retrieval is performed to obtain candidate historical intervention records. Then, based on the historical feedback results from each historical intervention record and the aforementioned emotional perception data, the standard medical intervention path is adjusted to generate a final health intervention plan adapted to the current intervention context. This ensures that the generated final health intervention plan is not only theoretically correct but also "executable" in the actual physical environment and user psychological state, thereby achieving high executability and dynamic adaptability of the intervention decision-making plan in real-life scenarios.

[0052] It should be noted that each implementation method of this application can be freely combined, rearranged, or executed individually, and does not need to rely on or depend on a fixed execution order.

[0053] According to the context-aware adaptive health intervention method provided by the present invention, the multimodal context data further includes historical intervention feedback data; based on the multimodal context data of the target object, an intervention context descriptor for characterizing the current intervention context is constructed, including: Feature mapping is performed on the environmental perception data to convert it into at least one discrete environmental state label; The emotion perception data is quantified to generate an emotion state vector; By fusing environmental state labels, emotional state vectors, and feedback parameters extracted from historical intervention feedback data, an intervention situation descriptor is generated.

[0054] Specifically, the multimodal contextual data also includes historical intervention feedback data (intervention history feedback). For example, users' physiological changes, such as whether heart rate variability (HRV) has improved, or objective performance data such as whether a standing motion was detected, can be automatically transmitted through smart wearable devices (such as wristbands) or environmental sensors (such as smart mattresses or infrared sensors). Another example is collecting explicit user feedback (such as "Did you find the suggestion helpful?" rating 1-5) through a terminal interface, such as application (App) pop-ups or voice assistant queries.

[0055] In some embodiments, step 101, constructing the intervention context descriptor (ICD), is achieved through the following steps: First, environmental perception data Feature mapping is performed to transform the environmentally perceived data into at least one discrete environmental state label. For example, a threshold function is set: for instance, when noise... When, it is mapped to a label , In decibels, For high noise; when illuminated When, it is mapped to a label . Lux, defined as lumens per square meter, is used to measure luminous intensity. Low light intensity.

[0056] Furthermore, regarding emotion perception data The process involves quantification to generate emotion state vectors. For example, raw emotion perception data is mapped to emotion state labels, and then these labels are encoded into numerical "emotion state vectors." .

[0057] Then, the environmental state labels, emotional state vectors, and feedback parameters extracted from historical intervention feedback data are concatenated into tensors to generate the intervention situation descriptor (ICD) for the current intervention situation.

[0058] The method provided in this invention performs structured modeling of unstructured community environmental physical quantities (noise, light, temperature, and humidity) and the emotional and psychological states of the elderly, transforming them into computable decision-making and regulatory variables—intervention context descriptors (ICDs). These ICDs are then used as the core query vector in the Retrieval Enhancement Generation (RAG) process, enabling personalized long-term adaptive intervention. This ensures that the generated final health intervention plan is not only theoretically correct but also "executable" in the actual physical environment and the user's psychological state.

[0059] According to the context-aware adaptive health intervention method provided by the present invention, based on historical feedback results and emotion perception data retrieved from various historical intervention records, the standard medical intervention path is adjusted to generate a final health intervention plan adapted to the current intervention context, including: Based on environmental perception data, the feasibility of at least one candidate intervention action in the standard medical intervention path is determined, and candidate intervention actions that are not feasible in the current environment are eliminated to obtain the remaining candidate intervention actions. Based on emotion perception data and retrieved historical intervention records, the execution priority weights of the remaining candidate intervention actions are dynamically adjusted to obtain the adjusted candidate intervention actions. Based on emotion perception data, the intensity of the execution parameters corresponding to each revised candidate intervention action is dynamically adjusted to generate the final health intervention plan.

[0060] In some embodiments, step 104 is implemented through the following steps: Step 1: Perform hard constraint pruning: Specifically, based on environmental perception data, the feasibility of at least one candidate intervention action in the standard medical intervention path is determined, and candidate intervention actions that are not feasible in the current environment are eliminated to obtain the remaining candidate intervention actions.

[0061] Standard medical intervention pathways include multiple candidate intervention actions. Hard constraint pruning involves determining the feasibility of candidate intervention actions based on current environmental perception data (environmental constraint labels). If the current environmental perception data (environmental constraint labels) is determined to be mismatched or conflicting with the physical environment parameters required for a candidate intervention action, then candidate intervention actions that are not executable in the current environment are eliminated; that is, logical pruning is performed.

[0062] Step 2, Adaptive update of soft constraint weights: Based on emotion perception data and historical intervention records, the execution priority weights of the remaining candidate intervention actions are dynamically adjusted to obtain the revised candidate intervention actions, in order to address the problem of low compliance with general recommendations of medical knowledge graphs (KG).

[0063] This regulatory mechanism is mainly achieved through the following two dimensions: 1) Emotional state gain / loss adjustment: By calculating the degree to which the real-time emotion vector (calculated based on emotion perception data) deviates from the baseline range, the system automatically identifies the potential success rate of users implementing intervention suggestions. For example, when the emotion vector is identified as "extremely tired" or "anxious," the system will reduce the original weight score of high-compliance actions (such as "long-distance walk") to avoid causing user resistance due to the output of high-intensity exercise suggestions.

[0064] 2) Preference correction based on negative gradient penalty.

[0065] By analyzing historical intervention records, a negative gradient signal is introduced to achieve precise convergence of user behavior preferences.

[0066] If feedback in historical intervention records shows that the target subject consistently exhibits similar situational descriptors (ICDs) If a certain type of candidate intervention action is rejected, the system will activate a negative gradient penalty mechanism, adjusting the priority weight of that candidate intervention action. Perform the following value update operation: in, The preset penalty coefficient is used to reduce the priority of the intervention action in the subsequent recommendation sequence through numerical decay, so as to ensure that the system outputs a more accurate solution that conforms to the individual's dynamic execution habits. For the first The original priority weights of each candidate intervention action, For the first The corrected priority weights for each candidate intervention action. The candidate intervention action was determined in the output instruction set. The sorting priority in the algorithm.

[0067] In practical applications, the corrected priority weight The lower bound is 0, that is, when the corrected priority weight When the value drops to 0, it is immediately removed from the candidate set. This is because if a user rejects too many times consecutively, It may become a negative number, in which case a lower limit value needs to be added.

[0068] Step 3, Parameter Strength Degradation: Based on emotion perception data, the intensity of the execution parameters corresponding to each revised candidate intervention action is dynamically adjusted to generate the final health intervention plan.

[0069] Medical knowledge graphs (KG) typically provide a "standard value" for the strength of a parameter. For example, it is recommended that patients with hypertension engage in 30 minutes of aerobic exercise daily. However, in real life, the physical and mental state of the elderly fluctuates: under standard conditions (pleasant mood, good physical strength), performing 30 minutes is not a problem, but when the elderly's emotional state changes (if the emotional vector deviates from the baseline range), they may resist the parameter strength of the standard value, leading to intervention failure.

[0070] Therefore, this invention presets a normal range (benchmark) for emotional states and detects the Euclidean distance between the emotional vector and the preset benchmark range in real time. If the emotional state is within the normal range, a standard value for parameter intensity is recommended; if it deviates from the normal range, the parameter intensity is linearly downgraded based on the degree of deviation. By lowering the execution threshold, this improves the elderly's compliance in negative states and prevents intervention interruptions.

[0071] After parameter adjustment through the above three steps, the final health intervention plan is obtained, such as the output instruction set. Output instruction set This is a structured strategy instruction refined through parameter adjustment. This parameter adjustment step reflects the "flexibility" of the plan. The final health intervention plan does not negate medical advice, but rather protects the feasibility of its implementation.

[0072] Optionally, the output instruction set is generated after the pruned policy set is generated. Following this, a constrained, structured intervention plan can be generated. For example, the process of generating and outputting the final health intervention plan (structured intervention plan) is achieved through the following steps: The core of this step lies in the "Prompt construction structure and input constraint mechanism," rather than the intelligence of the structured text generation engine itself.

[0073] Structured input constraint logic: The system is based on multi-source constraint information (standard medical intervention pathways). Historical intervention record collection Output instruction set Construct a specific input template for generation. This template, through preset logical guides, forces the generation model to logically deduce according to the following causal chain: [Environmental Awareness Layer]: Explicitly declares the currently perceived environmental physical indicators.

[0074] [Strategic Reflection Layer]: Explains the technical reasons for rejecting a specific solution based on punishment mechanisms and historical feedback.

[0075] [Decision Execution Layer]: Outputs the final adaptation recommendations (final health intervention plan) sorted by priority weight.

[0076] Enhanced safety directive: Add a mandatory hard constraint directive to the system prompt in the Prompt template: "Any environmental adaptation adjustment or intensity adjustment is strictly prohibited from touching the absolute contraindication boundary defined in the Medical Knowledge Graph (KG); when environmental adaptation recommendations conflict with medical contraindications, the solution must be tailored with medical safety as the highest principle." In practical applications, a mandatory Prompt template (strictly prohibiting contraindications) is first constructed based on the medical KG boundary constraints and the pruned strategy set (output instruction set). Then, a structured intervention plan output (including environmental interpretation, empathic comfort, and execution instructions) is generated based on the logical deduction of the large model.

[0077] The method provided in this invention introduces a parameterized control model before generating a decision. By using hard constraints (physical environmental limitations) and soft constraints (emotional acceptability), the retrieved candidate solutions are reordered, filtered, and their intensity is adjusted. This ensures that the generated final health intervention plan is not only correct in medical theory, but also "executable" in the actual physical environment and the user's psychological state.

[0078] According to a context-aware adaptive health intervention method provided by the present invention, the feasibility of at least one candidate intervention action in a standard medical intervention path is determined based on environmental perception data, including: If the environmental conditions required for the execution of any candidate intervention action conflict with the environmental perception data, the candidate intervention action is determined to be an unexecutable candidate intervention action under the current environment.

[0079] Specifically, in some embodiments, hard constraint pruning is achieved through the following steps: Output of path 2 The Boolean feasibility determination function is executed on the candidate action set. , This represents the physical environment parameters required for candidate intervention actions. This represents a candidate intervention action, and the output is a Boolean value. If the environmental conditions required for the execution of any candidate intervention action conflict with the environmental perception data, the candidate intervention action is determined to be an unexecutable candidate intervention action in the current environment.

[0080] In practical applications, if the environmental constraint label does not match the physical environment parameters required for the candidate intervention action (e.g., noise level exceeds the action tolerance threshold), the candidate intervention action is marked as "disabled" and logical pruning is performed. For example, when the environment is... When the mood is..., then prune suggestions like "meditation" and "audiobooks" (hard constraint pruning); when the mood is... If the activity is deemed too strenuous, it will be downgraded to "walking" and marked as requiring additional "encouraging rhetoric" (soft constraint pruning) to prevent physical unenforceability.

[0081] The method provided in this invention uses a "hard constraint pruning" strategy based on a parameterized control model to automatically eliminate solutions that conflict with the current physical environment (such as not recommending hearing intervention in noisy environments). This ensures that the generated intervention recommendations are not only correct in medical theory, but also "executable" in the actual physical environment and the user's psychological state.

[0082] According to a context-aware adaptive health intervention method provided by the present invention, the execution priority weights of the remaining candidate intervention actions are dynamically adjusted based on emotion perception data and historical intervention records to obtain the adjusted candidate intervention actions, including: If the retrieved historical intervention records show that, in a situation similar to the current intervention situation, the historical feedback result of any candidate intervention action is lower than a preset threshold, the current execution priority weight of the candidate intervention action is reduced. When emotion perception data indicates that the target object is in a preset negative emotional state, the execution priority weight of candidate intervention actions that require high compliance is reduced to obtain each modified candidate intervention action.

[0083] Specifically, in some embodiments, the step of adaptively updating the soft constraint weights includes: By analyzing historical intervention records, a negative gradient signal is introduced to achieve precise convergence of user behavior preferences. If historical intervention records show that, in a situation similar to the current intervention scenario, the historical feedback result of any candidate intervention action is lower than a preset threshold (for example, if the feedback log shows that the user has rejected a certain type of intervention n times consecutively under a similar Context Descriptor (ICD), the system will activate a negative gradient penalty mechanism), reducing the current execution priority weight of the candidate intervention action.

[0084] By analyzing emotion perception data, when the data indicates that the target audience is in a pre-defined negative emotional state, such as "extreme fatigue" or "anxiety," the execution priority weight of candidate intervention actions requiring high compliance (such as "long-distance walking") is reduced, resulting in the final health intervention plan. This avoids causing user resistance by providing high-intensity exercise suggestions.

[0085] The method provided in this invention dynamically adjusts the priority weights of candidate intervention actions based on emotion vectors and historical intervention records, thereby solving the problem of low compliance with general recommendations in medical knowledge graphs.

[0086] According to the present invention, a context-aware adaptive health intervention method dynamically adjusts the intensity of execution parameters corresponding to each modified candidate intervention action based on emotion perception data to generate a final health intervention plan, including: Determine the degree of deviation between the emotional state represented by the emotion perception data and the preset baseline emotional state; Based on the degree of deviation, a scaling factor is determined to adjust the intensity of the intervention; The standard execution parameters of each modified candidate intervention action are reduced using a scaling factor to obtain execution parameters that are adapted to the current emotional state. Based on the various execution parameters, a final health intervention plan is generated.

[0087] Specifically, in some embodiments, the parameter strength degradation process is achieved through the following steps: First, determine the degree of deviation between the emotional state represented by the emotion perception data and the preset baseline emotional state. Based on the degree of deviation, determine a scaling factor to adjust the intensity of the intervention. .

[0088] For example, if the emotion vector represented by the emotion perception data is within the range of this preset baseline emotional state, the execution coefficient... When the emotion vector (such as anxiety or fatigue) exceeds a critical threshold, the system automatically calculates the degree of deviation and determines a scaling factor using a preset linear interpolation mapping function or piecewise function. This, in turn, affects the intensity of the original action (such as time) from the medical pathway. ,frequency (etc.) Perform linear degradation operation, that is, use scaling factor to reduce the standard execution parameters of each modified candidate intervention action to obtain execution parameters that are adapted to the current emotional state.

[0089] in, Indicates standard execution parameters, This represents the execution parameters that adapt to the current emotional state. This represents the dynamic scaling factor that ultimately applies to the intervention plan, and its value ranges from... It floats within a continuous range and is used to downgrade the original intervention intensity (such as exercise duration and frequency) in real time; This indicates the preset emotional sensitivity attenuation coefficient (usually set). This is used to adjust the system's response sensitivity to fluctuations in user status. The higher the value, the more sensitive the system is to negative emotions, and the faster the intensity decays; Represents the current sentiment vector perceived in real time. With user-personalized baseline state vector The Euclidean distance between them is used to quantify the degree to which the current physical and mental state deviates from the standard comfort range; This represents the maximum theoretical distance normalization factor in the feature vector space, used as the denominator to map the absolute distance value to a relative deviation ratio, ensuring dimensionless calculation. Specifically, it sets... The lower bound is 0.1 instead of 0 to ensure that the intervention plan retains the "minimum feasible action." Even when the user's condition is extremely poor, the system encourages the user to perform small amounts of health-related activities (such as "exercising for 3 minutes"), rather than completely abandoning the intervention, thereby maintaining the continuity of behavioral habits. For example, if the user is feeling down, It might become 0.3, with the final execution parameter being 30 minutes × 0.3 = 9 minutes. This step dynamically switches the intervention recommendation from "standard mode" to "low-burden mode".

[0090] Finally, based on the revised execution parameters, a final health intervention plan is generated.

[0091] The method provided in this invention introduces an emotional sensitivity attenuation coefficient k and a linear interpolation function. When a user's emotional vector deviates from the baseline comfort range (e.g., extreme fatigue), a scaling factor is automatically calculated. The intensity of implementation of medical standard recommendations (such as exercise duration and frequency) is linearly downgraded to reduce the psychological burden on users while ensuring minimum feasibility.

[0092] According to the context-aware adaptive health intervention method provided by the present invention, the method further includes: Output the final health intervention plan and obtain the execution feedback data of the target subjects after implementing the final health intervention plan; the final health intervention plan is structured data and includes at least environmental adaptability explanation information to explain the reasons for the plan adjustment, emotional comfort information for empathy, and specific execution instructions; the execution feedback data includes objective physiological indicator change data and / or user subjective satisfaction rating data; The current intervention scenario, the final health intervention plan, and the implementation feedback data are treated as a new historical intervention record; New historical intervention records are stored in the experience base to update it.

[0093] Specifically, in some embodiments, the method further includes steps of closed-loop feedback and dynamic evolution of the experience base. Closed-loop feedback and dynamic evolution of the experience base are used to achieve "long-term effective intervention" of the system, which is a key difference from a one-time question-and-answer system and reflects the system's long-term self-learning capability. This process includes the following steps: First, the final health intervention plan is output, which consists of the following three structured parts: 1. Environmental compatibility explanation information used to explain the reasons for the solution adjustment: Inform the user why the adjustment was made based on the current environment (e.g., "The current light is too dim, so we have switched you to auditory guidance mode"). 2. Emotional comforting information for empathy: Based on the identified emotional state, generate empathetic messages (such as "We detected that you are a little tired, so we have appropriately reduced the intensity of your exercise...") to enhance the user's emotional receptiveness; 3. Specific execution instructions (final execution instructions): Specific action suggestions after parameter pruning and weight adjustment.

[0094] Simultaneously, it acquires real-time execution feedback data after the target subject implements the final health intervention plan. This feedback data includes objective execution data such as changes in the target subject's physiological indicators (e.g., whether heart rate variability (HRV) has improved) or behavioral actions (e.g., whether standing up was detected). It also collects explicit feedback from the target subject through the terminal interface (App pop-ups or voice assistant queries) (e.g., "Did you find the previous suggestion helpful?" rating 1-5).

[0095] Furthermore, the current intervention scenario, the final health intervention plan, and the implementation feedback data (i.e., the "scenario-intervention-feedback" triple) are stored as a new historical intervention record in the experience base to update the experience base and achieve long-term self-learning capability.

[0096] The method provided in this invention constructs a strategy self-evolution capability based on a long-term feedback closed loop, realizing a transformation from "one-time advice" to "long-term dynamic companionship." This invention changes the traditional system's unidirectional working mode of "generation ends," establishing an experience backtracking mechanism based on the "Context-Intervention-Feedback (SAR)" triple. This enables the system to possess "memory" and "learning" capabilities similar to a family doctor, continuously and accurately capturing and adapting to the dynamic drift of users' personalized preferences as interaction time increases, thereby providing long-term, precise health management services.

[0097] According to a context-aware adaptive health intervention method provided by the present invention, updating the experience base includes: If the feedback data indicates that the final health intervention plan is successful, the association vector established based on the current intervention situation and the executed actions will be moved by a preset step size towards the successful cluster center in the vector space corresponding to the experience base. If the feedback data indicates that the final health intervention plan has failed, the associated vector in the vector space will be moved by a preset step in the opposite direction to the successful cluster center.

[0098] Specifically, in some embodiments, the step of updating the experience base includes: To determine whether the current final health intervention plan has been successful, we specifically rely on real-time execution feedback data (changes in physiological indicators, behavioral actions, and explicit feedback from the target population).

[0099] Furthermore, if the feedback data indicates that the final health intervention plan was successful, the correlation vector (i.e., the "context-action" correlation vector) established based on the current intervention context and the executed action will be moved by a preset step size towards the successful cluster center in the vector space corresponding to the experience base. Successful cluster centers include historical intervention records of multiple successful interventions under similar ICD scenarios.

[0100] If the feedback data indicates that the final health intervention plan has failed, the correlation vector in the vector space will be moved in the opposite direction of the successful cluster centers by a preset step size. For example, applying a repulsive force in the vector space can reduce its Euclidean distance ranking under similar context ICD, thereby achieving convergence of the system to the user's personalized preferences.

[0101] The method provided in this invention can automatically reduce the weight of user-rejected solutions and converge towards the cluster center of successful solutions based on the user's historical performance feedback (objective physiological indicators and subjective scores). This enables the system to possess "memory" and "learning" capabilities similar to a family doctor, allowing it to continuously and accurately capture and adapt to the dynamic shifts in user's personalized preferences as interaction time increases, thereby providing long-term and precise health management services.

[0102] Figure 2 This is the second flowchart of the context-aware adaptive health intervention method provided by the present invention, as shown below. Figure 2 As shown, the method includes: Acquiring multimodal contextual data: Acquiring environmental perception data (Internet of Things (IoT) devices: noise, light, temperature, humidity, etc.), emotion perception data (Affective computing module: facial expressions, speech, and text features) and intervention history feedback (Wearable devices / terminals: physiological indicators HRV, subjective scores).

[0103] Step 1: Multimodal context feature perception and vectorization module.

[0104] Data denoising and normalization, parameterized mapping (based on a pre-defined mapping table T, such as...) → And generate intervention situation descriptors (ICDs) that combine environmental labels, emotion vectors, and feedback parameters.

[0105] Step 2: Hybrid retrieval system (based on knowledge graph and experience backtracking).

[0106] Path 1: Medical professional knowledge flow (hard constraint).

[0107] Based on medical knowledge graph (KG) retrieval, the standard treatment principles and contraindication boundaries are output.

[0108] Path 2: Contextual experience feedback flow (dynamic adaptation).

[0109] Construct a temporal heterogeneous behavior graph database (“context-intervention-feedback” triples), perform K-NN similarity retrieval based on ICD, and recall sets of similar contextual interaction records. .

[0110] Step 3: Strategy regulation and pruning based on parameterized control model.

[0111] Hard constraint pruning (Boolean decision) (Eliminating conflicting actions); soft-constraint weight updates (negative gradient penalty, ); and parameter intensity degradation (emotional adaptation, calculation of scaling factor). , ).

[0112] Step 4: Constrained Structured Intervention Schema Generation (LLM).

[0113] A mandatory Prompt template (strictly prohibiting contraindications) is constructed based on medical KG boundary constraints and a pruned strategy set; the large model is logically deduced to generate a structured intervention plan output (environmental interpretation, empathic comfort, and execution instructions). This step corresponds to user terminal execution and monitoring (stage).

[0114] Step 5: Closed-loop feedback and dynamic evolution of the experience base.

[0115] Collect feedback data from step 4, which is the actual execution data collection (completion rate, satisfaction level); conduct feedback evaluation based on the actual execution data (success / failure); If successful, it is considered positive feedback, and the vector moves towards the successful cluster center; If it fails, it is considered negative feedback, which applies a vector repulsion force (lowers the ranking). The experience base is updated in real time based on the feedback results (returning to path 2 in step 2).

[0116] The data in the experience base in step 2 is also entered into the user terminal for execution and monitoring, followed by feedback evaluation and updating of the experience base.

[0117] The retrieval-enhanced generative intervention decision-making system and method based on environmental and emotion perception proposed in this invention structures unstructured community environmental physical quantities and the emotional and psychological states of the elderly into decision variables, and introduces retrieval-enhanced generative (RAG) and closed-loop feedback processes. Compared with existing technologies, it has the following significant technical advantages: 1. It achieves high executability and dynamic adaptability of intervention decision-making schemes in real-life scenarios; 2. A strategy self-evolution capability based on a long-term feedback loop has been constructed, realizing the transformation from "one-time suggestions" to "long-term dynamic companionship"; 3. Through a “flexible” parameter adjustment mechanism, a balance is effectively struck between medical rigor and user compliance.

[0118] The core technical concept of this application lies in constructing a Retrieval Enhancement Generation (RAG) intervention decision-making system that integrates environmental physical quantities and user emotional state perception. By introducing a closed-loop mechanism based on execution feedback, and utilizing "context-intervention-feedback" data to drive the dynamic evolution of strategy weights, personalized long-term adaptive intervention is achieved. To realize this core concept, a specific implementation method for each step is described in the detailed embodiments. However, those skilled in the art should understand that, to achieve the same inventive objective, the following core technical features and their implementation methods have various alternatives or variations, all of which should fall within the protection scope of this invention: 1. Core Feature 1: Perception and Quantitative Representation of Multimodal Context Features (corresponding to Step 1 and Step 2).

[0119] Alternative Solution 1 (Sensing Source and Data Acquisition Method): This invention is not limited to relying on a specific combination of "IoT devices + multimodal emotion computing" to acquire data. and Any technical means that can acquire information about the environmental state and the user's psychological state can be used.

[0120] (a) Environmental data: In addition to dedicated IoT devices, temperature, humidity and air quality can be obtained by using built-in sensors (microphones, light sensors) of general smart terminals (mobile phones, smart speakers) or publicly available Internet data based on geographic location, such as real-time weather application programming interfaces (APIs).

[0121] (b) Emotional data (physiological / explicit): not limited to multimodal fusion of visual, speech, and text. In privacy-restricted scenarios, only a single modality (such as text semantic analysis only) may be used; or contact-based physiological sensors (such as EDA and EEG) may be used as input sources for emotion computing.

[0122] (c) Emotional Data (Implicit / Implicit Inference): To adapt to pure software application scenarios (such as mobile apps) where users do not wear sensors or refuse to turn on their cameras, the system can establish an implicit emotion inference model based on human-computer interaction behavior characteristics. Input characteristics include, but are not limited to: interaction timing characteristics: such as the delay time of user responses to the system, typing speed or keystroke intervals for text input; touch operation characteristics: such as the pressure value of screen buttons (using pressure touch technology), the smoothness of touch trajectory or the amplitude of jitter; editing behavior characteristics: such as the frequency of deletion and undoing during the input process (used to infer the user's hesitation, anxiety or cognitive load). This type of solution does not rely on external hardware and can achieve coarse-grained estimation of emotional state based solely on terminal interaction logs.

[0123] Alternative Solution 2 (Feature Mapping and ICD Construction Logic): This invention is not limited to using a pre-defined "feature mapping table". The intervention context descriptor (ICD) is generated using a simple "tensor splicing" method.

[0124] Mapping logic: Fuzzy logic can be used to map continuous numerical values ​​to membership degrees instead of hard discrete labels; or end-to-end deep neural network encoders (such as Transformer-based Encoders) can be used to directly embed the original environmental and emotional data into high-dimensional latent space vectors, replacing explicit label splicing.

[0125] Alternative Option 3 (Retrieval Strategy): In a hybrid retrieval system, the approach is not limited to using cosine similarity and K-nearest neighbor search (K-NN). Any metric capable of measuring contextual similarity can be used (such as Euclidean distance, Manhattan distance, or semantic distance learned from a neural network). The retrieval target is also not limited to "temporally heterogeneous behavioral graphs" but can be flattened records in a vector database or subgraph structures based on graph neural network (GNN) reasoning.

[0126] 2. Core Feature Two: Strategy Regulation Based on Parametric Control Model (corresponding to Step 3).

[0127] Alternative Solution 1 (Hard Constraint Pruning Logic): This invention is not limited to using Boolean decision functions. Perform hard pruning based on a "black and white" approach. This can be achieved using probabilistic models or soft masking mechanisms, which significantly reduce the probability distribution of conflicting actions in the environment rather than completely eliminating them, thus preserving the possibility of execution in special circumstances.

[0128] Alternative Solution 2 (Soft Constraints and Strength Degradation Algorithm): This invention is not limited to using a specific linear interpolation formula to calculate the scaling factor.

[0129] (a) Functional form: Nonlinear functions (such as Sigmoid, exponential decay function, logarithmic function) can be used to describe the effect curve of emotion change on executive strength.

[0130] (b) Degradation target: The controlled parameters are not limited to "exercise duration" or "frequency", but can be extended to any quantifiable performance attribute of the intervention content, such as "difficulty level", "interactive speech rate", "multimedia volume".

[0131] Alternative Solution 3 (Weight Update Mechanism): Regarding Negative Gradient Penalty This is not limited to linear deduction. Q-learning or Policy Gradient algorithms from reinforcement learning can be introduced to dynamically adjust the Q-value or probability weights of the policy based on long-term returns; or a forgetting factor can be introduced to cause the weights of long-term negative feedback to decay over time.

[0132] 3. Core Feature Three: Constrained Generation and Closed-Loop Evolution (corresponding to steps 4 and 5).

[0133] Alternative Solution 1 (Generative Model and Constraint Method): This is not limited to using a general large language model (LLM) in conjunction with a Prompt project. It can utilize specialized small models fine-tuned with medical domain knowledge, or rule-based template-based generative systems. Constraint mechanisms can also be implemented by introducing a Logit Processor during the model decoding stage, directly masking taboo words or dangerous suggestions at the underlying probability distribution level, rather than relying solely on Prompt hints.

[0134] Alternative Option 2 (Feedback and Evolution): The evolution of the experience base is not limited to "movement and repulsion in vector space". Evolution method: "Continual Learning" or LoRA (Low-Rank Adaptation) techniques can be used to periodically fine-tune and update the parameters of the model itself using feedback data, internalizing the experience into the model parameters, rather than just adjusting the ranking of search results.

[0135] The description of the aforementioned core technical features and their alternatives aims to encompass various possibilities for realizing the core concept of this invention, thereby defining a broader scope of protection. Combinations or further improvements made by those skilled in the art based on these alternatives, as long as their essence remains the same—driving retrieval enhancement through environmental and emotional perception and incorporating feedback loops for intervention decisions—should be considered to fall within the scope of this invention.

[0136] The context-aware adaptive health intervention device provided by the present invention will be described below. The context-aware adaptive health intervention device described below can be referred to in correspondence with the context-aware adaptive health intervention method described above.

[0137] Figure 3This is a schematic diagram of the structure of the context-aware adaptive health intervention device provided by the present invention, as shown below. Figure 3 As shown, the context-aware adaptive health intervention device 300 includes the following modules: The intervention situation descriptor construction module 310 is used to construct an intervention situation descriptor to represent the current intervention situation based on the multimodal situation data of the target object; the multimodal situation data includes environmental perception data and emotion perception data; The hybrid retrieval module 320 is used to retrieve at least one historical intervention record similar to the current intervention situation from the experience base, using the intervention situation descriptor as the query vector. The experience base pre-stores multiple historical intervention records, which include historical situation data, historical intervention plans corresponding to the historical situation data, and historical feedback results. Based on the physiological sign data of the target object, a retrieval is performed in the medical knowledge base to obtain the standard medical intervention path for the target object. The strategy control module 330 is used to control the standard medical intervention path based on the historical feedback results in each of the retrieved historical intervention records and the emotion perception data, and generate a final health intervention plan that is adapted to the current intervention situation.

[0138] The apparatus provided in this embodiment of the invention includes multimodal situational data, including environmental perception data and emotion perception data. An intervention situation descriptor construction module 310 is used to construct an intervention situation descriptor representing the current intervention situation based on the multimodal situational data of the target object. A hybrid retrieval module 320 is used to retrieve at least one historical intervention record similar to the current intervention situation from an experience base using the intervention situation descriptor as a query vector. Multiple historical intervention records are pre-stored in the experience base, and each historical intervention record includes historical situational data, corresponding historical intervention plans, and historical feedback results. Further, a retrieval is performed in a medical knowledge base based on the physiological characteristics data of the target object to obtain a standard medical intervention path for the target object. A strategy control module 330 is used to control the standard medical intervention path according to the historical feedback results and emotion perception data in each retrieved historical intervention record, generating a final health intervention plan adapted to the current intervention situation.

[0139] This invention overcomes the limitations of existing medical RAG systems, which rely solely on textual semantics and lack environmental and emotional dimensions. By constructing intervention context descriptors, the system maps environmental physical indicators (such as noise and light) and user psychological states (such as anxiety and fatigue) into "intervention context descriptors (ICDs)." Based on the intervention context descriptors (ICDs), a hybrid retrieval is performed to obtain candidate historical intervention records. Then, based on the historical feedback results from each historical intervention record and the aforementioned emotional perception data, the standard medical intervention path is adjusted to generate a final health intervention plan adapted to the current intervention context. This ensures that the generated final health intervention plan is not only theoretically correct but also "executable" in the actual physical environment and user psychological state, thereby achieving high executability and dynamic adaptability of the intervention decision-making plan in real-life scenarios.

[0140] According to the present invention, a context-aware adaptive health intervention device 300 is provided, wherein the intervention context descriptor construction module 310 is specifically used for: The environmental perception data is feature-mapped and converted into at least one discrete environmental state label. The emotion perception data is quantified to generate an emotion state vector; The environmental state labels, the emotional state vector, and the feedback parameters extracted from the historical intervention feedback data are fused to generate the intervention situation descriptor.

[0141] According to the present invention, a context-aware adaptive health intervention device 300 is provided, wherein the strategy control module 330 is specifically used for: Based on the environmental perception data, the feasibility of at least one candidate intervention action in the standard medical intervention path is determined, and candidate intervention actions that are not feasible in the current environment are eliminated to obtain the remaining candidate intervention actions. Based on the emotion perception data and the retrieved historical intervention records, the execution priority weights of the remaining candidate intervention actions are dynamically adjusted to obtain the adjusted candidate intervention actions. Based on the emotion perception data, the intensity of the execution parameters corresponding to each of the modified candidate intervention actions is dynamically adjusted to generate the final health intervention plan.

[0142] According to the context-aware adaptive health intervention device 300 provided by the present invention, the strategy control module 330 is further configured to: If the environmental conditions required to execute any of the candidate intervention actions conflict with the environmental perception data, the candidate intervention action is determined to be an unexecutable candidate intervention action under the current environment.

[0143] According to the context-aware adaptive health intervention device 300 provided by the present invention, the strategy control module 330 is further configured to: If the retrieved historical intervention records show that, in a situation similar to the current intervention situation, the historical feedback result of any candidate intervention action is lower than a preset threshold, the current execution priority weight of the candidate intervention action is reduced. When the emotion perception data indicates that the target object is in a preset negative emotional state, the execution priority weight of the candidate intervention action that requires high compliance is reduced to obtain each of the modified candidate intervention actions.

[0144] According to the context-aware adaptive health intervention device 300 provided by the present invention, the strategy control module 330 is further configured to: Determine the degree of deviation between the emotional state represented by the emotional perception data and the preset baseline emotional state; Based on the degree of deviation, a scaling factor is determined to adjust the intensity of the intervention; The scaling factor is used to reduce the standard execution parameters of each of the modified candidate intervention actions to obtain execution parameters that are adapted to the current emotional state. Based on the execution parameters, the final health intervention plan is generated.

[0145] According to the present invention, a context-aware adaptive health intervention device 300 is provided, the device further comprising an update module; The update module is specifically used for: Output the final health intervention plan and obtain the execution feedback data after the target object implements the final health intervention plan; the final health intervention plan is structured data, and the final health intervention plan includes at least environmental adaptability explanation information for explaining the reasons for the plan adjustment, emotional comfort information for empathy, and specific execution instructions; the execution feedback data includes objective physiological indicator change data and / or user subjective satisfaction rating data; The current intervention scenario, the final health intervention plan, and the execution feedback data are treated as a new historical intervention record; The new historical intervention records are stored in the experience base to update the experience base.

[0146] According to a context-aware adaptive health intervention device 300 provided by the present invention, the updating module is further configured to: If the execution feedback data indicates that the final health intervention plan is successful, the association vector established based on the current intervention situation and the executed action will be moved by a preset step size towards the successful cluster center in the vector space corresponding to the experience base. If the execution feedback data indicates that the final health intervention plan has failed, the correlation vector in the vector space is moved by the preset step size in the opposite direction to the successful cluster center.

[0147] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 4 As shown, the electronic device may include: a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a context-aware adaptive health intervention method, which includes: Based on the multimodal contextual data of the target object, an intervention context descriptor is constructed to represent the current intervention context; the multimodal contextual data includes environmental perception data and emotion perception data; Using the intervention scenario descriptor as the query vector, at least one historical intervention record similar to the current intervention scenario is retrieved from the experience base; the experience base stores multiple historical intervention records in advance, and the historical intervention records include historical scenario data, historical intervention plans corresponding to the historical scenario data, and historical feedback results; Based on the physiological signs data of the target object, a search is performed in the medical knowledge base to obtain the standard medical intervention path for the target object; Based on the historical feedback results in the retrieved historical intervention records and the emotional perception data, the standard medical intervention path is adjusted to generate a final health intervention plan adapted to the current intervention situation.

[0148] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0149] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the context-aware adaptive health intervention method provided by the above methods, the method comprising: Based on the multimodal contextual data of the target object, an intervention context descriptor is constructed to represent the current intervention context; the multimodal contextual data includes environmental perception data and emotion perception data; Using the intervention scenario descriptor as the query vector, at least one historical intervention record similar to the current intervention scenario is retrieved from the experience base; the experience base stores multiple historical intervention records in advance, and the historical intervention records include historical scenario data, historical intervention plans corresponding to the historical scenario data, and historical feedback results; Based on the physiological signs data of the target object, a search is performed in the medical knowledge base to obtain the standard medical intervention path for the target object; Based on the historical feedback results in the retrieved historical intervention records and the emotional perception data, the standard medical intervention path is adjusted to generate a final health intervention plan adapted to the current intervention situation.

[0150] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the context-aware adaptive health intervention method provided by the methods described above, the method comprising: Based on the multimodal contextual data of the target object, an intervention context descriptor is constructed to represent the current intervention context; the multimodal contextual data includes environmental perception data and emotion perception data; Using the intervention scenario descriptor as the query vector, at least one historical intervention record similar to the current intervention scenario is retrieved from the experience base; the experience base stores multiple historical intervention records in advance, and the historical intervention records include historical scenario data, historical intervention plans corresponding to the historical scenario data, and historical feedback results; Based on the physiological signs data of the target object, a search is performed in the medical knowledge base to obtain the standard medical intervention path for the target object; Based on the historical feedback results in the retrieved historical intervention records and the emotional perception data, the standard medical intervention path is adjusted to generate a final health intervention plan adapted to the current intervention situation.

[0151] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0152] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A context-aware adaptive health intervention method, characterized in that, include: Based on the multimodal contextual data of the target object, an intervention context descriptor is constructed to represent the current intervention context; the multimodal contextual data includes environmental perception data and emotion perception data; Using the intervention scenario descriptor as the query vector, at least one historical intervention record similar to the current intervention scenario is retrieved from the experience base; the experience base stores multiple historical intervention records in advance, and the historical intervention records include historical scenario data, historical intervention plans corresponding to the historical scenario data, and historical feedback results; Based on the physiological signs data of the target object, a search is performed in the medical knowledge base to obtain the standard medical intervention path for the target object; Based on the historical feedback results in each of the retrieved historical intervention records and the emotional perception data, the standard medical intervention path is adjusted to generate a final health intervention plan that is adapted to the current intervention situation. The step of adjusting the standard medical intervention path based on historical feedback results from the retrieved historical intervention records and the emotion perception data to generate a final health intervention plan adapted to the current intervention context includes: determining the feasibility of at least one candidate intervention action in the standard medical intervention path based on the environmental perception data, and eliminating candidate intervention actions that are not feasible in the current environment to obtain the remaining candidate intervention actions; dynamically adjusting the execution priority weight of each remaining candidate intervention action based on the emotion perception data and the retrieved historical intervention records to obtain the adjusted candidate intervention actions; and adjusting the execution priority weight of each remaining candidate intervention action based on the emotion perception data. The process of dynamically adjusting the intensity of execution parameters corresponding to the modified candidate intervention actions to generate the final health intervention plan, based on the emotion perception data, includes: determining the degree of deviation between the emotional state represented by the emotion perception data and a preset benchmark emotional state; determining a scaling factor for adjusting the intervention intensity based on the degree of deviation; reducing the standard execution parameters of each modified candidate intervention action using the scaling factor to obtain execution parameters adapted to the current emotional state; and generating the final health intervention plan based on each execution parameter.

2. The context-aware adaptive health intervention method according to claim 1, characterized in that, The multimodal context data also includes historical intervention feedback data; the multimodal context data based on the target object constructs an intervention context descriptor to characterize the current intervention context, including: The environmental perception data is feature-mapped and converted into at least one discrete environmental state label. The emotion perception data is quantified to generate an emotion state vector; The environmental state labels, the emotional state vector, and the feedback parameters extracted from the historical intervention feedback data are fused to generate the intervention situation descriptor.

3. The context-aware adaptive health intervention method according to claim 1, characterized in that, The step of determining the feasibility of at least one candidate intervention action in the standard medical intervention pathway based on the environmental perception data includes: If the environmental conditions required to execute any of the candidate intervention actions conflict with the environmental perception data, the candidate intervention action is determined to be an unexecutable candidate intervention action under the current environment.

4. The context-aware adaptive health intervention method according to claim 1, characterized in that, The step involves dynamically adjusting the execution priority weights of the remaining candidate intervention actions based on the emotion perception data and the retrieved historical intervention records, resulting in adjusted candidate intervention actions, including: If the retrieved historical intervention records show that, in a situation similar to the current intervention situation, the historical feedback result of any candidate intervention action is lower than a preset threshold, the current execution priority weight of the candidate intervention action is reduced. When the emotion perception data indicates that the target object is in a preset negative emotional state, the execution priority weight of the candidate intervention action that requires high compliance is reduced to obtain each of the modified candidate intervention actions.

5. The context-aware adaptive health intervention method according to any one of claims 1-4, characterized in that, The method further includes: Output the final health intervention plan and obtain the execution feedback data after the target object implements the final health intervention plan; the final health intervention plan is structured data, and the final health intervention plan includes at least environmental adaptability explanation information for explaining the reasons for the plan adjustment, emotional comfort information for empathy, and specific execution instructions; the execution feedback data includes objective physiological indicator change data and / or user subjective satisfaction rating data; The current intervention scenario, the final health intervention plan, and the execution feedback data are treated as a new historical intervention record; The new historical intervention records are stored in the experience base to update the experience base.

6. The context-aware adaptive health intervention method according to claim 5, characterized in that, The updating of the experience base includes: If the execution feedback data indicates that the final health intervention plan is successful, the association vector established based on the current intervention situation and the executed action will be moved by a preset step size towards the successful cluster center in the vector space corresponding to the experience base. If the execution feedback data indicates that the final health intervention plan has failed, the correlation vector in the vector space is moved by the preset step size in the opposite direction to the successful cluster center.

7. A context-aware adaptive health intervention device, characterized in that, include: An intervention context descriptor construction module is used to construct an intervention context descriptor to represent the current intervention context based on the multimodal context data of the target object; the multimodal context data includes environmental perception data and emotion perception data; The hybrid retrieval module is used to retrieve at least one historical intervention record similar to the current intervention situation from the experience base, using the intervention situation descriptor as the query vector. The experience base stores multiple historical intervention records in advance, and each historical intervention record includes historical situation data, historical intervention plans corresponding to the historical situation data, and historical feedback results. Based on the physiological characteristics data of the target object, a retrieval is performed in the medical knowledge base to obtain the standard medical intervention path for the target object. The strategy control module is used to control the standard medical intervention path based on the historical feedback results in each of the retrieved historical intervention records and the emotion perception data, and generate a final health intervention plan that is adapted to the current intervention situation. The strategy control module is specifically used to determine the feasibility of at least one candidate intervention action in the standard medical intervention path based on the environmental perception data, and to eliminate candidate intervention actions that are not executable in the current environment, so as to obtain the remaining candidate intervention actions. Based on the emotion perception data and the retrieved historical intervention records, the execution priority weights of the remaining candidate intervention actions are dynamically adjusted to obtain the adjusted candidate intervention actions. Based on the emotion perception data, the intensity of the execution parameters corresponding to each of the modified candidate intervention actions is dynamically adjusted to generate the final health intervention plan. The step of dynamically adjusting the intensity of the execution parameters corresponding to each of the modified candidate intervention actions based on the emotion perception data to generate the final health intervention plan includes: determining the degree of deviation between the emotional state represented by the emotion perception data and a preset benchmark emotional state; determining a scaling factor for adjusting the intervention intensity based on the degree of deviation; reducing the standard execution parameters of each of the modified candidate intervention actions using the scaling factor to obtain execution parameters adapted to the current emotional state; and generating the final health intervention plan based on each execution parameter.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the context-aware adaptive health intervention method as described in any one of claims 1 to 6.

Images