Sleep health dynamic assessment method and system based on user feedback driving, and medium

Abstract

The present application relates to the field of information technology, in particular to a sleep health dynamic evaluation method and system based on user feedback driving and a medium. The method comprises: constructing an individual knowledge graph reflecting individual health causal relationship by using time series causal analysis method; constructing a group knowledge graph reflecting group health causal relationship by combining agglomerative hierarchical clustering method and hierarchical federated learning method; triggering individual knowledge graph update and triggering group knowledge graph delayed incremental reconstruction based on real-time acquired user feedback information and health factor data; dynamically fusing the weight of the updated double graph based on the cumulative number of user feedback information to obtain the comprehensive weight of the user health factor, and then performing weighted calculation on the newly acquired health factor data to obtain the sleep health evaluation score. Thus, the real-time feedback data is deeply utilized to realize dynamic, accurate and interpretable sleep health evaluation, and the technical prejudice that "group and individual are irreconcilable" is broken through.

Description

Technical Field

[0001] This invention relates to the field of information technology, and specifically to a method, system, and medium for dynamic assessment of sleep health based on user feedback. Background Technology

[0002] With the rapid development of smart home technology, sleep health assessment technology has also seen rapid growth in the field of smart health management in recent years, and is widely used in wearable devices, smart mattresses, and related mobile applications. Currently, existing health assessment technologies have achieved some success in automated monitoring. For example, Chinese patent CN118098545A, a big data-driven method for assessing the health status of middle-aged and elderly patients, shows progress in device adaptability by dynamically adjusting wearable device usage time through habit analysis and physiological indicator stability judgment. However, this patent is essentially still a one-way monitoring mode. Its core flaw lies in the disconnect between group profiles and individual characteristics. Specifically, this patent only integrates physiological data collected by the device, ignoring semantic information from user feedback. User acceptance surveys are only used to adjust device usage time. Furthermore, the health prediction model is based on binary classification (normal / abnormal) using support vector machines, relying on predefined threshold alarms, and cannot correlate with real-time changes in health status. In reality, it still lacks the ability to predict potential risks, and the assessment results lack interpretability.

[0003] For example, Chinese patent CN119851951A, a menopausal syndrome health management platform based on big data, integrates multi-source data (marker monitoring, health profile modeling, etc.) to achieve more accurate prediction of menopausal health status and trend visualization, and improves the accuracy of assessing the severity of menopausal symptoms. However, the patent's integrated traditional Chinese and Western medicine solution only matches constitution and physiological indicators through static association tables, lacking dynamic association analysis and failing to capture individual dynamic changes. Moreover, the traditional federated learning used in this patent is only used for data privacy protection, ignoring the importance of subjective feedback, resulting in a separation between subjective experience and objective data. At the same time, it relies on pre-trained reinforcement learning models to generate intervention plans, and model updates require manual triggering of retraining, making it unable to respond to changes in user behavior in real time.

[0004] In summary, existing health assessment technologies still have the following technical problems:

[0005] 1. Lack of subjective user feedback limits assessment accuracy: Existing technologies rely on one-way data collection, but ignore the semantic value of user-initiated feedback, which can easily lead to misjudgment of inter-class similarity and bias in health risk attribution.

[0006] 2. Insufficient personalization and dynamic adaptability: The device strategy adjustment in patent CN118098545A (such as reducing the duration of remote monitoring) and the health profile modeling (Transformer-LSTM-tSNE process) in patent CN119851951A cannot achieve real-time weight optimization. After long-term use, the technical problem of "individual feature passivation" will inevitably occur - that is, the model cannot adapt to the evolution of user behavior, thereby reducing the accuracy of intervention.

[0007] 3. Separation of group and individual knowledge: Traditional solutions assume that group models and individual models are substitutes. Patent CN118098545A only uses SVM for classification, while patent CN119851951A relies entirely on federated learning. Summary of the Invention

[0008] To address the aforementioned technical issues, this invention proposes a method, system, and medium for dynamic sleep health assessment based on user feedback. The aim is to make deep use of real-time feedback data to achieve dynamic, accurate, and interpretable sleep health assessment, overcoming the technical bias of "incompatibility between the group and the individual."

[0009] Firstly, this application provides a user-feedback-driven dynamic assessment method for sleep health, including the following steps:

[0010] Using time series causal analysis, an individual knowledge graph reflecting the causal relationships of individual health is constructed;

[0011] By combining agglomerative hierarchical clustering and hierarchical federated learning methods, a group knowledge graph reflecting the causal relationships of group health is constructed.

[0012] Real-time acquisition of user feedback and health factor data;

[0013] Based on newly acquired user feedback information and health factor data, the individual knowledge graph is updated to reflect changes in the causal relationship of the user's individual health in real time, and the weights of the updated individual knowledge graph are obtained.

[0014] Based on newly acquired user feedback information and health factor data, a delayed incremental reconstruction of the group knowledge graph is triggered to obtain the weights of the updated group knowledge graph.

[0015] The weights of individual knowledge graphs and group knowledge graphs are dynamically integrated and updated based on the cumulative amount of user feedback information to obtain the comprehensive weight of user health factors.

[0016] Based on the comprehensive weight of user health factors, the newly acquired health factor data is weighted and calculated to obtain the user's sleep health assessment score.

[0017] In some embodiments, a time-series causal analysis method is used to construct an individual knowledge graph reflecting the causal relationships of individual health, including:

[0018] Obtain users' historical health factor time-series data and abnormal event sequences;

[0019] Time series causal analysis was used to examine the causal relationship between health factors and abnormal events.

[0020] Based on the causal relationship test results between health factors and abnormal events, the weights between health factors and abnormal events are calculated as edge weights.

[0021] Based on the causal relationship test results between health factors and abnormal events, as well as the edge weights, a directed graph structure is constructed to form an individual knowledge graph.

[0022] In some embodiments, time series causal analysis is used to examine the causal relationship between health factors and abnormal events, including:

[0023] The Granger causality test algorithm was used to calculate the significance score between health factors and abnormal events;

[0024] The causal relationship between health factors and abnormal events is determined based on the significance scores between health factors and abnormal events.

[0025] In some embodiments, a group knowledge graph reflecting the causal relationships of group health is constructed by combining agglomerative hierarchical clustering methods and hierarchical federated learning methods, including:

[0026] The system acquires historical feedback information, historical health factor data, and abnormal event sequences of user groups, and performs structured transformation to obtain the health feature vector of each user.

[0027] Based on the health feature vectors of each user, an agglomerative hierarchical clustering method is used to cluster the user group, resulting in a clustering tree structure that reflects the differences in the health levels of the group.

[0028] Each leaf node of the clustered tree structure is treated as an independent participant, while the intermediate nodes and root node of the clustered tree structure are treated as coordinators, forming a hierarchical federated architecture.

[0029] Based on a hierarchical federated architecture, hierarchical federated learning is performed on the health feature vectors of each user, and a group knowledge graph reflecting the causal relationship of group health is constructed based on the federated learning results.

[0030] In some embodiments, hierarchical federated learning is performed on the health feature vectors of each user based on a hierarchical federated architecture, and a group knowledge graph reflecting the causal relationship of group health is constructed based on the federated learning results, including:

[0031] The participants are used as group labels in the group knowledge graph;

[0032] The coordinator distributes the health feature vectors of all users to the participating parties based on a hierarchical federated architecture;

[0033] Each participant builds an adaptive random forest model locally and incrementally trains the model based on the health feature vectors of the users in the partition.

[0034] The SHAP method is used to score the importance of health features in the trained adaptive random forest model, obtain the health feature weight vector of each participant, and transform it into the edge weight between health factors and group labels.

[0035] The health feature weight vectors of each participant are aggregated upwards based on the hierarchical federation architecture to obtain the feature weight vectors of each coordinator, forming a hierarchical weight matrix.

[0036] The coordinator counts the frequency of anomalous events among each participant and converts it into edge weights between group labels and anomalous events;

[0037] Based on the edge weights between group labels, health factors and group labels, as well as the edge weights between group labels and abnormal events, a directed graph structure is constructed to form a group knowledge graph.

[0038] In some embodiments, based on newly acquired user feedback information and health factor data, an individual knowledge graph is triggered to reflect changes in the causal relationships of a user's individual health in real time, resulting in the weights of the updated individual knowledge graph, including:

[0039] Based on the newly acquired health factor data, the weights between health factors and abnormal events are calculated using the method described above, and are denoted as the weight increment based on the new data.

[0040] A time decay factor is used to decay the edge weights between historical health factors and abnormal events, and the weight increment based on new data is superimposed to obtain the first weight.

[0041] Based on the newly acquired user feedback information, the frequency of user feedback information is converted into a feedback intensity coefficient, and the first weight is increased based on the feedback intensity coefficient to obtain the second weight, which is used as the weight of the updated individual knowledge graph.

[0042] In some embodiments, based on newly acquired user feedback information and health factor data, a delayed incremental reconstruction of the group knowledge graph is triggered to obtain the updated weights of the group knowledge graph, including:

[0043] The newly acquired user feedback information and health factor data are structurally transformed to obtain newly acquired health feature vectors;

[0044] Determine whether the newly acquired health feature vector can be classified into any leaf node of the clustering tree structure.

[0045] If a newly acquired health feature vector can be categorized into a leaf node of a clustering tree structure, the coordinator assigns it to the corresponding participant based on a hierarchical federated architecture. When the number of newly acquired health feature vectors received by a participant exceeds a preset threshold, the coordinator incrementally trains the participant's local adaptive random forest model using the method described above, updates the participant's health feature weight vector, aggregates it upwards, and updates the hierarchical weight matrix layer by layer to obtain the updated weights of the group knowledge graph.

[0046] If it cannot be classified into any leaf node of the clustering tree structure, then perform the following extended review steps:

[0047] Calculate the distance between the newly acquired health feature vector and the cluster center of the group. When the distance is greater than a preset distance threshold, determine the nearest neighbor intermediate node based on the distance and create a corresponding new leaf node as a new participant.

[0048] The coordinator assigns the newly acquired health feature vectors to the new participants based on a hierarchical federated architecture;

[0049] The aforementioned method is used to construct a local adaptive random forest model for the new participants. The model of the new participants is incrementally trained to complete the initialization. After calculating the health feature weight vector of the new participants, it is aggregated upwards and the hierarchical weight matrix is ​​updated layer by layer to obtain the weight of the updated group knowledge graph.

[0050] In some embodiments, the weights of the updated individual knowledge graph and the group knowledge graph are dynamically fused based on the cumulative amount of user feedback information to obtain the comprehensive weight of the user health factor, including:

[0051] It can determine in real time whether the cumulative number of user feedback messages exceeds a preset feedback threshold.

[0052] If the cumulative number of user feedback messages is less than or equal to the preset feedback threshold, then the edge weight between health factors and group tags in the updated group knowledge graph will be used as the comprehensive weight of the user's health factors.

[0053] If the cumulative number of user feedback messages exceeds a preset feedback threshold, the cumulative number of user feedback messages is converted into an individual weight fusion coefficient. The individual weight fusion coefficient is proportional to the cumulative number of user feedback messages. Based on the individual weight fusion coefficient, the weights of the updated individual knowledge graph and the weights of the group knowledge graph are weighted and summed to obtain the comprehensive weight of the user health factor.

[0054] Secondly, this application provides a user-feedback-driven dynamic sleep health assessment system, including:

[0055] The individual knowledge graph construction module is used to construct an individual knowledge graph that reflects the causal relationships of individual health using time series causal analysis methods.

[0056] The group knowledge graph construction module is used to combine agglomerative hierarchical clustering methods and hierarchical federated learning methods to construct a group knowledge graph that reflects the causal relationship of group health.

[0057] The data acquisition module is used to acquire user feedback information and health factor data in real time;

[0058] The dual-graph weight update module is used to trigger the update of the individual knowledge graph based on newly acquired user feedback information and health factor data to reflect the changes in the causal relationship of the user's individual health in real time, obtain the weight of the updated individual knowledge graph, and trigger the delayed incremental reconstruction of the group knowledge graph to obtain the weight of the updated group knowledge graph.

[0059] The weight fusion module is used to dynamically fuse the weights of the updated individual knowledge graph and the group knowledge graph based on the cumulative amount of user feedback information to obtain the comprehensive weight of the user health factor.

[0060] The sleep health assessment module is used to perform weighted calculations on newly acquired health factor data based on the comprehensive weight of user health factors, and obtain the user's sleep health assessment score.

[0061] Thirdly, a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned user feedback-driven dynamic sleep health assessment method.

[0062] The beneficial technical effects of the present invention include at least the following:

[0063] 1. This paper proposes a user-feedback-driven dynamic sleep health assessment method, system, and medium. It systematically addresses existing technical problems through a user-feedback-driven dual-graph collaborative architecture: First, it constructs a bidirectional optimized individual knowledge graph and a group knowledge graph—the health causal relationships in the individual knowledge graph provide fine-grained correlation evidence for the group knowledge graph, while the group knowledge graph provides baseline weights for cold-start users. Second, it acquires user feedback information in real time, directly integrating subjective semantic information into the assessment process. This solves the problem of limited assessment accuracy caused by the lack of subjective feedback in existing technologies (such as misjudging inter-class similarity or failing to distinguish the causes of similar signals). User-initiated data directly drives subsequent graph updates and weighted calculations. Third, it updates the individual knowledge graph weights in real time based on new data, capturing changes in user health causal relationships with millisecond-level response. Combined with delayed incremental reconstruction of the group knowledge graph (avoiding frequent computational burden while maintaining dynamic adaptability), it solves the problems of individual feature passivation and individual... The problem of insufficient personalization was addressed. Ultimately, based on the cumulative feedback, individual and group weights were dynamically integrated, transforming the group model from a "data averager" into an "individualized foundation" (using a group knowledge graph constructed through hierarchical federated learning as the foundation, and adapting to individual differences through reconstruction). At the same time, the individual model was upgraded from an "isolated analyzer" to a "group evolution driver" (real-time feedback from individual updates triggers group reconstruction, driving the evolution of the group model). Through this bidirectional data flow (individual update → group reconstruction → weight fusion) and mathematical fusion mechanism (adjusting the weight ratio according to the cumulative feedback), the mutual reinforcement of individual characteristics and group knowledge was achieved, thus overcoming the problem of separation between the group and the individual in existing technologies (such as the inability of a single model to adapt to physical differences). This achieved a technical effect of "1+1>2"—both improving efficiency through large-scale group data and achieving precise personalization through individual feedback, significantly improving the real-time performance, accuracy, and interpretability of sleep health assessment, and realizing the unity of scale and personalization in the field of sleep health assessment.

[0064] 2. Traditional sleep health assessment systems rely on static correlation analysis, which fails to capture individual-specific health causal relationships, resulting in a lack of personalized interpretability in the assessment results. To address this, this application employs a user feedback-driven dual-graph collaborative architecture design. User feedback is transformed into structured health feature vectors, constructing individual and group knowledge graphs. The individual knowledge graph uses Granger causality tests to update the edge weights of "health factors → abnormal events" in real time, quantifying the contribution of health factors and capturing true causal relationships through time-series lag effects. This resolves the risk of misjudgment due to "correlation ≠ causation," providing a mathematical basis for interpretable intervention recommendations. The group knowledge graph provides a physical organizational framework through clustering trees, a computational engine through a hierarchical federated architecture, and a semantic expression layer. These three elements form a "data-computation-knowledge" closed loop, overcoming the technical dilemma of the trade-off between scale and accuracy in traditional health assessments. This enables the evolution of group health knowledge under privacy protection, providing a dynamic and reliable group benchmark for sleep health assessments. It systematically solves the core problems of the disconnect between group profiles and individual characteristics, and insufficient dynamic adaptability in existing technologies. By correcting the statistical correlation of the group knowledge graph through real-time causal weights of individual knowledge graphs, and providing a benchmark for cold-start users through hierarchical federated learning of the group knowledge graph, and then dynamically adjusting the weights of the updated individual knowledge graph and the weights of the group knowledge graph according to the feedback volume, a smooth transition from group to individual is achieved, which systematically solves the data fragmentation problem and breaks through the technical biases of "incompatible groups and individuals" and "difficulty in deeply utilizing feedback data".

[0065] 3. Existing technologies typically treat feedback simply as a label overlay (e.g., a "drinking" label), while this application creatively transforms feedback into a driving force for graph weight adjustment, and achieves complementary enhancement of group knowledge and individual causality through a hierarchical mechanism. First, at the individual level, the weight increment based on new data is calculated, historical weights are decayed by combining a time decay factor, the human forgetting curve is simulated to prioritize recent data, and the user feedback frequency is converted into a feedback intensity coefficient F(t) to gain weights, directly integrating subjective semantic information, solving the problem of limited evaluation accuracy, and achieving minute-level dynamic optimization, avoiding "individual feature passivation." Next, at the group level, stability is ensured through delayed triggering, dynamic evolution is achieved through extended review, and knowledge consistency is guaranteed through regularized aggregation—a triple synergy that overcomes the response lag bottleneck of traditional health systems (requiring global model retraining), achieving dynamic evolution of the group knowledge graph. Specifically, based on new data, it is determined whether the leaf nodes can be categorized and clustered: if categorized, incremental training of the hierarchical federated architecture adaptive random forest model is triggered, weights are updated locally, and aggregation occurs upwards; if not categorized, extended review creates new nodes and initializes the model, with local incremental updates (no retraining required), combined with regularization... By aggregating and updating the hierarchical weight matrix layer by layer, the dynamic expansion of the group knowledge graph is supported, adapting to individual changes and avoiding feature passivation. This collaborative mechanism enables individual updates to provide real-time input for group reconstruction (such as individual graphs that trigger gain weights based on user feedback serving as group feature sources), and the group knowledge graph to provide initial parameters for cold-start users. Group updates (such as weight matrix and edge weight adjustments) are then integrated with individual updates, forming a closed-loop optimization cycle that breaks down the separation between group and individual knowledge. In particular, the mathematical reinforcement of the time decay factor and feedback intensity coefficient ensures the dynamic integration of subjective experience and objective data, while the delayed increment strategy combined with structural expansion (node ​​addition) achieves real-time weight optimization and semantic integrity, breaking through the limitations of static models. This significantly improves evaluation accuracy, prevents "individual feature passivation," and enhances intervention precision (real-time generalization of group knowledge adapts to individual conditions), thus comprehensively solving the bottlenecks of existing technologies in semantic utilization, dynamic adaptability, and knowledge separation.

[0066] 4. By using the individual weight fusion coefficient μ as a dynamic adjustment lever, the weights of the group knowledge graph serve as a robust foundation, while the weights of the individual knowledge graphs are used for fine-grained correction. This achieves a smooth transition from the group baseline to the personalized model, achieving an optimal balance between common group patterns and personalized accuracy. It becomes the "adaptive nerve center" of sleep health assessment, enabling continuous evolution through "learning while using," completely overcoming the technical problem of rigid static models in existing technologies. This improves the accuracy of long-term sleep health assessment. At the same time, through the cold start resistance design, it can still provide reliable assessments based on the group baseline when user feedback is insufficient.

[0067] Other features and advantages of the present invention will be disclosed in detail in the following detailed description and accompanying drawings. Attached Figure Description

[0068] The invention will be further described below with reference to the accompanying drawings:

[0069] Figure 1 This is a flowchart of a user feedback-driven dynamic assessment method for sleep health according to an embodiment of the present invention.

[0070] Figure 2 This is a schematic diagram of the structure of a user feedback-driven dynamic sleep health assessment system according to an embodiment of the present invention. Detailed Implementation

[0071] The technical solutions of the embodiments of the present invention will be explained and described below with reference to the accompanying drawings. However, the following embodiments are only preferred embodiments of the present invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments in the implementation methods without creative effort are all within the protection scope of the present invention.

[0072] In the following description, terms such as “inner,” “outer,” “upper,” “lower,” “left,” and “right” are used only to indicate orientation or positional relationship for the convenience of describing the embodiments and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention.

[0073] Example 1:

[0074] Please see the appendix Figure 1 , Figure 1 A flowchart illustrating a user feedback-driven dynamic assessment method for sleep health provided in one embodiment of this specification is shown.

[0075] like Figure 1 As shown, this user feedback-driven dynamic assessment method for sleep health may include at least the following steps:

[0076] S1 uses time series causal analysis to construct an individual knowledge graph that reflects the causal relationships of individual health.

[0077] It is understood that the goal of the individual knowledge graph in this embodiment is to construct the association between personal health factors and abnormal events for individual users who have accumulated health factor data over a long period of time and frequently provide interactive feedback, and to represent the strength and personalized characteristics of these associations in the form of a graph structure.

[0078] Specifically, in this embodiment, a time-series causal analysis method is used to construct an individual knowledge graph reflecting the causal relationships of individual health, including:

[0079] S11, obtain the user's historical health factor time series data and abnormal event sequence.

[0080] This embodiment can be achieved by time-series processing of local multi-dimensional data and extracting numerical data types from it to obtain multi-dimensional time series data, which can then be used as the user's health factor time series data. The local multi-dimensional data includes basic user information (age, gender, body mass index (BMI), medical history, etc.), sleep data (heart rate, respiratory rate, body movement, and other physiological signals), and APP behavior logs (function usage frequency, click paths, and behavior analysis data). The types of health factor data include, but are not limited to, sleep duration (total duration), sleep efficiency (deep sleep duration / total duration × 100%), awakenings (number of awakenings at night), respiratory rate (breaths / minute), apnea index (number of apneas per hour), number of blood oxygen saturation drop events, sleep heart rate, resting heart rate (mean), heart rate variability (RMSSD value of HRV), number of abnormal heart rate fluctuations at night, body movement frequency (number of times turning over per unit time), body movement intensity (variance value of accelerometer data), etc., which are not limited in this embodiment. The user's historical health factor time series data... ,in This represents the value of the i-th health factor at time point t, and n represents the total number of health factor types.

[0081] The types of abnormal events include, but are not limited to, insomnia, early awakening, difficulty falling asleep, sleep apnea, snoring, nocturnal arrhythmias, orthostatic tachycardia, etc., and this embodiment does not limit these. Based on user feedback, the abnormal event type corresponding to the feedback content can be automatically identified using existing natural language processing models. Abnormal event sequence ,in Indicates the state of an abnormal event at time point t (e.g.) (Insomnia occurs at time t), m represents the total number of abnormal event types.

[0082] S12 uses time series causal analysis to examine the causal relationship between health factors and abnormal events.

[0083] In this embodiment, the time series causal analysis method is used to examine the causal relationship between health factors and abnormal events. This can be achieved by using information theory methods (such as transfer entropy) or by using machine learning causal discovery (such as PCMCI, LiNGAM).

[0084] Preferably, in this embodiment, a time series causal analysis method is used to examine the causal relationship between health factors and abnormal events, including:

[0085] The Granger causality test algorithm was used to calculate the significance score between health factors and abnormal events;

[0086] The causal relationship between health factors and abnormal events is determined based on the significance scores between health factors and abnormal events.

[0087] The mathematical essence of the Granger causality test is the improvement of predictive power—if we add... Historical data can significantly improve the understanding of The prediction accuracy, then yes The reason is that the basic formula for Granger causality testing (i.e., the time lag model) is:

[0088] ,

[0089] Where K represents the maximum lag order. For example, K=3 means analyzing the impact of the past 3 days. This can be determined based on the actual time period of historical data. This embodiment does not limit this. , This represents the preset regression coefficient. This represents a candidate causal variable with a lag of k periods. This represents the random error term.

[0090] The specific method for calculating the significance score between health factors and abnormal events based on the Granger causality test formula is similar to the method for calculating the p-value using the Granger causality test algorithm in existing technologies, and will not be described again in this embodiment.

[0091] For example, taking the collection of a user's sleep heart rate (HR) over the past 30 days and whether the abnormal event of "frequent awakening" occurred each night (represented by 0 / 1), the Granger causality test analysis is used to determine whether the health factor "sleep heart rate" can significantly predict "frequent awakening". If the significance score between the health factor and the abnormal event is calculated and it is found that the sleep heart rate of the past few nights has a significant impact on the current awakening state (p value < 0.05), then it can be considered that there is a causal relationship between "sleep heart rate" and "frequent awakening".

[0092] It is understood that this embodiment preferably uses the Granger causality test algorithm to test the causal relationship between health factors and abnormal events. The p-value is calculated by the improvement in the predictive ability of the time lag model, which provides a mathematically rigorous causal reasoning basis for the individual knowledge graph of this embodiment. It breaks through the misjudgment trap of "correlation is causation" in traditional health assessment, and can clearly output the causal direction rather than a black box conclusion. It is suitable for short time series and sleep factor data with significant linear relationship, and has low computational complexity, making it suitable for lightweight deployment on wearable devices.

[0093] S13. Based on the causal relationship test results between health factors and abnormal events, the weights between health factors and abnormal events are calculated as edge weights.

[0094] The method for calculating the weight between health factors and abnormal events as edge weights can be as follows: normalize the inverse of the p-value, calculate the causal strength and use it directly as the edge weight to represent the correlation strength between health factors and abnormal events, thereby eliminating the rigidity of statistical thresholds (i.e., the technical problem of traditional binary judgment easily missing weak correlations).

[0095] Furthermore, to avoid ignoring the indirect causality of correlation, this embodiment calculates the edge weights between health factors and abnormal events based on the causal relationship test results, and also includes:

[0096] The causal strength between health factors and anomalous events is calculated based on the significance scores between health factors and anomalous events.

[0097] Calculate the Pearson correlation coefficient between health factors and abnormal events;

[0098] The weighted summation of the causal strength and correlation coefficient between health factors and abnormal events yields the edge weights between them.

[0099] The method for calculating the Pearson correlation coefficient between health factors and abnormal events is similar to the method for calculating correlation in existing technologies, and will not be described again in this embodiment.

[0100] S14. Based on the causal relationship test results between health factors and abnormal events and the edge weights, a directed graph structure is constructed to form an individual knowledge graph.

[0101] The individual knowledge graph includes health factor nodes, abnormal event nodes, and edge weights between health factors and abnormal events.

[0102] Understandably, traditional sleep health assessment systems rely on static association analysis, which fails to capture individual-specific health causal relationships, resulting in a lack of personalized interpretability in the assessment results. To address this, this embodiment designs a dynamic causal modeling individual knowledge graph. It employs the Granger causality test to update the edge weights of "health factors → abnormal events" in real time, quantifying the contribution of health factors and capturing true causal relationships through time-series lag effects. This resolves the risk of misjudgment due to "correlation ≠ causation," providing a mathematical basis for interpretable intervention recommendations. Furthermore, the Granger causality test in this embodiment does not operate independently but rather transforms traditional hypothesis testing into a dynamic and interpretable causal engine through significance score quantification and the fusion of causality and correlation weights. This resolves inter-class misjudgments in existing technologies and deeply collaborates with the dual-graph architecture. New feedback triggers changes in individual health causal relationships in real time, avoiding "individual characteristic blunting." Individual causal changes correct group biases, in turn enhancing the group knowledge graph. The group baseline accelerates cold-start user assessments (e.g., new users directly use the causal weights of the "30-40 year old drinking group"), becoming a core hub for overcoming existing technological bottlenecks.

[0103] S2 combines agglomerative hierarchical clustering and hierarchical federated learning to construct a group knowledge graph that reflects the causal relationships of group health.

[0104] It is understood that the group knowledge graph in this embodiment aims to reveal the correlation mechanism between health factors and abnormal events of a specific user group, and to provide group data support and theoretical basis for health risk modeling, intervention strategy formulation and explanation of potential causal relationships.

[0105] In this embodiment, the combination of agglomerative hierarchical clustering and hierarchical federated learning (using agglomerative clustering to generate a tree structure, with upper-level group nodes acting as coordinators and lower-level subgroup nodes acting as participants) is key to resolving the conflict between privacy and personalization. This structure is naturally adapted to the group differences in health data and is more efficient than traditional federated learning.

[0106] Specifically, in this embodiment, a group knowledge graph reflecting the causal relationships of group health is constructed by combining agglomerative hierarchical clustering and hierarchical federated learning methods, including:

[0107] S21. Obtain historical feedback information, historical health factor data, and abnormal event sequences of the user group, and perform structured transformation to obtain the health feature vector of each user.

[0108] The feedback information may be presented in various forms, including but not limited to text, voice, images, and combinations thereof, and this embodiment does not limit the specific forms.

[0109] The implementation of structured transformation of historical feedback information, historical health factor data, and abnormal event sequences of user groups is as follows: For fields with semantic attributes in feedback information and health factor data (such as basic user information), methods such as rule mapping, keyword extraction, and dictionary encoding are used to transform them into structured labels; for health factor data (such as numerical features in sleep data and APP behavior logs), methods such as standardization and normalization are used to eliminate dimensional differences; for categorical variables such as abnormal event sequences, One-Hot encoding or label encoding is used to uniformly represent them, transforming unstructured feedback into semantic labels, supplementing the missing context of physiological data, and finally constructing a structured feature vector with unified dimensions and clear semantics, which serves as the basic input for subsequent clustering modeling and group knowledge graph construction.

[0110] S22, Based on the health feature vectors of each user, the user group is clustered using an agglomerative hierarchical clustering method to obtain a clustering tree structure that reflects the differences in the health levels of the group;

[0111] Agglomerative hierarchical clustering methods typically merge the most similar samples or groups stepwise until a hierarchical structure is formed. In this embodiment, an agglomerative hierarchical clustering method is used to perform cluster analysis on the health feature vectors of user groups, merging user groups with similar health features from bottom to top, and finally constructing a clustering tree structure that reflects the hierarchical differences in user health features.

[0112] Specifically, in this embodiment, the agglomerative hierarchical clustering method is used to cluster the user group based on the health feature vectors of each user.

[0113] 1. Initialization: Treat each user's health feature vector as a separate cluster (i.e., leaf node).

[0114] 2. Calculate similarity: Use a preset distance metric (such as Euclidean distance) to calculate the similarity between all user pairs. The smaller the distance, the higher the similarity.

[0115] 3. Merge the most similar clusters: Gradually merge the two closest clusters (users or subgroups) to form a new cluster. Merging strategies can employ different connection criteria (such as single connection, full connection, or average connection).

[0116] 4. Update the distance matrix: After each merge, update the distance matrix between clusters.

[0117] 5. Termination condition: Repeat the merging process until all users are merged into a single root node (the entire user group), forming a hierarchical clustering tree.

[0118] In the clustering tree structure of this embodiment, leaf nodes represent user subgroups with similar health characteristics and behavioral patterns (e.g., sleep patterns of diabetic patients aged 40-50), intermediate nodes represent the aggregation relationship of multiple similar subgroups (e.g., a set of people with metabolic abnormalities), and the root node corresponds to the generalized model of the overall user group (global sleep health benchmark).

[0119] It is understandable that the clustering tree structure in this embodiment not only accurately reflects the population stratification pattern related to health risks, but also provides a clear hierarchical division and organizational basis for subsequent hierarchical federated learning and population knowledge graph construction, effectively improving the generalization ability of the sleep health assessment system and the accuracy of population difference modeling.

[0120] S23 treats each leaf node of the clustered tree structure as an independent participant, and the intermediate nodes and root node of the clustered tree structure as coordinators, forming a hierarchical federated architecture.

[0121] Understandably, the clustering tree structure determines the architecture of hierarchical federated learning. In hierarchical federated learning, the upper-level nodes (coordinators) correspond to the middle or root nodes in the clustering tree and are responsible for aggregating the parameters of the lower-level nodes. The lower-level nodes (participants) correspond to the leaf nodes (i.e., specific user subgroups) in the clustering tree, with each leaf node serving as a participant.

[0122] S24. Based on the hierarchical federated architecture, hierarchical federated learning is performed on the health feature vectors of each user, and a group knowledge graph reflecting the causal relationship of group health is constructed based on the federated learning results.

[0123] It is understood that the group knowledge graph in this embodiment is constructed based on the hierarchical clustering tree structure generated by user group classification. Combined with the hierarchical federated collaboration mechanism in horizontal federated learning, it simulates the synchronization and fusion process of feature parameters between multiple participants and coordinators, and realizes the layer-by-layer transfer and generalization of feature weights across groups.

[0124] Specifically, in this embodiment, S24, hierarchical federated learning is performed on the health feature vectors of each user based on a hierarchical federated architecture, and a group knowledge graph reflecting the causal relationship of group health is constructed based on the federated learning results, including:

[0125] S241 uses participants as group labels for the group knowledge graph. .

[0126] It is understood that in this embodiment, the group label nodes in the group knowledge graph are directly derived from the leaf nodes in the clustering tree (each leaf node is a group label). The group label nodes in the group knowledge graph and the leaf nodes in the clustering tree structure are essentially two representations of the same entity. The leaf nodes generated by clustering are used for data partitioning and federated architecture organization, while the group label nodes are used for knowledge storage and interpretable querying of the group knowledge graph.

[0127] S242, the coordinator divides the health feature vectors of all users into each participating party based on a hierarchical federated architecture;

[0128] It is understandable that in this embodiment, the upper-level node (coordinator) initializes the global feature weight vector. (Initial parameters can be determined by the statistical average features of all users or the cluster center features), and the user's health feature vector is divided into subgroups (corresponding to the leaf nodes of the cluster tree structure) according to the cluster tree structure. Each lower-level node (participant) only obtains the user subset within its own group and its corresponding health feature vector data, and independently learns the sparse feature parameters. While ensuring privacy, personalized modeling is achieved, thereby enabling differentiated modeling and knowledge representation of group characteristics.

[0129] S243, each participant builds an adaptive random forest model locally and incrementally trains the model based on the health feature vectors of the users in the partition.

[0130] In this embodiment, each participant constructs its own locally-based adaptive random forest model. The input is the health feature vector of users in its respective subgroup, and the output is the predicted probability distribution of sleep health status. The model adapts to dynamic changes in user data (such as new feedback) through incremental training. The implementation method of incremental training of the local adaptive random forest model by each participant is similar to the training method of adaptive random forest models in existing technologies, and will not be elaborated further in this embodiment.

[0131] S244 uses the SHAP method to score the importance of health features on the trained adaptive random forest model, obtains the health feature weight vector of each participant, and transforms it into the edge weight between health factors and group labels.

[0132] In this embodiment, the SHAP method is used to score the health feature importance of the trained adaptive random forest model, which can be expressed as:

[0133] ,

[0134] in, The average contribution of health factor i (after normalization) is represented by F, and F represents the number of trees in the adaptive random forest. This represents the marginal contribution of health factor i in the e-th tree. The resulting health feature weight vector of the participating parties is generated accordingly. This represents the importance of each health factor within the subgroup. Next, the health characteristic weight vectors of the participants are... Transformed into the marginal weights between health factors and group labels To achieve machine-readable knowledge storage.

[0135] S245, the health feature weight vectors of each participant are aggregated upwards based on the hierarchical federation architecture to obtain the feature weight vectors of each coordinator, forming a hierarchical weight matrix.

[0136] For example, in this embodiment, the method of aggregating the health feature weight vectors of each participant to obtain the feature weight vector of each coordinator is as follows: the health feature weight vectors of each participant are passed from bottom to top, each coordinator performs weighted fusion of the health feature weight vectors of its child nodes, and then regularization optimization is used to avoid small sample distortion, finally obtaining the feature weight vector of the coordinator. The root node coordinator integrates the importance of all group features to form a hierarchical weight matrix. , where L represents the number of layers in the clustered tree structure.

[0137] The design of the hierarchical weight matrix in this embodiment is based on the following considerations:

[0138] 1. Hierarchical Nature of Knowledge Graphs: Group knowledge graphs are not single-level but hierarchical. Each intermediate node and root node in the clustering tree structure corresponds to a different level of group aggregation. Therefore, the feature weight vector of the root node coordinator represents the feature importance from a global perspective. Realizing the hierarchical and multi-scale modeling of the group knowledge graph is crucial for the generalization ability of the entire system.

[0139] 2. Cold start support: For new users or users with little feedback data, the system needs to rely on higher-level group knowledge. The global weight of the root node coordinator provides a benchmark evaluation basis for this situation, becoming the computational basis for supporting the dynamic updating of the group knowledge graph.

[0140] 3. Model stability: Hierarchical aggregation and regularization optimization can prevent weight distortion in small sample subgroups and improve system robustness.

[0141] 4. Dynamic update mechanism: When user feedback triggers a group graph update, this hierarchical aggregation process allows knowledge to spread layer by layer from individuals to subgroups to the main group to the global level, maintaining knowledge consistency.

[0142] Understandably, although the group knowledge graph directly stores the edge weights at the leaf node level, its core function is to achieve interpretable analysis of sleep health risks through structured associations, while the aggregated weight vectors of the upper-level nodes... It is the computational foundation that supports dynamic updates of the knowledge graph. When a user is classified into a subgroup (participant), if the data for that subgroup is insufficient, the feature weight vector of the upper-level node (coordinator) can be used as a supplementary basis. The globally integrated hierarchical weight matrix actually constitutes a multi-scale indexing framework for the group knowledge graph. The importance of health features at different levels corresponds to different granularities of group knowledge, supporting semantic connectivity from the micro (subgroup) to the macro (global) level. It resolves the contradiction between "group difference modeling" and "global knowledge accumulation", upgrading the group knowledge graph from static storage to a dynamic decision engine.

[0143] S246, the coordinator counts the frequency of anomalous events among each participant and converts it into the edge weights between group labels and anomalous events.

[0144] Understandably, Granger causality tests on individual knowledge graphs provide evidence for the association of anomalous events in group knowledge graphs. In this embodiment, the frequency of an anomalous event is the quotient of the number of times a certain anomalous event occurs and the total number of anomalous events occurring in the participants, representing the distribution of anomalous events in that subgroup. Then, the frequency of anomalous events occurring in the participants is converted into the edge weights between group labels and anomalous events. This enables machine-readable knowledge storage, and the edge weights between group labels and abnormal events can reveal the types of high-incidence abnormal events in a subgroup.

[0145] S247. Based on the group label, the edge weights between health factors and the group label, and the edge weights between the group label and abnormal events, a directed graph structure is constructed to form a group knowledge graph.

[0146] The group knowledge graph includes health factor nodes, group tag nodes, abnormal event nodes, and edge weights between health factors and group tags. and the boundary weights between group labels and anomalous events .

[0147] It is understood that the entire process of constructing the group knowledge graph in this embodiment is as follows: agglomerative clustering to construct a tree structure → the tree structure determines the hierarchical architecture of federated learning (leaf nodes are participants, and their parent nodes and above are coordinators) → the results of federated learning (feature weights of subgroups and statistics of abnormal events) are used to construct the group knowledge graph (where the group label node is the leaf node of the clustering tree).

[0148] The technical concept of this embodiment is as follows: First, a cohesive hierarchical clustering method is used to merge users' health feature vectors from bottom to top into a clustering tree structure, where leaf nodes represent fine-grained user subgroups, intermediate nodes aggregate similar groups, and the root node carries the global health benchmark—this hierarchical division naturally adapts to the group differences in health data; Second, the clustering tree structure is mapped to a hierarchical federated architecture: leaf nodes act as participants independently training adaptive random forest models, while intermediate and root nodes act as coordinators hierarchically aggregating the parameters of corresponding child nodes—this design breaks through the flat architecture of traditional federated learning, achieving a balance between privacy protection (data not leaving the local machine) and group knowledge accumulation (i.e., extracting common patterns across groups); Next, participants extract health feature weight vectors using the SHAP method and transform them into edge weights of the group knowledge graph (health factor → group label), while simultaneously generating another type of edge weight (group label → abnormal event) by statistically analyzing the frequency of abnormal events; Finally, the coordinator constructs a hierarchical weight matrix by aggregating the health feature weight vectors of the participants, forming a group knowledge graph containing a multi-scale indexing framework.

[0149] It is understandable that this embodiment provides a physical organizational framework through clustering trees, a computational engine through hierarchical federated architecture, and a semantic expression layer through a group knowledge graph. These three elements form a closed loop of "data-computation-knowledge," breaking through the technical dilemma of the incompatibility between scale and accuracy in traditional health assessments. It achieves the evolution of group health knowledge under privacy protection, providing a dynamic and reliable group benchmark for sleep health assessments. This systematically solves the core problems of the disconnect between group profiles and individual characteristics, and insufficient dynamic adaptability in existing technologies. Specifically, this is reflected in the following three aspects:

[0150] 1. Dynamic Adaptive Multiplication: When new user feedback triggers incremental training of the local model of the participants, the SHAP health feature weights are updated in real time, and the coordinator weights are synchronously corrected through hierarchical upward aggregation. Compared with the periodic federated learning of patent CN119851951A (which requires global retraining) and the static SVM of patent CN118098545A, the response speed can be greatly improved.

[0151] 2. Group-Individual Knowledge Integration: Fine-grained weights of leaf nodes provide accurate benchmarks for individuals; as individual feedback accumulates, the system dynamically integrates and updates the weights of the individual knowledge graph with the weights of the group knowledge graph to achieve a complementary relationship between "group benchmark robustness and individual causal specificity"—solving the mechanical matching problem of the static TCM constitution table in patent CN119851951A and the disconnect between equipment strategy and health model in patent CN118098545A.

[0152] 3. Enhanced interpretability: The marginal weights between health factors and group labels Quantifying the contribution of health factors to a specific group, and the marginal weights between group labels and anomalous events. Revealing high-incidence anomalies in a group allows subsequent assessment conclusions to be traced back to specific characteristics (such as "in the user's group, high blood sugar contributes 75% to insomnia")—overcoming the semantic poverty of the reinforcement learning black box of patent CN119851951A and the binary classification alarm of patent CN119851951A.

[0153] In summary, one of the innovations of this embodiment lies in the design of a user feedback-driven dual-graph collaborative architecture. This architecture transforms user feedback into structured health feature vectors, constructing individual and group knowledge graphs. The individual knowledge graph uses Granger causality tests to analyze the association between health factors and abnormal events. The group knowledge graph dynamically aggregates group feature weights through agglomerative hierarchical clustering and hierarchical federated learning. Real-time causal weights from the individual knowledge graph correct the statistical correlation of the group knowledge graph. Hierarchical federated learning in the group knowledge graph provides a benchmark for cold-start users. Furthermore, by dynamically adjusting the weights of the updated individual and group knowledge graphs based on subsequent feedback volume, a smooth transition from group to individual knowledge is achieved. This systematically solves the data fragmentation problem and overcomes the technical biases of "incompatible groups and individuals" and "difficulty in deeply utilizing feedback data."

[0154] S3 can obtain user feedback information and health factor data in real time.

[0155] S4. Based on newly acquired user feedback information and health factor data, trigger the individual knowledge graph update to reflect changes in the causal relationship of the user's individual health in real time, and obtain the weight of the updated individual knowledge graph.

[0156] Specifically, in this embodiment, based on newly acquired user feedback information and health factor data, an individual knowledge graph update is triggered to reflect changes in the causal relationships of a user's individual health in real time, resulting in the weights of the updated individual knowledge graph, including:

[0157] S41, Based on the newly acquired health factor data, the weights between health factors and abnormal events are calculated using the method described above, and denoted as the weight increment based on the new data. .

[0158] S42 uses a time decay factor to decay the edge weights between historical health factors and abnormal events, and then adds weight increments based on new data to obtain the first weight. , can be represented as:

[0159] ,

[0160] in, Indicates the health factors at the previous moment and abnormal events The weights between them This represents the time decay factor, i.e., the preset decay rate (e.g., 0.05 / day, which means the weight decays exponentially over time, simulating the human memory curve). This indicates the time interval between the current moment and the last update (e.g., in days).

[0161] S43, based on the newly acquired user feedback information, convert the frequency of user feedback information into a feedback intensity coefficient. The second weight is obtained by gaining the first weight based on the feedback intensity coefficient. As the weights of the updated individual knowledge graph, they can be represented as:

[0162] ,

[0163] In this embodiment, the feedback intensity coefficient The weighting of relevant abnormal events is determined by the severity and frequency of the feedback content. The frequency and intensity coefficient of user feedback information are related. The conversion rules between these values ​​can be set according to the actual application, and this embodiment does not limit them. For example, taking F(t)∈[0,0.2] as an example, the conversion rules can be: F(t)=0.05 for a single ordinary feedback (such as occasional "insomnia"), F(t)=0.15 for a medically confirmed event (such as "diagnosed sleep apnea"), and F(t)=0.2 for a high-frequency severe feedback (such as "waking up ≥3 times per night").

[0164] Furthermore, this embodiment can also dynamically optimize the individual knowledge graph structure, including:

[0165] 1. Edge pruning: Remove Weakly correlated edges are used to eliminate incidental noise (such as temporary insomnia during travel) and ensure that individual knowledge graphs retain only significant correlations, thereby improving the accuracy of assessment.

[0166] 2. Adding new nodes: When a new abnormal event exceeds the existing category, a new abnormal event node is created and its weight is initialized (e.g., the first occurrence of "long-term insomnia after a cold"), thereby achieving a minute-level response speed (traditional solutions require at least days of retraining).

[0167] Understandably, existing technologies typically treat feedback simply as a label overlay (such as a "drinking" label). However, this embodiment creatively transforms user feedback information (i.e., subjective user input) into a mathematical reinforcement engine for causal knowledge, solving the long-standing technical bottleneck of "insufficient utilization of semantic information" in the field of sleep health. At the same time, it simulates the forgetting curve of human memory through a time decay factor (high weight for recent data and low weight for long-term data), and strengthens the influence of the user's subjective experience by combining the feedback intensity coefficient F(t). This ensures that the individual knowledge graph always reflects the user's latest and most relevant sleep health status. This design breaks through the rigidity of traditional static models and effectively solves the problem of assessment distortion caused by dynamic changes in individual health status, providing real-time basis for precise intervention.

[0168] S5, based on newly acquired user feedback information and health factor data, triggers delayed incremental reconstruction of the group knowledge graph to obtain the updated weights of the group knowledge graph.

[0169] Understandably, in order to adapt to changes in the causal relationship of individual user health, this embodiment introduces a dynamic update strategy in the group knowledge graph. By combining delayed incremental training with structural expansion, the continuous generalization ability and semantic integrity of the group knowledge graph are ensured.

[0170] Specifically, in this embodiment, based on newly acquired user feedback information and health factor data, a delayed incremental reconstruction of the group knowledge graph is triggered to obtain the updated weights of the group knowledge graph, including:

[0171] The newly acquired user feedback information and health factor data are structurally transformed to obtain newly acquired health feature vectors;

[0172] Determine whether the newly acquired health feature vector can be classified into any leaf node of the clustering tree structure.

[0173] If a health feature vector can be categorized into a leaf node of a clustering tree structure, the coordinator assigns the newly acquired health feature vector to the corresponding participant based on a hierarchical federated architecture. When the number of newly acquired health feature vectors received by a participant exceeds a preset threshold, the coordinator incrementally trains the participant's local adaptive random forest model using the method described above, updates the participant's health feature weight vector, aggregates it upwards, and updates the hierarchical weight matrix layer by layer to obtain the updated weights of the group knowledge graph.

[0174] If it cannot be classified into any leaf node of the clustering tree structure, then perform the following extended review steps:

[0175] Calculate the distance between the newly acquired health feature vector and the cluster center of the group. When the distance is less than or equal to a preset distance threshold, ignore the non-standard data. When the distance is greater than the preset distance threshold, determine the nearest neighbor intermediate node based on the distance and create the corresponding new leaf node as a new participant.

[0176] The coordinator assigns the newly acquired health feature vectors to the new participants based on a hierarchical federated architecture;

[0177] The aforementioned method is used to construct a local adaptive random forest model for the new participants. The model of the new participants is incrementally trained to complete the initialization. After calculating the health feature weight vector of the new participants, it is aggregated upwards and the hierarchical weight matrix is ​​updated layer by layer to obtain the weight of the updated group knowledge graph.

[0178] In this embodiment, the method for determining whether a newly acquired health feature vector can be classified into any leaf node of the clustering tree structure can be achieved by calculating the distance between feature vectors as the basis for judgment. This embodiment will not elaborate on this.

[0179] The implementation method of this embodiment is as follows: When a user generates new feedback information and health factor data, it is first determined whether it matches the existing group classification system. If it can be classified into a certain leaf node (such as "alcoholic insomnia" belonging to the "metabolic abnormal sleep disorder" subgroup), then the adaptive random forest model incremental training of that subgroup node is triggered to update the local health feature weight vector. The data is then uploaded to the upper-level node for weighted aggregation to obtain the health feature weight vector of the coordinator. If the group exceeds the existing group classification system (e.g., the first appearance of "long-term insomnia after a cold"), an extended review mechanism is initiated. This mechanism verifies the group's innovativeness by calculating the distance between feature vectors. Once confirmed, the group is inserted into a suitable position in the clustering tree. A new subgroup node is created, and an adaptive random forest model is initialized. At the same time, a time decay factor is introduced to adjust the influence of historical weights. Finally, regularized aggregation is used to update the group layer by layer from bottom to top until the root node coordinator completes the global knowledge graph reconstruction. During this process, the edge weights of "health factor → group label" (derived from the transformation) and "group label → abnormal event" (re-statistical abnormal event frequency) in the group knowledge graph are updated simultaneously.

[0180] Understandably, this embodiment creatively transforms feedback into the driving force for graph weight adjustment and achieves complementary enhancement of group knowledge and individual causality through a hierarchical mechanism. Firstly, at the individual level, it calculates the weight increment based on new data, combines a time decay factor to decay historical weights, simulates the human forgetting curve to prioritize recent data, converts user feedback frequency into a feedback intensity coefficient F(t) to gain weights, directly integrates subjective semantic information, solves the problem of limited evaluation accuracy, and achieves minute-level dynamic optimization, avoiding "individual feature passivation." Next, at the group level, it achieves stability through delayed triggering, dynamic evolution through extended review, and knowledge consistency through regularized aggregation—a triple synergy that overcomes the response lag bottleneck of traditional health systems (requiring global model retraining) and realizes the dynamic evolution of the group knowledge graph. Specifically, based on new data, it determines whether the leaf nodes can be categorized and clustered: if categorized, it triggers incremental training of the hierarchical federated architecture adaptive random forest model, updates weights locally, and aggregates upwards; if not categorized, it extends review to create new nodes and initializes the model, performing local incremental updates (without retraining), combined with regularization... By aggregating and updating the hierarchical weight matrix layer by layer, the dynamic expansion of the group knowledge graph is supported, adapting to individual changes and avoiding feature passivation. This collaborative mechanism enables individual updates to provide real-time input for group reconstruction (such as individual graphs that trigger gain weights based on user feedback serving as group feature sources), and the group knowledge graph to provide initial parameters for cold-start users. Group updates (such as weight matrix and edge weight adjustments) are integrated with individual updates, forming a closed-loop optimization that breaks down the separation between group and individual knowledge. In particular, the mathematical reinforcement of the time decay factor and feedback intensity coefficient ensures the dynamic integration of subjective experience and objective data. The delayed increment strategy combined with structural expansion (node ​​addition) achieves real-time weight optimization and semantic integrity, breaking through the limitations of static models. This significantly improves evaluation accuracy, prevents "individual feature passivation," and enhances intervention precision (real-time generalization of group knowledge adapts to individual conditions). Overall, it solves the bottlenecks of existing technologies in semantic utilization, dynamic adaptability, and knowledge fragmentation.

[0181] S6. Based on the cumulative amount of user feedback information, the weights of the updated individual knowledge graph and the group knowledge graph are dynamically integrated to obtain the comprehensive weight of the user health factor.

[0182] Specifically, in this embodiment, the weights of the updated individual knowledge graph and the group knowledge graph are dynamically fused based on the cumulative amount of user feedback information to obtain the comprehensive weight of the user health factor, including:

[0183] It can determine in real time whether the cumulative number of user feedback messages exceeds a preset feedback threshold.

[0184] If the cumulative number of user feedback messages is less than or equal to the preset feedback threshold, then the edge weight between health factors and group tags in the updated group knowledge graph will be used as the comprehensive weight of the user's health factors.

[0185] If the cumulative number of user feedback messages exceeds a preset feedback threshold, the cumulative number of user feedback messages is converted into an individual weight fusion coefficient. The individual weight fusion coefficient is proportional to the cumulative number of user feedback messages. Based on the individual weight fusion coefficient, the weights of the updated individual knowledge graph and the weights of the group knowledge graph are weighted and summed to obtain the comprehensive weight of the user health factor, which can be expressed as:

[0186] ,

[0187] Where i represents the index of the health factor, This represents the overall weight of the i-th health factor. The individual weight fusion coefficient can be dynamically adjusted according to the richness of user feedback information (for example, the more feedback is accumulated, the larger μ becomes) to achieve a smooth transition. This embodiment does not limit this. That is, the edge weight between the i-th health factor and the group label in the updated group knowledge graph. , That is, the weight between the i-th health factor and the abnormal event in the updated individual knowledge graph. .

[0188] Understandably, in this embodiment, when the cumulative number of user feedback messages is less than or equal to a preset feedback threshold, i.e., for users with no feedback (initial feedback or few feedback records), the path to obtain the comprehensive weight of the health factor is as follows: matching layer by layer along the clustering tree structure of the group knowledge graph to determine the group label, and then obtaining the edge weight between the group label and the health factor as the comprehensive weight of the user's health factor. When the cumulative number of user feedback messages is greater than the preset feedback threshold, i.e., for users with feedback (users have accumulated some individual feedback records), the path to obtain the comprehensive weight of the health factor is as follows: dual weighting of the individual knowledge graph and the group knowledge graph, dynamically adjusting the ratio of the weight of the group knowledge graph to the weight of the individual knowledge graph through the individual weight fusion coefficient μ to obtain the comprehensive weight of the user's health factor.

[0189] Understandably, this embodiment uses the individual weight fusion coefficient μ as a dynamic adjustment lever, taking the weights of the group knowledge graph as a robust foundation and the weights of the individual knowledge graph as fine-grained corrections. This achieves a smooth transition from the group baseline to the personalized model, achieving an optimal balance between common group patterns and personalized accuracy. It becomes the "adaptive neural center" of sleep health assessment, enabling continuous evolution through "learning while using," completely overcoming the rigidity of static models in existing technologies. This improves the accuracy of long-term sleep health assessment. Furthermore, through a cold-start resistance design, it can still provide reliable assessments based on the group baseline even when user feedback is insufficient. Moreover, the fusion formula can clearly separate the contributions of group patterns and individual characteristics, improving the interpretability of sleep health assessment (e.g., "Your heart rate score is 60% based on personal characteristics (40 feedbacks) and 40% referenced to similar populations").

[0190] S7 calculates the user's sleep health assessment score by weighting the newly acquired health factor data based on the comprehensive weight of the user's health factors.

[0191] Furthermore, this embodiment can also quantify the causal strength based on the edge weights of the individual knowledge graph to generate interpretable intervention suggestions, that is, retrieve the high-weight individual health causal relationships in the individual knowledge graph, generate a graph-visualized causal chain, and solve the technical problem of lack of evidence for intervention in the prior art.

[0192] In summary, this embodiment systematically solves the problems of existing technologies through a user feedback-driven dual-graph collaborative architecture: First, it constructs a bidirectional optimized individual knowledge graph and a group knowledge graph—the health causal relationship of the individual knowledge graph provides fine-grained correlation basis for the group knowledge graph, and the group knowledge graph provides benchmark weights for cold-start users; Second, it acquires user feedback information in real time, directly integrating subjective semantic information into the evaluation process, solving the problem of limited evaluation accuracy caused by the lack of subjective feedback in existing technologies (such as misjudging inter-class similarity or being unable to distinguish the causes of similar signals), and directly driving subsequent graph updates and weighted calculations through user-initiated input data; Next, it updates the individual knowledge graph weights in real time based on new data, capturing changes in user health causal relationships with millisecond-level response, and combines this with delayed incremental reconstruction of the group knowledge graph (avoiding frequent computational burden while maintaining dynamic adaptability), solving the problems of individual feature passivation and insufficient personalization caused by static models in existing technologies; Finally, it... By dynamically integrating individual and group weights based on the cumulative feedback, the group model is transformed from a "data averager" into an "individualized foundation" (using a group knowledge graph constructed through hierarchical federated learning as a foundation, and adapting to individual differences through reconstruction). At the same time, the individual model is upgraded from an "isolated analyzer" to a "driving force for group evolution" (real-time feedback from individual updates triggers group reconstruction, driving the evolution of the group model). Through this bidirectional data flow (individual update → group reconstruction → weight fusion) and mathematical fusion mechanism (adjusting the weight ratio according to the cumulative feedback), the mutual reinforcement of individual characteristics and group knowledge is achieved, thereby overcoming the problem of separation between the group and the individual in existing technologies (such as the inability of a single model to adapt to physical differences). This achieves a technical effect of "1+1>2"—both improving efficiency through large-scale group data and achieving precise personalization through individual feedback, significantly improving the real-time performance, accuracy, and interpretability of sleep health assessment, and realizing the unity of scale and personalization in the field of sleep health assessment.

[0193] Example 2:

[0194] Please see the appendix Figure 2 , Figure 2 This is a schematic diagram of a user feedback-driven dynamic sleep health assessment system provided in one embodiment of this specification.

[0195] like Figure 2 As shown, this user feedback-driven dynamic sleep health assessment system may include at least:

[0196] Individual knowledge graph construction module 1 is used to construct an individual knowledge graph that reflects the causal relationships of individual health using time series causal analysis methods;

[0197] Group Graph Construction Module 2 is used to combine agglomerative hierarchical clustering methods and hierarchical federated learning methods to construct a group knowledge graph that reflects the causal relationship of group health.

[0198] Data acquisition module 3 is used to acquire user feedback information and health factor data in real time;

[0199] The dual-graph weight update module 4 is used to trigger the update of the individual knowledge graph based on newly acquired user feedback information and health factor data to reflect the changes in the causal relationship of the user's individual health in real time, obtain the weight of the updated individual knowledge graph, and trigger the delayed incremental reconstruction of the group knowledge graph to obtain the weight of the updated group knowledge graph.

[0200] The weight fusion module 5 is used to dynamically fuse the weights of the updated individual knowledge graph and the group knowledge graph based on the cumulative amount of user feedback information to obtain the comprehensive weight of the user health factor.

[0201] The sleep health assessment module 6 is used to perform weighted calculations on newly acquired health factor data based on the comprehensive weight of user health factors, and obtain the user's sleep health assessment score.

[0202] It is understood that the technical concept of the sleep health dynamic assessment system based on user feedback provided in this embodiment is similar to the technical concept of the sleep health dynamic assessment method based on user feedback provided in the foregoing embodiments, and will not be repeated here.

[0203] Example 3:

[0204] Another embodiment of this specification provides a computer-readable storage medium storing instructions that, when executed on a computer or processor, cause the computer or processor to perform one or more steps in the embodiments of the user feedback-driven dynamic sleep health assessment method described above. If the constituent modules of the above-described electronic device are implemented as software functional units and used as independent downstream task predictions or applications, they can be stored in a computer-readable storage medium.

[0205] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this specification is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., Digital Versatile Discs (DVDs)), or semiconductor media (e.g., Solid State Disks (SSDs)).

[0206] The above description is merely a preferred embodiment disclosed in this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of protection involved in this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this disclosure.

[0207] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

Claims

1. A sleep health dynamic assessment method based on user feedback driven, characterized in that, The method comprises the following steps: adopting time series causal analysis method, constructing individual knowledge graph reflecting individual health causal relationship; combining the condensed hierarchical clustering method and the hierarchical federated learning method, constructing a group knowledge graph reflecting the causal relationship of group health; real-time acquisition of user feedback information and health factor data; based on the newly acquired user feedback information and health factor data, triggering the individual knowledge graph update to reflect the individual health causal relationship changes of the user in real time, obtaining the weight of the updated individual knowledge graph; based on the newly acquired user feedback information and health factor data, triggering the group knowledge graph delayed incremental reconstruction, obtaining the weight of the updated group knowledge graph; based on the cumulative number of user feedback information, dynamically fusing the weight of the updated individual knowledge graph and the weight of the group knowledge graph to obtain the comprehensive weight of the user health factor; based on the comprehensive weight of the user health factor, performing weighted calculation on the newly acquired health factor data to obtain the sleep health evaluation score of the user; wherein, combining the condensed hierarchical clustering method and the hierarchical federated learning method, constructing a group knowledge graph reflecting the causal relationship of group health, comprising: obtaining the historical feedback information, historical health factor data and abnormal event sequence of the user group, and performing structured conversion to obtain the health feature vector of each user; based on the health feature vector of each user, adopting the condensed hierarchical clustering method to cluster the user group, obtaining a clustering tree structure reflecting the hierarchical difference of the group health; each leaf node of the clustering tree structure is regarded as an independent participant, and the intermediate nodes and root nodes of the clustering tree structure are regarded as coordinators, forming a hierarchical federated architecture; based on the hierarchical federated architecture, performing hierarchical federated learning on the health feature vector of each user, and based on the federated learning result, constructing a group knowledge graph reflecting the causal relationship of group health; wherein, based on the newly acquired user feedback information and health factor data, triggering the group knowledge graph delayed incremental reconstruction, obtaining the weight of the updated group knowledge graph, comprising: performing structured conversion on the newly acquired user feedback information and health factor data to obtain the newly acquired health feature vector; determining whether the newly acquired health feature vector can be classified into any leaf node of the clustering tree structure, if it can be classified into a leaf node of the clustering tree structure, the coordinator divides the newly acquired health feature vector to the corresponding participant based on the hierarchical federated architecture, and when the number of newly added health feature vectors received by the participant exceeds the preset threshold, the adaptive random forest model of the participant is incrementally trained, the health feature weight vector of the participant is updated, the hierarchical weight matrix is updated layer by layer, and the weight of the updated group knowledge graph is obtained, if it cannot be classified into any leaf node of the clustering tree structure, the following extension audit steps are performed: calculate the distance between the newly acquired health feature vector and the clustering center of the group, when the distance is greater than the preset distance threshold, determine the nearest neighbor intermediate node based on the distance and create a new leaf node as a new participant; the coordinator divides the newly acquired health feature vector to the new participant based on the hierarchical federated architecture; The local adaptive random forest model is constructed for the new participant, the model of the new participant is incrementally trained to complete initialization, the health feature weight vector of the new participant is calculated and aggregated upwards, the hierarchical weight matrix is updated layer by layer, and the weight of the updated group knowledge graph is obtained.

2. The sleep health dynamic evaluation method driven by user feedback according to claim 1, wherein, An individual knowledge graph reflecting the causal relationship of individual health is constructed by using a time series causal analysis method, including: Obtaining the time series data of historical health factors and the sequence of abnormal events of the user; Using a time series causal analysis method, the causal relationship between health factors and abnormal events is tested; Based on the test results of the causal relationship between health factors and abnormal events, the weight between health factors and abnormal events is calculated as an edge weight; Based on the test results of the causal relationship between health factors and abnormal events and the edge weight, a directed graph structure is constructed to form an individual knowledge graph.

3. The sleep health dynamic evaluation method driven by user feedback according to claim 2, wherein, The causal relationship between health factors and abnormal events is tested by using a time series causal analysis method, including: Using Granger causal test algorithm, the significance score between health factors and abnormal events is calculated; Based on the significance score between health factors and abnormal events, the causal relationship between health factors and abnormal events is determined.

4. The sleep health dynamic evaluation method driven by user feedback according to claim 1, wherein, Based on the hierarchical federated architecture, the health feature vectors of each user are subjected to hierarchical federated learning, and based on the federated learning results, a group knowledge graph reflecting the causal relationship of the group health is constructed, including: The participants are taken as group labels of the group knowledge graph; The coordinator divides the health feature vectors of all users into each participant based on the hierarchical federated architecture; Each participant locally constructs an adaptive random forest model, and incrementally trains the model based on the health feature vectors of the users divided therein; Using the SHAP method, the health feature importance score of the trained adaptive random forest model is obtained, and the health feature weight vector of each participant is obtained, and is converted into the edge weight between health factors and group labels; The health feature weight vectors of each participant are aggregated upwards based on the hierarchical federated architecture, and the feature weight vectors of each coordinator are obtained to form a hierarchical weight matrix; The coordinator statistics the frequency of abnormal events in each participant, and converts it into the edge weight between the group label and the abnormal event; Based on the group label, the edge weight between the health factor and the group label, and the edge weight between the group label and the abnormal event, a directed graph structure is constructed to form a group knowledge graph.

5. The sleep health dynamic evaluation method driven by user feedback according to claim 2, wherein, Based on the newly obtained user feedback information and health factor data, the individual knowledge graph is updated to reflect the changes of the individual health causal relationship of the user in real time, and the weight of the updated individual knowledge graph is obtained, including: Based on the newly acquired health factor data, the weight between the health factor and the abnormal event is calculated, denoted as the weight increment based on new data; The edge weight between the historical health factor and the abnormal event is decayed by using a time decay factor, and the weight increment based on new data is superimposed to obtain a first weight; Based on the newly acquired user feedback information, the frequency of the user feedback information is converted into a feedback intensity coefficient, and the first weight is gain based on the feedback intensity coefficient to obtain a second weight as the weight of the updated individual knowledge graph.

6. The sleep health dynamic evaluation method based on user feedback driving according to claim 1, wherein, The weight of the updated individual knowledge graph and the weight of the group knowledge graph are dynamically fused based on the cumulative number of user feedback information to obtain a comprehensive weight of the user health factor, comprising: determining in real time whether the cumulative number of user feedback information is greater than a preset feedback threshold, if the cumulative number of user feedback information is less than or equal to the preset feedback threshold, the edge weight between the health factor and the group label in the updated group knowledge graph is directly taken as the comprehensive weight of the user health factor, if the cumulative number of user feedback information is greater than the preset feedback threshold, the cumulative number of user feedback information is converted into an individual weight fusion coefficient, the individual weight fusion coefficient is proportional to the cumulative number of user feedback information, and the weight of the updated individual knowledge graph and the weight of the group knowledge graph are weighted and summed based on the individual weight fusion coefficient to obtain the comprehensive weight of the user health factor.

7. A sleep health dynamic assessment system driven by user feedback, characterized in that, including: an individual graph construction module for constructing an individual knowledge graph reflecting individual health causal relationship by using time series causal analysis method; a group graph construction module for constructing a group knowledge graph reflecting group health causal relationship by combining agglomerative hierarchical clustering method and hierarchical federated learning method; a data acquisition module for acquiring user feedback information and health factor data in real time; a double graph weight updating module for triggering individual knowledge graph updating to reflect real-time changes of individual health causal relationship of the user based on newly acquired user feedback information and health factor data, obtaining the weight of the updated individual knowledge graph, and triggering group knowledge graph delayed incremental reconstruction, obtaining the weight of the updated group knowledge graph; a weight fusion module for dynamically fusing the weight of the updated individual knowledge graph and the weight of the group knowledge graph based on the cumulative number of user feedback information to obtain the comprehensive weight of the user health factor; a sleep health evaluation module for performing weighted calculation on the newly acquired health factor data based on the comprehensive weight of the user health factor to obtain the sleep health evaluation score of the user; wherein the group graph construction module is used to perform the following steps: obtaining historical feedback information, historical health factor data and abnormal event sequence of the user group, and performing structured conversion to obtain health feature vectors of each user; based on the health feature vectors of each user, using agglomerative hierarchical clustering method to cluster the user group to obtain a clustering tree structure reflecting the health level difference of the group; Each leaf node of the clustering tree structure is taken as an independent participant, and the intermediate nodes and the root node of the clustering tree structure are taken as coordinators, to form a hierarchical federation architecture; Based on the hierarchical federation architecture, hierarchical federated learning is performed on the health feature vectors of each user, and a group knowledge graph reflecting the group health causal relationship is constructed based on the federated learning result; The double-graph weight updating module is configured to perform the following steps: Structurally converting the newly obtained user feedback information and health factor data to obtain a newly obtained health feature vector; Determine whether the newly obtained health feature vector can be classified into any leaf node of the clustering tree structure, If the newly obtained health feature vector can be classified into a leaf node of the clustering tree structure, the coordinator divides the newly obtained health feature vector to the corresponding participant based on the hierarchical federation architecture, and when the number of newly added health feature vectors received by the participant exceeds a preset threshold, incrementally trains the adaptive random forest model of the participant, updates the health feature weight vector of the participant, aggregates the health feature weight vector of the participant backward, updates the hierarchical weight matrix layer by layer, and obtains the weight of the updated group knowledge graph, If the newly obtained health feature vector cannot be classified into any leaf node of the clustering tree structure, the following extension audit steps are performed: Calculate the distance between the newly obtained health feature vector and the clustering center of the group, and when the distance is greater than a preset distance threshold, determine the nearest neighbor intermediate node based on the distance and create a corresponding new leaf node as a new participant; The coordinator divides the newly obtained health feature vector to the new participant based on the hierarchical federation architecture; Construct a local adaptive random forest model for the new participant, incrementally train the model of the new participant to complete initialization, calculate the health feature weight vector of the new participant, aggregate the health feature weight vector of the new participant backward, update the hierarchical weight matrix layer by layer, and obtain the weight of the updated group knowledge graph.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the sleep health dynamic evaluation method based on user feedback driving according to any one of claims 1 to 6.

Images