A sleep assessment intervention method and system of a hierarchical dynamic state space model

Abstract

The application discloses a sleep evaluation intervention method and system of a layered dynamic state space model, and belongs to the technical field of sleep health monitoring. The method comprises the following steps: collecting multi-modal data through adaptive question setting and information entropy control, and restoring original semantic states by using a semantic mapping matrix; a characteristic state vector containing 22 dimensions is constructed, and the characteristic state vector is divided into four function levels; a stable baseline vector is obtained, and characteristic weights are dynamically modulated to eliminate subjective statement bias; a dynamic Bayesian network is used to perform time series modeling on a sleep state evolution track, a progress latent variable is introduced to monitor rehabilitation or degradation trends in real time; after the characteristic state vector and the stable baseline vector are spliced, a risk probability is predicted, an image label is matched, and a structured report is output, and an optimal intervention path is generated based on Monte Carlo tree search. The application solves the problems of low efficiency, strong subjectivity and lack of dynamic tracking of traditional scales, and realizes high-precision and individualized sleep dynamic evaluation.

Description

Technical Field

[0001] This invention relates to the fields of medical information processing, bio-information feedback and big health monitoring technology, and in particular to a sleep assessment and intervention method and system based on a hierarchical dynamic state space model. Background Technology

[0002] With the rapid pace of modern life and fundamental changes in lifestyle, sleep disorders are no longer merely physiological phenomena, but have evolved into a global public health challenge affecting the physical and mental well-being of the public. In the fields of clinical medicine and mobile healthcare, various sleep assessment schemes have been developed to achieve accurate assessment and early intervention of sleep quality. Among them, traditional clinical scales, such as the Pittsburgh Sleep Quality Index (PSQI) and the Insomnia Severity Index (ISI), hold an authoritative position in current clinical diagnosis and scientific research due to their long-term clinical validation, rigorous indicator system, and broad academic consensus. These scales typically use a series of standardized questionnaire items to comprehensively score an individual's sleep status from multiple dimensions, including subjective sleep quality, sleep latency, sleep efficiency, sleep disorders, medication use, and daytime dysfunction, providing healthcare professionals with preliminary diagnostic references.

[0003] However, in practical applications, especially in long-term home health monitoring and universal health management scenarios, existing assessment technologies still face significant technical bottlenecks. First, traditional scales often suffer from severe semantic redundancy and low assessment efficiency. Because these scales aim to cover the broadest possible range of clinical indicators to ensure diagnostic rigor, they typically have a large number of items, requiring users to spend considerable time completing a full scale assessment. In the era of mobile internet and fragmented information, this high-intensity interaction easily leads to severe assessment fatigue and psychological burden on users, resulting in a significant decrease in data collection integrity and long-term adherence. This highly invasive assessment method not only reduces user experience but also fundamentally restricts the large-scale application of sleep monitoring technology in digital health management, making high-frequency sleep state tracking difficult to achieve.

[0004] Secondly, existing assessment methods generally lack the ability to dynamically track and temporally model the evolutionary trajectory of sleep states. Sleep health itself is a complex dynamic evolutionary process influenced by circadian rhythms, environmental stress, lifestyle interventions, and psychological fluctuations, exhibiting both continuity and time-varying characteristics. However, current static scale assessments essentially only provide "cross-sectional data" at specific points in time, failing to depict the complex trajectory of sleep state evolution over time. Within the existing technological framework, due to the lack of deep mathematical modeling of the temporal relationship between historical states, baseline characteristics, and current observations, the system cannot accurately capture changes in the slope of individual sleep recovery, nor can it identify early, subtle warning signals of health state decline. This static inference lacking a time dimension results in assessment results with significant lag, failing to provide forward-looking decision support for individualized interventions.

[0005] Furthermore, the lack of an effective individualized calibration mechanism is another core factor leading to assessment errors and compromising objectivity. Because self-report scales are inherently highly dependent on users' subjective statements, their assessment results are easily influenced by complex psychological effects such as individual response styles, emotional susceptibility, and social expectation bias. For example, in actual testing, some users tend to choose extreme rating scales, while others exhibit a clear tendency towards centralization. This systematic bias is difficult to identify and eliminate in traditional linear weighted models lacking baseline calibration. Because existing assessment systems cannot decouple an individual's long-term psychological baseline from immediate sleep fluctuations, they often produce "inflated" or "inflated" rating conclusions, severely compromising the objectivity of the assessment and the scientific rigor of the diagnosis.

[0006] Finally, existing assessment methods still have significant room for improvement in terms of the accuracy of risk prediction and the coupling of multidimensional features for inference. Traditional assessment methods are mostly based on simple weighted summation, logical discrimination, or fixed threshold determinations, lacking deep inference mechanisms that combine an individual's long-term baseline with multidimensional heterogeneous features. When dealing with complex sleep data exhibiting high individual variability, existing models often overlook the interactions and causal relationships between circadian rhythms, cognitive emotions, physical sleep function, and bodily regulatory abilities due to the lack of scientific hierarchical division of feature dimensions. This flattened assessment model struggles to achieve accurate individualized risk modeling and cannot generate structured intervention recommendations targeting specific underlying causes in each user.

[0007] In summary, in the current field of sleep health monitoring, how to develop an assessment scheme that can significantly reduce the burden of answering questions for users and improve assessment efficiency, while taking into account individual differences and eliminating subjective psychological biases through scientific mathematical models, and achieving accurate characterization and forward-looking prediction of the dynamic evolution of sleep states, has become a core technical challenge and research focus that urgently needs to be addressed in this field. Summary of the Invention

[0008] To address the aforementioned technical issues, a sleep assessment and intervention method and system based on a hierarchical dynamic state space model is provided. This technical solution solves the problems mentioned in the background art, such as the low efficiency of traditional sleep scales, the single data dimension, the lack of individualized calibration mechanisms, and the inability to achieve long-term dynamic evolution tracking.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] In a first aspect of the present invention, a sleep state assessment and intervention method based on a hierarchical dynamic state space model is provided, comprising the following steps:

[0011] S100, Multimodal Data Acquisition and Semantic Uncertainty Control: An adaptive dynamic question generation program is initiated via an intelligent analysis terminal, and the risk label... This is a probability distribution vector characterizing the type and severity of a user's sleep disorder. The risk label is calculated in real-time during the assessment process. posterior distribution entropy value This is used as a quantitative indicator for assessing uncertainty. Based on the expected information gain (IG), the question bank is dynamically sorted in real time, and the optimal questions that can minimize the current assessment uncertainty are selected and pushed to the user until the uncertainty index drops to a preset threshold. Automatic termination logic is triggered afterward.

[0012] S200, Multi-source heterogeneous vector construction and original semantic recovery: Obtaining users' immediate response feedback, basic demographic characteristics, and past medical history to construct high-dimensional multi-source heterogeneous vectors. And using a pre-defined semantic mapping matrix Restore it to the original semantic state vector ;

[0013] S300, Hierarchical Feature Space Mapping and Dimensional Partitioning: Utilizing the Feature Mapping Matrix The original semantic state vector Transformed into a feature state vector with 22 dimensions It is divided into four functional levels: circadian rhythm, sleep function, cognitive emotion, and physical regulation. The score of each functional level is calculated by fitting the dimensional indicators within the level according to the preset weight ratio.

[0014] S400, Personalized Baseline Calibration and Dynamic Weight Modulation: Obtain a stable baseline vector encompassing user personality traits, emotional susceptibility, and regulatory capacity. Obtain the initial weight vectors corresponding one-to-one with the 22 feature dimensions. Among the 22 feature dimensions, each feature dimension belonging to the same functional level is initially weighted according to the preset weight ratio of its functional level (the preset weight ratio for each level is: 0.25 for the circadian rhythm layer, 0.35 for the sleep function layer, 0.20 for the cognitive-emotional layer, and 0.20 for the physical regulation layer); using the stable baseline vector Initial weights for each feature dimension Dynamic modulation is performed to generate corrective weights that can eliminate subjective statement bias. .

[0015] S500, Temporal evolution modeling based on dynamic Bayesian networks: combining the aforementioned feature state vectors The stable baseline vector And historical state trajectory, using state evolution equations to characterize the migration trajectory of sleep state over time, and introducing personality modulation terms to ensure that the inference process is deeply modulated by the individual's inherent psychological baseline;

[0016] S600, Rehabilitation Pathway Monitoring and Health Deterioration Early Warning: Introducing Latent Variables of Time Progress It characterizes the degree to which an individual's state approaches the ideal health baseline and captures sleep recovery slope or early warning signals of decline in real time;

[0017] S700, Multidimensional Risk Inference and Imagery Portrait Representation: The feature state vector is compared with the stable baseline vector. After concatenation, the data is input into a machine learning model to predict the risk probability, and the cosine similarity between the user state and the prototype vector is calculated to match image tags with connotations of traditional Chinese culture.

[0018] S800, Optimal Intervention Path Selection Based on Monte Carlo Tree Search: Based on a unified semantic space mapping relationship, the Monte Carlo tree search algorithm is used to simulate the implementation effects of different intervention schemes under different execution orders and intensities, select the optimal intervention path, and generate personalized intervention schemes. (Based on feature state vectors) Stable baseline vector The root node of the Monte Carlo tree is formed by the risk probability predicted by S700. The intervention types corresponding to the 22 dimensions and four functional levels—nutrition, traditional Chinese medicine, psychology, and behavior regulation—are used as first-level nodes. Specific intervention methods under each type are used as second-level nodes, and the intervention execution order, intensity, frequency, and duration are used as leaf node attributes. The UCT upper confidence interval algorithm is used to complete path selection, node expansion, effect simulation, and value backtesting. A comprehensive scoring function is constructed by combining intervention effect, execution difficulty, and user compliance to select the optimal intervention path and generate personalized intervention plans that accurately match the 22-dimensional features and each functional level, ensuring compatibility with all input data and feature vectors. A closed loop is formed.

[0019] S900, Structured Report Generation: Automatically outputs a structured assessment report containing three dimensions: numerical, typological, and inference layers, and incorporating the S800 optimal intervention path and personalized intervention plan; the numerical layer includes scores for each of the 22 features and the total score for each functional level; the typological layer includes imagery labels and sleep state types; the inference layer includes risk probability and rehabilitation progress (based on...). Include intervention recommendations and ensure the report is complete and logically clear.

[0020] Preferably, the step of initiating the multimodal data acquisition and semantic recovery program through the intelligent analysis terminal and utilizing the adaptive dynamic question generation mechanism specifically includes the following steps:

[0021] S101, Question Bank Integration and Initialization: Traverse the preset integrated question bank in the background. This question bank is generated by integrating multiple clinical sleep scales through semantic clustering algorithm, covering multiple semantic dimensions such as difficulty falling asleep, sleep maintenance, and daytime function.

[0022] S102, Calculation of Expected Information Gain: For each unanswered question item, calculate the expected information gain, that is, simulate the impact on the risk label after the question item is answered. The degree of correction to the probability distribution;

[0023] S103. Dynamic sorting and push: The integrated question bank is dynamically sorted in real time according to the calculated information gain, and the question with the first ranking is pushed to the user first, so as to achieve the goal of obtaining the maximum amount of information with the least interaction.

[0024] S104. Real-time Probability and Entropy Updates: As users complete each question, the system updates the probability distribution of risk labels based on each real-time feedback. And according to the formula Reassess the posterior distribution entropy of the population;

[0025] S105, Trigger Automatic Termination: When the uncertainty index Decline to a pre-set clinically acceptable threshold If the information collected is deemed sufficient to accurately infer the user's sleep state, the question generation process will immediately cease and the subsequent data processing will commence.

[0026] Preferably, the step of mapping the compressed response vector to the original semantic state vector using a preset semantic mapping matrix specifically includes the following steps:

[0027] Establish multi-source heterogeneous vectors In addition to the instant score, it also includes unstructured multimodal information such as age, gender, sleep habits and past health records;

[0028] Through formula To achieve semantic recovery, It is a mapping operator constructed using semantic clustering algorithms based on massive clinical corpora and expert-annotated data;

[0029] Semantic completion of clinical information lost in compressed interactions within a mathematical space enables a small amount of interactive data to have diagnostic efficacy comparable to traditional complete scales (such as PSQI and ISI).

[0030] Preferably, the original semantic state vector recovered using the feature mapping matrix is... The transformation into a feature state vector with 22 dimensions includes the following steps:

[0031] For the original semantic state vector Perform high-dimensional reconstruction to generate a feature state vector containing 22 dimensions. ;

[0032] The 22 feature dimensions are divided into four functional levels. Each functional level uses a method of "independently calculating each dimension's index within the level and then fitting it according to a preset weight ratio" to obtain a comprehensive score for that level. Specifically:

[0033] The circadian rhythm layer contains multiple rhythm-related dimensional indicators. Each dimensional indicator is scored independently. After all dimensional scores are calculated, the circadian rhythm layer is fitted according to the preset weights within the layer to obtain the comprehensive score of the circadian rhythm layer.

[0034] The sleep function layer contains multiple sleep function-related dimensional indicators. Each dimensional indicator is calculated independently and then fitted according to the preset weights within the layer. Among them, the core dimension extracts sleep firmness by calculating the ratio of actual sleep onset time to bedtime.

[0035] The cognitive-emotional layer contains multiple psychological stress-related dimensions and indicators. Each dimension and indicator is calculated independently and then fitted according to the preset weights within the layer. The core logic is to extract scores of individual physical and mental regulation abilities by analyzing users' psychological stress responses in stressful scenarios.

[0036] The physical conditioning layer contains multiple individual trait-related dimension indicators. Each dimension indicator is calculated independently and then fitted according to the preset weights within the layer. The core is to provide an individualized reference background for the evaluation model by capturing the individual's sensitivity, susceptibility, and baseline rate of physical recovery.

[0037] Of the 22 feature dimensions, 5 are categorized into the circadian rhythm layer, 8 into the sleep function layer, 4 into the cognitive-emotional layer, and 5 into the physical regulation layer, consistent with the S300 mentioned earlier; simultaneously, the feature mapping matrix is ​​the same preset matrix. .

[0038] Preferably, the acquisition of a stable baseline vector encompassing user personality traits, emotional susceptibility, and self-regulation ability is... The specific steps for dynamically modulating the initial weights of the aforementioned dimensions using this baseline include the following:

[0039] S401. Based on clinical expert consensus and the hierarchical architecture of sleep medicine, the system obtains initial weight vectors that correspond one-to-one with the 22 feature dimensions. In the process of constructing the initial weight vector, the weight values ​​of each feature dimension belonging to the same functional level are quantitatively allocated according to the preset weight ratio of the level. The preset weight ratios for each level are: 0.25 for the circadian rhythm layer, 0.35 for the sleep function layer, 0.20 for the cognitive-emotional layer, and 0.20 for the physical regulation layer. This allocation mechanism ensures that... It can reflect the differences in importance of macro-functional levels, and also accurately apply to the 22 micro-level feature dimensions;

[0040] S402, Stable Baseline Vector Acquisition: A stable baseline vector is obtained by mining users' historical assessment behavior data or performing initial personality tendency tests. .in Representative personality traits (such as central tendency or extreme tendencies). Represents emotional susceptibility. It represents self-regulation ability;

[0041] S403, Weighted Dynamic Modulation Calculation: Using the formula Calculate the corrected weights ;in , , The preset modulation coefficients are used to adjust the contribution intensity of different baseline features to the final weights;

[0042] S404, Subjective Bias Filtering: Utilizing the aforementioned corrected weights For the feature state vector The original feature values ​​(i.e., instant feature values) of the 22 dimensions at the current moment are weighted and calculated to identify and filter extreme or moderate scoring biases caused by the user's specific answering tendencies, so that the final evaluation conclusion returns to the objective sleep function level.

[0043] Preferably, the step of using a dynamic Bayesian network to perform temporal modeling of continuous changes in sleep state specifically includes the following steps:

[0044] Construct a dynamic Bayesian network model, the state evolution equation of which is expressed as:

[0045]

[0046] in, This is the feature state vector at the current moment. This is the historical feature state vector from the previous moment; The vector of observations obtained through the adaptive dynamic question generation program; This is the state transition matrix, used to describe the spontaneous evolutionary inertia of the sleep state; The observation contribution weight matrix is ​​used to determine the degree to which the current observation contributes to the state update; This is a personality modulation matrix used to ensure that the evolution of states is deeply modulated by the individual's inherent psychological baseline; For a stable baseline vector; For random noise;

[0047] Through the state transition matrix Describe the spontaneous evolutionary inertia of sleep states under the absence of significant external disturbances;

[0048] Through personality modulation matrix With stable baseline vector The product term ensures that the evolution of the state is deeply modulated by the individual's inherent psychological baseline;

[0049] Preferably, the step of introducing a time-progress latent variable to characterize the degree to which the user's state approaches the ideal health baseline specifically includes the following steps:

[0050] Calculate the latent variables of the time schedule The formula for calculating the latent variable of the time schedule is:

[0051]

[0052] in, Based on the recovery rate; This is a weight vector used to incorporate individual baseline features. Correction to recovery speed; For risk feedback coefficient, This represents the probability of past risks, used to describe the hindering effect of risks on current progress. For random noise;

[0053] The value is strictly limited by the clip function. Within the timeframe, the individual's recovery trajectory is continuously monitored in numerical form;

[0054] Real-time analysis of a user's sleep status reveals whether it is in a recovery zone where sleep is steadily converging towards a healthy state, or in a fluctuating risk zone, or even in a warning zone showing a continuous deterioration trend.

[0055] Preferably, the generation mechanism based on fixed templates and label slots, which automatically outputs a structured health assessment report containing three dimensions—a numerical layer, a type layer, and an inference layer—specifically includes the following steps:

[0056] Compare the characteristic state vector with the stable baseline vector Perform feature concatenation to construct an enhanced input vector;

[0057] The enhanced vector is input into a pre-trained machine learning model to infer the user's current sleep risk probability;

[0058] Calculate the user's current comprehensive state vector, filter the optimal intervention path, and accurately match personalized intervention plans;

[0059] Calculate the cosine similarity between the vectors and a preset set of prototype vectors to match image tags for the user;

[0060] Based on the fixed templates of the numerical layer, typological layer, and inference layer, the total score, functional level score, imagery label, individualized intervention path, and plan are filled into the corresponding slots to generate a structured health assessment report.

[0061] In a second aspect of the invention, a sleep state assessment system based on a hierarchical dynamic state space model is also provided, comprising:

[0062] The data acquisition and semantic recovery module is used to execute adaptive question generation logic and obtain multimodal response vectors.

[0063] The hierarchical feature construction module is configured to construct the original semantic state vector. Mapped to a four-functional-level space containing 22 dimensions;

[0064] A baseline calibration module, which is used to acquire a stable baseline and dynamically modulate feature weights;

[0065] A dynamic evolution modeling module, which is used to construct a dynamic Bayesian network and track state transition trajectories;

[0066] Time progress latent variable monitoring module: used to capture sleep recovery slope or regression warning signals in real time, and to determine whether the user's sleep state is in the recovery zone, risk zone or warning zone.

[0067] Risk Prediction and Imagery Profiling Module: Used to predict risk probability, match imagery tags, and complete the imagery representation of the user's sleep state.

[0068] Intervention Path Optimization Module: The intervention path optimization module is used to screen the optimal intervention path and generate personalized solutions.

[0069] Structured report generation module: The structured report generation module is used to output a complete structured assessment report that clearly presents intervention recommendations.

[0070] In a third aspect of the invention, an electronic device is also provided. The electronic device includes at least one processor; and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of the first aspect of the invention.

[0071] Compared with existing technologies, the present invention provides a sleep assessment and intervention method and system based on a hierarchical dynamic state space model, which has the following beneficial effects:

[0072] This invention first reduces the average time users spend completing assessments from the traditional 22.6 minutes to 6.7 minutes through the synergistic application of an adaptive dynamic question generation mechanism and semantic mapping technology, reducing the burden by approximately 70%. This lightweight assessment experience greatly improves user compliance during long-term health monitoring.

[0073] Secondly, despite the significant reduction in the number of items, it still maintained extremely high clinical consistency and measurement reliability. The correlation coefficient between the assessment results and the gold standard scale reached 0.879, and the AUC value for risk prediction was as high as 0.906, ensuring the medical authority of the assessment conclusions.

[0074] Furthermore, the modeling approach based on dynamic Bayesian networks can accurately depict the evolutionary trajectory of an individual's sleep state, enabling quantitative tracking of rehabilitation progress or regression trends, which has significant technical advantages compared to static assessments that can only provide cross-sectional data.

[0075] Finally, the baseline calibration mechanism effectively filters out systematic biases in subjective statements, ensuring the objectivity and stability of the measurement results. Its unique hierarchical structure and image tag matching mechanism make the assessment report both rigorous scientific values ​​and highly interpretable, better assisting non-professional users in understanding and managing their own sleep health. Attached Figure Description

[0076] Figure 1 is a flowchart of a sleep state assessment and intervention method based on a hierarchical dynamic state space model provided by the present invention;

[0077] Figure 2 is a decision logic diagram of the adaptive dynamic question generation mechanism in this invention;

[0078] Figure 3 shows the feature state vector in this invention. A logical diagram of the four functional levels;

[0079] Figure 4 is a block diagram of a sleep state assessment and intervention system based on a hierarchical dynamic state space model provided by the present invention. Detailed Implementation

[0080] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0081] Example 1

[0082] refer to Figure 1 This invention is a sleep state assessment and intervention method based on a hierarchical dynamic state-space model, applicable to electronic devices or systems. This embodiment demonstrates the operational logic of each step of the invention through continuous monitoring of an individual with potential sleep disorders.

[0083] 1. Data Acquisition and Semantic Recovery Stage

[0084] S100, Multimodal Data Acquisition and Semantic Uncertainty Control: An adaptive dynamic question generation program calculates the posterior distribution entropy of risk labels in real time and selects the optimal questions based on expected information gain to push to users. Risk labels are calculated in real time based on multimodal data. posterior distribution entropy In this implementation, multimodal data includes physiological monitoring data, environmental perception data, lifestyle data, etc., and risk labels are used to characterize the degree of sleep disorder in users, which can be mapped to a preset level range (e.g., a 7-level classification). By calculating the expected information gain (IG) of unanswered items in the fusion question bank, the system prioritizes pushing items that minimize uncertainty to users (i.e., optimal items). When the entropy value drops to the clinical threshold... If the current information collected is deemed sufficient to accurately infer the user's sleep state, the question generation process is immediately stopped and the subsequent data processing flow begins, thereby reducing the assessment time from over 20 minutes to approximately 6 minutes. In one embodiment, S100 includes:

[0085] Multimodal data acquisition: Multimodal data includes users' real-time responses, physiological monitoring data, environmental perception data, lifestyle data, basic demographic characteristics and past medical history, etc., which can be obtained through user responses, wearable devices, IoT sensor data, demographic data and other means.

[0086] Question Bank Fusion and Initialization: A pre-defined fusion question bank, generated by fusing multiple clinical sleep scales using a semantic clustering algorithm, covering multiple semantic dimensions such as difficulty falling asleep, sleep maintenance, and daytime functioning; setting clinical thresholds.

[0087] Semantic uncertainty filtering: This involves filtering multimodal data. For example, subjective user responses are often ambiguous (e.g., "sometimes I don't sleep well"). The semantic ambiguity score (SAS) can be used to quantify the uncertainty, allowing only high-confidence data to enter the inference process. The formula for calculating the semantic ambiguity score is as follows:

[0088]

[0089] in, , representing the set of words in the user's answer; , represents the set of words in the response that match the "standard description of sleep disorders" (e.g., "difficulty falling asleep" and "early awakening" from the DSM-5 diagnostic criteria); f(w): the frequency of word w in the standard corpus. The closer the SAS is to 1, the more ambiguous the response; the closer it is to 0, the more specific the response. In one implementation, if the SAS exceeds a preset value (e.g., 0.7), the user needs to provide further clarification, or the ambiguous data needs to be deleted. In other implementations, semantic uncertainty filtering may not be necessary if the multimodal data meets the requirements.

[0090] Expected information gain calculation: To prioritize question recommendations, the posterior distribution is calculated using Bayesian methods, and the entropy value is calculated using the Shannon entropy formula. The posterior distribution is calculated as follows:

[0091]

[0092] in, This indicates the collection of multimodal data at time t. back, The probability distribution, Represent the prior distribution The likelihood function represents the probability of observed data given a risk r. The probability is calculated using a machine learning model trained on historical data, such as a random forest. This represents the normalization constant.

[0093] Entropy calculation as follows:

[0094]

[0095] It can be abbreviated as: ;

[0096] Where r corresponds to the risk label level. When the uncertainty index... Decline to a pre-set clinically acceptable threshold If the information collected is deemed sufficient to accurately infer the user's sleep state, the question generation process will immediately cease and the subsequent data processing will commence.

[0097] Iterate through the fusion question bank and calculate the expected information gain (IG) for each unanswered question, that is, simulate the impact on the risk label after the question is answered. The degree of correction of the probability distribution :

[0098]

[0099] in, This indicates a question item that has not been answered (i.e., a candidate question item). Indicates user's opinion Possible answers, It expresses expectation.

[0100] Dynamic sorting and push: The integrated question bank is dynamically sorted in real time based on the calculated information gain, and the top-ranked question is pushed to the user first, in order to achieve the goal of obtaining the maximum amount of information with the least interaction. :

[0101]

[0102] in, This represents the candidate item library.

[0103] Real-time probability and entropy updates: As users complete each question, the system updates the probability distribution of risk labels based on each real-time feedback. .

[0104] S200. Construction of Multi-Source Heterogeneous Vectors and Recovery of Original Semantics: A high-dimensional integrated response vector is constructed, and using a pre-defined semantic mapping matrix, it is recovered into the original semantic state vector. The system integrates user's immediate feedback, age, basic demographic characteristics, gender, and medical history to construct a high-dimensional multi-source heterogeneous vector. Using a pre-defined semantic mapping matrix Execute formula That is, using a pre-defined semantic mapping matrix Restore high-dimensional multi-source heterogeneous vectors to their original semantic state vectors In mathematical space, semantic completion is performed on the clinical information lost in the interaction, so that a small amount of interactive data has the diagnostic power of the gold standard scale.

[0105] 2. Feature Construction and Baseline Calibration Stage

[0106] S300, Hierarchical Feature Space Mapping and Dimension Partitioning: The system utilizes the feature mapping matrix The original semantic state vector Transformed into a feature state vector with 22 dimensions It is divided into four functional levels. These dimensions are scientifically divided into four functional levels: the circadian rhythm level (quantifying social jet lag), the sleep function level (extracting sleep firmness), the cognitive-emotional level (stress response), and the physical regulation level (restoring baseline rate).

[0107] S400, Personalized Baseline Calibration and Dynamic Weight Modulation: Personalized Baseline Calibration and Dynamic Weight Modulation: Obtain a stable baseline vector, and use this stable baseline vector to modulate the initial weights of the feature dimensions to generate corrected weights. Specifically, the system obtains the user's stable baseline vector. ( Indicates personality traits, Indicates emotional susceptibility, (Indicates regulatory capacity). Using the formula... Dynamically generate corrected weights It identifies and filters extreme rating biases in users' subjective statements to ensure that assessment conclusions reflect objective sleep function.

[0108] Among them, the initial weights It is a weight vector that corresponds one-to-one with the 22 feature dimensions. It is quantitatively allocated according to its functional level (0.25 for the circadian rhythm layer, 0.35 for the sleep function layer, 0.20 for the cognitive emotion layer, and 0.20 for the physical regulation layer) to quantify the basic contribution of each dimension in the standard evaluation model.

[0109] Subsequently, the system utilizes adjusted weights For the feature state vector The system performs a weighted calculation on the raw feature values ​​(i.e., instantaneous feature values) across 22 dimensions at the current moment. Through this process, the system can identify and filter out extreme or moderate scoring biases caused by the user's specific answering tendencies, ensuring that the final evaluation conclusion returns to the objective level of sleep function.

[0110] 3. Dynamic Modeling and Progress Monitoring Phase

[0111] S500, Temporal Evolution Modeling of Dynamic Bayesian Networks: A state evolution equation is used to characterize the migration trajectory of sleep states over time, and a personality modulation term is introduced to ensure that the inference process is deeply modulated by the individual's inherent psychological baseline. The system constructs the state evolution equation as follows: Through the state transition matrix Describing evolutionary inertia, and composed of a personality modulation matrix. Ensure that the evolutionary process is influenced by the individual's psychological baseline. Depth adjustment.

[0112] S600, Rehabilitation Pathway Monitoring and Health Degradation Early Warning: Introduces a time-progression latent variable to characterize the degree to which an individual's state approaches the ideal health baseline, capturing sleep recovery slope or degradation early warning signals in real time. The system incorporates a time-progression latent variable. The calculation formula is: This variable represents in real time how close a user's condition is to the ideal health baseline, accurately capturing the slope of sleep recovery or warning signals of decline.

[0113] 4. Risk Prediction, Intervention Path, and Report Output Stage

[0114] S700, Multidimensional Risk Inference and Imagery-Based Profile Representation: The feature vector and baseline vector are concatenated and input into a machine learning model to predict risk probability, matching it with imagery tags representing cultural connotations. The system will... and Concatenate the input vectors to construct an enhanced input vector, and then use a machine learning model to predict the probability of risk. Simultaneously calculate the current synthesis state and prototype vector. The cosine similarity is used to match image tags with connotations of traditional Chinese culture.

[0115] S800, Optimal Intervention Path Selection Based on Monte Carlo Tree Search: Based on a unified semantic space mapping relationship, the system uses the Monte Carlo tree search algorithm to simulate the implementation effects of different intervention schemes under different execution orders and intensities, selects the optimal intervention path, and generates personalized intervention schemes. The system uses feature state vectors... Stable baseline vector The root node of the Monte Carlo tree is composed of risk probabilities. The intervention types corresponding to the 22 dimensions and four functional levels—nutrition, traditional Chinese medicine, psychology, and behavior regulation—are used as first-level nodes. Specific intervention methods under each type are used as second-level nodes, and the intervention execution order, intensity, frequency, and duration are used as leaf node attributes. The UCT upper confidence interval algorithm is used to complete path selection, node expansion, effect simulation, and value backtesting. A comprehensive scoring function is constructed by combining intervention effects, execution difficulty, and user compliance to screen the optimal intervention path and generate personalized intervention plans that accurately match the 22-dimensional features and each functional level, ensuring compatibility with all input data and feature vectors. This forms a closed loop. The unified semantic space mapping relationship includes feature state vectors. Stable baseline vector Risk probability, first-level nodes, second-level nodes, and leaf nodes.

[0116] S900 Structured Report Generation: Automatically outputs a structured assessment report containing three dimensions: numerical, typological, and inference layers, incorporating the S800 optimal intervention path and personalized intervention plan. The numerical layer includes scores for each of the 22 features and the total score for each functional level; the typological layer includes imagery labels and sleep state types; the inference layer includes risk probability and rehabilitation progress (based on...). Include intervention recommendations and ensure the report is complete and logically clear.

[0117] Example 2

[0118] refer to Figure 2 In the sleep state assessment and intervention method based on a hierarchical dynamic state space model provided by this invention, Example 2 demonstrates in detail how to transform traditional cumbersome and redundant subjective scales into structured data vectors that can be efficiently processed by computers without losing clinical diagnostic accuracy through innovative multimodal data acquisition technology and semantic information recovery mechanism.

[0119] This implementation process is not only the data foundation for the entire stratified model, but also a key step in addressing core pain points such as poor user compliance and assessment fatigue from a technical perspective. In traditional clinical practice, doctors typically rely on scales such as the Pittsburgh Sleep Quality Index (PSQI) or the Insomnia Severity Index (ISI) for diagnosis. However, these scales often contain dozens of items, and users often need to spend more than 20 minutes to complete a single assessment, which is extremely difficult to maintain in long-term home monitoring scenarios.

[0120] This embodiment revolutionizes the data collection process by employing an adaptive dynamic question generation mechanism. When the assessment program starts, the system does not present questions in a fixed order, but instead calculates the current risk label in real time through an intelligent analysis terminal. The posterior distribution entropy value is used as a quantitative indicator for assessing uncertainty.

[0121] In the initial stages of the assessment, due to a lack of sufficient observations, the system's uncertainty is at its highest level. At this point, the system traverses a pre-set question bank generated by fusion of multiple clinical sleep scales (such as PSQI, ISI, ESS, etc.) using a semantic clustering algorithm, calculating the expected information gain for each unanswered item. This information gain calculation logic is similar to active learning strategies in machine learning, aiming to find the key dimensions that can minimize the entropy of the current assessment result. For example, if initial signs indicate that a user has difficulty falling asleep, the system dynamically adjusts the weights, prioritizing items related to pre-sleep anxiety or environmental disturbances, while automatically skipping semantically redundant parts irrelevant to the user's current state.

[0122] In the specific algorithm implementation, the system dynamically sorts the question bank in real time based on the calculated information gain, and prioritizes pushing the question with the highest ranking. As users complete the answers to each question, the system updates the risk tags based on each real-time feedback. The system calculates the probability distribution and reassesses the overall posterior distribution entropy. Once this uncertainty index drops below a preset clinically acceptable threshold, the system determines that the currently collected information is sufficient to support an accurate inference of the user's sleep state and immediately triggers automatic termination logic. This information theory-based decision-making approach reduces the average assessment completion time from the traditional 22.6 minutes to 6.7 minutes, significantly reducing the user burden by approximately 70% while ensuring the high quality and relevance of the collected data.

[0123] After obtaining dynamic user responses, the system enters the multimodal data vector construction phase. The system will then proceed to the next stage... Continuously extract multi-dimensional information from users to build a dimensional response vector This vector is a highly integrated feature space that not only includes the user's immediate score on the integrated questions, but also deeply integrates the user's basic demographic characteristics, such as age and gender, as well as the user's preset sleep habits, past health history, and long-term lifestyle records.

[0124] Multimodal fusion of this unstructured background information with the instantaneous response results makes the vector This approach provides a more comprehensive picture of an individual's sleep baseline than a single questionnaire score. It takes into account the differences in how users of different ages and health backgrounds respond to the same sleep problems, providing an individual-sensitive input source for subsequent precise modeling.

[0125] To address the risk of information loss caused by item compression, this embodiment introduces a core semantic information recovery algorithm. Since the fused items are significantly simplified, their direct scores cannot be directly compared to clinical gold standards. Therefore, the system utilizes a pre-set semantic mapping matrix. The compressed response vector Mapped to the original semantic state vector Its rigorous mathematical expression is defined as:

[0126]

[0127] Mapping matrix The system is built upon a massive clinical corpus and expert-annotated data, using semantic clustering algorithms to establish deep semantic connections between fusion items and original items from traditional clinical scales. Essentially, this step performs "super-resolution" semantic completion of missing clinical information within a mathematical space, enabling the system to infer a user's potential score on the full version of the PSQI or ISI scale through only a few key interactions.

[0128] Experimental data further confirms the effectiveness of this semantic recovery mechanism. Even in the extreme case where the number of items is reduced by 70%, after... Matrix mapping recovery The vector still demonstrated extremely high diagnostic consistency. Clinical validation showed that the assessment results generated by this system had correlation coefficients of 0.879 and 0.828 with the gold standard scales ISI and PSQI, respectively, and the area under the curve (AUC) for risk prediction was as high as 0.906.

[0129] This means that the algorithm successfully strikes a balance between a lightweight user experience and rigorous clinical needs. This semantic information recovery not only ensures measurement accuracy but also enables horizontal comparison of assessment results with internationally recognized standards, greatly enhancing the medical authority and practical application value of the assessment conclusions.

[0130] Finally, the original semantic state vector generated in this embodiment This lays the foundation for subsequent feature space mapping. It's not merely a set of values, but rather represents a nascent "sleep profile" encompassing multi-dimensional semantic dimensions. In this way, the system transforms raw subjective statements into standardized mathematical vectors, eliminating noise and ambiguity inherent in traditional dialogue. This structured data processing approach allows the invention to transform a subjective assessment process, which previously relied heavily on expert experience, into a computationally achievable, traceable, and dynamically evolving mathematical process over time, thus providing technical support for achieving high-frequency, predictive sleep health monitoring.

[0131] Example 3

[0132] In the sleep state assessment and intervention method based on a hierarchical dynamic state space model provided by this invention, Example 3 elaborates on how to transform the original semantic information into a sleep feature atlas with medical interpretability and multidimensional depth by constructing a hierarchical feature state space. The core of this process lies in utilizing the feature mapping matrix. The original semantic state vector recovered in the previous steps Perform high-dimensional reconstruction to generate a feature state vector containing 22 dimensions. .

[0133] To enable those skilled in the art to more clearly understand the feature state vector The following table shows the composition of the 22 dimensions across the four functional levels and their corresponding technical indicators:

[0134] Functional hierarchy Dimension number Technical Indicators / Dimension Names Explanation and Algorithm Basis Circadian rhythm layer (weight 0.25) 1-5 Diurnal pattern tendency, rhythm consistency, phase synchronicity, light regulation balance, social jet lag Each dimension is first scored using an independent scoring algorithm, and then a weighted fitting algorithm (based on the preset weights of each functional level) is used to calculate the comprehensive score for each functional level, ensuring the scientific validity and interpretability of the feature stratification. For example, the score for Dimension 1, diurnal pattern tendency, is the total score of the MEQ (Morning-Evening Questionnaire). The score for Dimension 2, circadian rhythm consistency, is calculated as 100 - (average wake-up time difference × 10). Sleep functional layer (weight 0.35) 6-13 Subjective sleep quality, sleep latency, sleep maintenance stability, sleep efficiency, sleep duration adequacy, daytime recovery level, sleep regulation ability, and tendency to be sleepy. For example: Dimension 6 Subjective Sleep Quality Score = 100 - [(PSQI Q1 standardized × 0.6) + (ISIQ4 standardized × 0.4)]. Dimension 7 Sleep Latency Score = (PSQI Component 2 standardized × 0.6) + (ISIQ1 standardized × 0.4). Cognitive Emotional Layer (Weight 0.20) 14-17 Sleep belief biases, sense of control over sleep, sleep-related anxiety, and mood regulation ability For example: Dimension 14 Sleep Control Score = (DBAS Control Score Inverse Standardization × 0.6) + (ISI_Q5 Inverse Standardization × 0.4). Dimension 16 Sleep-Related Anxiety Score = 100 - (PSAS Total Score ÷ Maximum Score × 100). Body constitution regulation layer (weight 0.20) 18-22 Innate constitution type, dynamic imbalance pattern, sleep constitution characteristics, dream integration ability, and physical and mental regulation ability For example: Dimension 20 Sleep Constitution Characteristics; (A) Ayurvedic Sleep Type Identification (Nidra Type) 1. Average score for each category (Vata / Pitta / Kapha) (range 0–4). 2. The highest score is the dominant sleep type. 3. If the difference between the two is <0.5, it is determined to be a dual sleep constitution. NidraType = Max(Vata,Pitta, Kapha) 4. If the difference between the three is <0.3, it is a balanced sleep constitution. (B) Traditional Chinese Medicine Sleep Constitution Weighted Integration (TCM Sleep Modulation) 1. Take the items related to sleep from the TCM constitution table: Yin deficiency, Yang deficiency, Qi stagnation, phlegm dampness, and damp heat as the five main influencing factors. 2. Calculate the "Sleep Bias Score": Bias Score = Balance Score - Average (Yin Deficiency, Yang Deficiency, Qi Stagnation, Damp-Heat, Phlegm-Dampness) 3. Standardize the bias score to 0–100: TCM_SleepIndex = 50 + Bias Score × 5 (C) Calculate the Comprehensive Sleep Quality Index (SCI): SCI = (Nidra Balance × 0.5) + (TCM_SleepIndex × 0.5)

[0135] Circadian rhythm layer: This layer includes not only social time difference It also uses five fine-grained indicators to characterize the homeostasis of the biological clock; for example, dimension 1 is specifically used to quantify the difference in the amplitude of work and rest on weekdays and rest days, in order to predict the risk of long-term insomnia.

[0136] Sleep functional layer: Contains 8 core physical indicators, through a feature mapping matrix The extracted "sleep firmness" dimension can effectively distinguish between physiological disorders and psychological sleep expectations, providing a physical basis for clinical intervention;

[0137] Cognitive-emotional layer: Includes four dimensions of psychological mediating variables. Through detailed analysis of user stress feedback, it identifies whether users are trapped in a vicious cycle of "anxiety-insomnia-more anxiety", thus providing a basis for judgment of cognitive behavioral therapy (CBT-I) for the report output layer;

[0138] Constitutional regulation layer: It includes 5 deep profile dimensions and introduces features similar to biological typology; for example, it can identify low-quality recovery of users with weak constitution through the baseline rate dimension, and accurately assess it through the weight modulation mechanism even if the duration is within the normal range.

[0139] In mathematical terms, this evolutionary logic is rigorously defined as , where the mapping matrix It is not a simple linear transformation, but a weighted mapping operator built on a sleep medicine knowledge base. It is responsible for transforming discrete semantic responses into a coupled biopsychological feature space. This transformation mechanism breaks through the limitations of traditional scales that merely "summarize scores," achieving a qualitative change from superficial description to structured representation, and providing fine-grained data support for subsequent individualized and precise assessments.

[0140] To ensure the assessment model captures both macroscopic sleep patterns and delves into underlying pathological and psychological factors, this embodiment scientifically divides the 22 feature dimensions into four independent yet interdependent functional levels. The first level is the circadian rhythm layer, which is assigned a weight of 0.25 within the overall assessment framework. It is primarily used to quantify an individual's diurnal pattern tendency and the consistency of their biological clock. In practice, the system calculates the circadian rhythm deviation index by collecting the difference between the user's wake-up time and bedtime on weekdays and free days (such as weekends). The core scoring algorithm for this level is defined as follows:

[0141]

[0142] The formula is adept at capturing the potential erosion of sleep quality caused by "social jet lag." As the score increases, the score decreases linearly, which directly reflects the disruption of biological rhythms caused by irregular work and rest schedules and is an important precursor indicator for predicting the risk of long-term insomnia.

[0143] The second layer is the sleep function layer, which serves as the core of the assessment system and has its weight set to the highest at 0.35. This layer focuses on the physical level of sleep recovery efficiency, encompassing key indicators such as sleep latency, sleep duration sufficiency, sleep efficiency, and subjective sleep quality. In the specific algorithm processing, the system not only focuses on the duration of a single sleep episode but also extracts the "firmness" of sleep through a feature mapping matrix, that is, correcting for single duration deviations by calculating the ratio of actual sleep onset time to bedtime. The innovation of this layer lies in its ability to transform the user's subjective "whether they slept well or not" into a quantifiable functional recovery index. Through dynamic monitoring of sleep latency, it can effectively distinguish between physiological difficulty falling asleep and psychological sleep expectation, thus providing the most direct physical basis for clinical intervention.

[0144] The third level, defined as the cognitive-emotional layer, is assigned a weight of 0.20. Its initial purpose is to address the issue of psychological mediating variables that are easily overlooked in traditional assessments. This level focuses on describing users' biases in sleep beliefs, such as excessively compulsive demands regarding sleep duration, and negative emotional tendencies strongly associated with sleep, such as anxiety and depression. During feature extraction, the system analyzes users' psychological stress responses in stressful scenarios to extract indicators of individual physical and mental regulation abilities. The introduction of this dimension has significant clinical value because the root cause of many chronic insomnia patients is not physiological disorder, but rather "learned insomnia" caused by erroneous sleep cognitions. At this fine-grained dimension of the vector, the model can identify whether a user is caught in a vicious cycle of "anxiety-insomnia-more anxiety," and thus provide accurate recommendations for cognitive behavioral therapy (CBT-I) in the generated report.

[0145] The fourth layer is the constitution and regulation layer, with a weight set at 0.20. It aims to characterize an individual's innate constitution type, dynamic imbalance patterns, and the body's self-regulatory boundaries in response to the external environment. The innovation of this layer lies in its introduction of a deep profiling dimension similar to "Traditional Chinese Medicine constitution" or "Ayurvedic medicine." By capturing an individual's sensitivity, susceptibility, and baseline rate of physical recovery, it provides a crucial individualized reference background for the evaluation model. For example, for users with a weak constitution and poor regulatory capacity, even if their sleep duration is within the normal range, the system will identify their lower recovery quality through a weighted modulation mechanism. This meticulously designed layered architecture ensures... The feature state vector is not only a static snapshot, but also a "sleep hologram" with a multi-dimensional dynamic view, which enables the assessment system to maintain extremely high resolution and diagnostic reliability when faced with highly individualized sleep data.

[0146] Example 4

[0147] In the sleep state assessment intervention method based on a hierarchical dynamic state space model provided by this invention, Example 4 details the specific implementation scheme of the individualized baseline calibration mechanism. This mechanism aims to eliminate psychological biases generated by users' subjective responses through mathematical means, ensuring the objectivity of the assessment conclusions.

[0148] To effectively decouple an individual's long-term psychological baseline from immediate sleep fluctuations, this embodiment introduces a stable baseline vector. .in, Representative personality traits (such as central tendency or extreme tendencies). Represents emotional susceptibility. These dimensions represent self-regulation ability. They are not derived from a single assessment, but rather from historical assessment data or initial personality tendency tests, providing the model with a "scale" to measure users' subjective statements.

[0149] In the specific algorithm implementation process, the system first obtains the initial weight vectors that correspond one-to-one with the 22 feature dimensions. The quantization allocation of this vector follows a hierarchical principle: dimensions belonging to the circadian rhythm layer are assigned a weight of 0.25, the sleep function layer 0.35, the cognitive-emotional layer 0.20, and the physical regulation layer 0.20. The system utilizes a stable baseline vector. For the initial weight vector Dynamic modulation is performed to generate a corrected weight vector that reflects the true state of the individual and also has 22 dimensions. .

[0150] The revised weight calculation formula is rigorously defined as follows:

[0151]

[0152] in, , , The preset modulation coefficients are used to adjust the contribution intensity of different baseline features to the weight correction.

[0153] In practical sleep assessment, the system utilizes adjusted weights. The "instantaneous feature values" are weighted and calculated. Here, "instantaneous feature values" specifically refer to those at the current time. From the original semantic state vector

[0154] via the feature mapping matrix The original values ​​of the 22 dimensions after transformation, before baseline calibration. Through this weighting process, the system can identify and filter out inflated or under-scoring biases caused by specific user response tendencies, thus bringing the assessment conclusions back to the objective sleep function level. The object of correction here is the feature state vector. The instantaneous raw feature values ​​of each dimension (i.e., the initial scores directly mapped and generated after the user answers).

[0155] For example, a person who is chronically anxious (high emotional susceptibility) Users who consistently rate their sleep efficiency very low, even when their sleep duration is normal; in such cases, the calibration mechanism will identify their... The system dynamically reduces the weight of the user's original scores for the relevant dimensions of "sleep function." By lowering the contribution weight of these original dimensions, which are heavily influenced by subjective emotions, the system can filter out "emotional rating bias," allowing the final assessment conclusion to return to the objective level of sleep function.

[0156] This bias filtering logic not only improves the accuracy of the assessment but also significantly enhances the stability of the measurement results. Experimental data show that after introducing this mechanism, the system's test-retest reliability (ICC) reached an extremely high level of 0.901.

[0157] Furthermore, the baseline calibration mechanism provides crucial background features for subsequent risk prediction. During the assessment report generation process, the system not only focuses on the immediate feature state vector... I will also refer to The individual's self-regulation ability reflected in the data If a user's short-term sleep metric declines, but their... The baseline data shows that the user possesses good psychological resilience and adjustment capabilities. When the system outputs suggestions at the inference layer, it correspondingly lowers the warning level and provides more guiding self-repair suggestions. This deep, individualized calibration ensures that each assessment report accurately reflects the user's true physical and psychological limits.

[0158] Example 5

[0159] In the sleep state assessment and intervention method based on a hierarchical dynamic state space model provided by this invention, Example 5 details how to use a dynamic Bayesian network (DBN) to accurately model and track the complex trajectory of an individual's sleep state evolution over time. This step is a key aspect of this invention in addressing the core pain point of existing technologies that lack dynamic tracking capabilities and cannot characterize the trajectory of sleep state evolution over time. Traditional sleep assessments are mostly static observations in a cross-sectional manner, making it difficult to capture the dynamic evolution process under the combined effects of biological rhythms, environmental stress, and intervention measures. This embodiment, by introducing mathematical derivation in the time dimension, views an individual's sleep health state as a dynamic process that continuously migrates in a characteristic state space, thereby achieving real-time capture of sleep recovery slope or early warning signals of health decline.

[0160] In the core state transition calculation process, this invention adopts a dynamic Bayesian network model that describes the continuous changes in sleep state, and its state evolution equation is rigorously expressed as:

[0161]

[0162] In this equation, Represents the current number Each time step contains a feature state vector with 22 dimensions, while This represents the historical state at the previous moment. Matrix As a state transition matrix, it describes the spontaneous evolutionary inertia of sleep states without significant external disturbances; This represents the vector of observations obtained at the current moment through the adaptive question-generating mechanism, while the matrix... This determines the contribution weight of the current observation to the state update. The most innovative part is the introduction of a value derived from the personality modulation matrix. With stable baseline vector The product term ensures that the evolution of the state depends not only on current performance but also on the deep modulation of an individual's inherent personality traits, emotional susceptibility, and self-regulation abilities. Furthermore, the term at the end of the formula... This represents the random noise term, used to absorb sudden environmental disturbances or measurement errors not covered by the model, ensuring the robustness of the model in practical applications.

[0163] To more intuitively quantify a user's recovery progress or regression trend, this embodiment also introduces a key latent variable: time progress. This variable is used to characterize how close a user's current state is to an ideal health baseline. Its dynamic updates are influenced by the combined effects of multiple parameters, and its calculation formula is as follows:

[0164]

[0165] In this formula, This represents the basic recovery rate and reflects the self-repair potential of the human sleep system; Items through weight vectors The introduction of individual baseline characteristics to adjust for recovery speed means that users with stronger self-regulation abilities will exhibit a faster recovery slope under the same intervention. The item describes the hindering effect of past risk probabilities on the current recovery progress, in which This is the risk feedback coefficient. The clip function is used to... The value is strictly limited to Within the timeframe, the system can continuously monitor an individual's recovery trajectory in numerical form.

[0166] In practical sleep assessment, this modeling approach based on dynamic Bayesian networks has demonstrated significant technological advantages. It can analyze in real time whether a user's sleep state is in a "recovery zone" steadily converging towards a healthy state, a fluctuating "risk zone," or even a "warning zone" showing a continuous deterioration trend. Experimental data shows that, because this model can accurately depict the evolutionary trajectory and provide quantitative intervention guidance, patients receiving personalized intervention recommendations from this system experienced a significantly greater decrease in their Insomnia Severity Index (ISI) compared to the control group that only received conventional intervention.

[0167] Specifically, clinical controlled studies showed that the experimental group experienced a decrease in ISI of -9.0 after 8 weeks, while the control group only saw a decrease of -5.0. This fully demonstrates the clinical effectiveness of dynamic evolutionary modeling in assisting individualized decision-making. This paradigm shift from "static scoring" to "dynamic trajectory tracking" not only improves the predictability of the assessment but also provides a scientific mathematical basis for early intervention in sleep disorders.

[0168] The logical architecture of this embodiment ensures that sleep health management is no longer a blind, point-by-point recording process, but a structured, continuous, and forward-looking inference process. Through the... By coupling the 22 features in the vector along the time axis, the system can keenly identify changes in dominant risk factors, such as the evolution from initial "environmental disturbances" to later "cognitive biases" or "rhythm disorders," thereby dynamically adjusting intervention strategies. This deep dynamic monitoring capability, combined with the aforementioned semantic information recovery and baseline calibration mechanisms, constitutes the mathematical foundation of the high-precision assessment system of this invention.

[0169] Example 6

[0170] refer to Figure 3 In the sleep state assessment and intervention method based on a hierarchical dynamic state space model provided by this invention, Example 6 demonstrates how to transform the high-dimensional, dynamic feature state vector extracted by the dynamic Bayesian network in the aforementioned steps into a structured health assessment report with high-precision predictive ability and humanistic care. This stage is not only the end point of data processing, but also the starting point for decision-making to achieve individualized and precise intervention. The system first executes the risk prediction logic, taking the current moment... Feature state vector With a stable baseline vector that reflects the user's psychological baseline Feature concatenation is performed to construct an enhanced input vector. Its mathematical expression is The innovative significance of this splicing mechanism lies in its deep, non-linear coupling of a user's instantaneous sleep function performance with long-term personality traits and emotional susceptibility.

[0171] Subsequently, enhanced vector The data is fed into a pre-trained machine learning model, such as a random forest or a deep neural network, to infer the user's current probability of sleep risk. Experimental data shows that this predictive model, which combines dynamic features with individual baselines, exhibits excellent performance, with an area under the curve (AUC) of 0.906 for risk prediction, significantly outperforming traditional methods that rely solely on static scores. This means that the system can not only inform users of their current sleep quality but also accurately predict the potential risk of it developing into chronic insomnia or leading to mental health deterioration.

[0172] To bridge the gap between professional medical numerical values ​​and the understanding of ordinary users, this embodiment introduces a culturally meaningful "image tag matching" mechanism. The system calculates the user's current comprehensive state vector and compares it with a preset set of prototype vectors. We use the cosine similarity between them to find the profile representation that best fits the user's current state.

[0173] By calculating the current user feature state vector With each prototype vector cosine similarity , the system can quantify the degree of approximation between the user's current sleep profile and the standard pathological / psychological model. The one with the highest similarity is selected The corresponding imagery tag is used as the output, aiming to compress the complex 22-dimensional values into intuitive, highly interpretable cultural imageries to assist non-professional users in understanding their underlying sleep logic. The calculation formula of cosine similarity is defined as:

[0174]

[0175] where, the preset prototype vector is a standard state center vector obtained by performing feature extraction and cluster analysis on massive clinically confirmed sleep disorder sample data. Specifically, the system uses historical desensitized data, for different sleep disorder types (such as difficulty falling asleep type, early awakening type, rhythm disorder type, etc.) and different psychological baseline characteristics, through or a hierarchical clustering algorithm to calculate the cluster center of each type, and uses it as the prototype vector stored in the prototype knowledge base of the system.

[0176] Each prototype vector is pre-associated with an imagery tag with connotations of traditional Chinese culture. The system will match the user with the "Shanhaijing" imagery prototype tag with traditional Chinese cultural connotations according to the result with the highest similarity. For example, the prototype vector is characterized by high time rhythm deviation ( ) but the sleep function layer score is acceptable. The associated imagery tag of the prototype vector corresponding to this state is "Bifang", which symbolizes that the individual is in a compensatory state of being tough but with rhythm overdraft. For users with biological rhythm disorder but acceptable physical fitness, they may be matched with the "Bifang" tag, suggesting that they are in a tough but overdrawn compensatory state; while for users with difficulty falling asleep and high anxiety, a more appropriate imagery may be matched. This tagged representation not only increases the intuitiveness and interestingness of the assessment, but also endows the report with a sense of humanistic care, thereby better assisting non-professional users in understanding complex sleep physiological indicators.

[0177] The imagery tag described in the present invention is technically defined as a category semantic index in the high-dimensional sleep state vector space. Its technical attributes are reflected in: mathematical uniqueness, each imagery tag has a one-to-one mapping relationship with a 22-dimensional prototype vector generated by clustering historical clinical samples ; data compression feature, its essence is to convert the complex, non-intuitive 22-dimensional continuous feature vector Transforming into discrete semantic categories enables feature compression and dimensionality reduction representation of high-dimensional sleep profiles; human-computer interaction optimization reduces the cognitive load for non-professional users to understand the 22-dimensional numerical distribution by using semantically suggestive labels (such as "Jingwei" representing circadian rhythm overdraft), thereby guiding users to more accurately execute the individualized intervention suggestions given by the inference layer.

[0178] Ultimately, the system automatically outputs a structured health assessment report based on a "fixed template + tag slot" generation mechanism. The generation formula is as follows: The report is meticulously divided into three levels to meet varying information needs. The first level is the numerical level, visually displaying the overall score, four functional level scores, and 22 dimensions of sleep characteristics, providing professionals with detailed quantitative data. The second level is the typological level, focusing on presenting the matched image archetypes and the current sleep state type, helping users build a concrete self-awareness. The third level is the inferential level, the most clinically valuable part of the report. The system analyzes dominant risk factors (such as specific cognitive biases or circadian rhythm fluctuations) and provides targeted, individualized intervention recommendations.

[0179] Clinical controlled studies have further confirmed the remarkable effectiveness of this structured reporting system and its accompanying intervention recommendations. Data shows that patients who received personalized recommendations generated by this system and underwent intervention experienced a decrease in their Insomnia Severity Index (ISI) of -9.0 after 8 weeks, significantly better than the -5.0 in the group receiving only conventional intervention. This significant intervention effect fully demonstrates the immense value of the hierarchical dynamic modeling and structured reporting output mechanism employed in this invention in practical clinical applications. Through this end-to-end digital management, this invention successfully constructs a closed loop for sleep health management, from accurate data collection and dynamic modeling to scientific feedback.

[0180] Example 7

[0181] In one embodiment, to provide intervention plans based on sleep assessment results, this embodiment also establishes cross-domain intervention features corresponding to feature state vectors and functional levels. These cross-domain intervention features are stored in a cross-domain intervention feature library. Each cross-domain intervention feature includes information such as a unique code for the intervention type (including Traditional Chinese Medicine, nutrition, psychology, yoga, behavior, and environment), the intervention unit, a unique code, the core function, and the applicable scenario. For example, an example of cross-domain intervention features stored in the cross-domain intervention feature library is as follows:

[0182] Intervention type Intervention Unit Unique Code core role Nutrition Magnesium supplementation intervention N001 Regulate neurotransmitters and improve difficulty falling asleep Traditional Chinese Medicine Head meridian regulation T003 Relieves sleep anxiety and stabilizes mood psychology Cognitive Behavioral Training (CBT-I) P002 Correcting sleep cognitive biases and prolonging deep sleep

[0183] In other embodiments, each intervention type in the cross-domain intervention feature library may include one or more intervention methods, such as traditional Chinese medicine, nutrition, yoga, CBT-I, light therapy, massage, etc.

[0184] This embodiment establishes a first-level mapping table associated with the feature state vector, and also establishes a second-level mapping table associated with the functional hierarchy. The first-level mapping table includes the feature state vector. The table includes the relevant intervention units and matching coefficients for each of the 22 dimensions. An example of the first-level mapping table is shown below:

[0185] Feature state vector dimension Related intervention unit (coding) Matching coefficient Clinical validation basis Sleep firmness (sleep functional layer) N001 (Magnesium Supplement) 0.85 Clinical trials have shown that users with a score of <60 on this trait experienced an average score increase of 15 points after magnesium supplementation. Sleep anxiety (cognitive-emotional level) T003 (Head Meridian Regulation) 0.92 Validated by 300+ samples, anxiety characteristic scores decreased by an average of 20 points after meridian regulation.

[0186] The secondary mapping table includes cross-domain intervention features corresponding to each functional level. An example of a secondary mapping table is shown below:

[0187] Functional hierarchy Corresponding intervention type Related basis Time rhythm layer Nutrition + Environmental Intervention This hierarchical characteristic (such as social jet lag) is mainly related to nutrient regulation and light intervention. Cognitive Emotional Layer Psychology + Traditional Chinese Medicine This level of characteristic (such as sleep anxiety) requires coordinated intervention through emotional guidance and meridian regulation.

[0188] To provide intervention plans based on sleep assessment results, this embodiment extracts dimensions from the feature state vector that are below a preset score (dimensions with scores < 70 are selected, indicating the need for intervention) as the dimensions to be intervened. It then obtains the associated intervention units and matching coefficients corresponding to these dimensions from a first-level mapping table. These units are sorted in descending order of matching coefficients, and the top three are selected to form a candidate intervention set. Next, based on a second-level mapping table, the candidate intervention set is selected to cover cross-domain intervention types at the corresponding level (for example, after determining the candidate intervention set, the corresponding intervention type can be determined from the cross-domain intervention feature library using the corresponding code, and then, combined with the association between the 22-dimensional feature state vector and the four functional levels, the corresponding intervention type is obtained from the second-level mapping table to ensure that the candidate intervention set covers the cross-domain intervention types at the corresponding level. For example, when the dimension to be intervened belongs to the cognitive-emotional layer, the second-level mapping table indicates that the cognitive-emotional layer must include psychology + traditional Chinese medicine interventions; therefore, it is necessary to ensure that the intervention type of the candidate intervention set belongs to psychology + traditional Chinese medicine). The standardized intervention matching results are output (including intervention unit codes, matching degree, usage suggestions, intervention type, intervention method, etc.).

[0189] Example 8

[0190] To evaluate the effectiveness of the intervention program, this embodiment will simulate the implementation effects of the intervention program under different execution sequences and intensities based on Monte Carlo tree search, in order to screen the optimal intervention path.

[0191] The Monte Carlo tree has a four-layer structure, with the root node containing the feature state vector, stable baseline vector, and risk probability; the first-level nodes are the intervention types included in the intervention plan; the second-level nodes are the intervention methods under each intervention type; and the leaf nodes are the execution order, execution intensity, execution days, and expected effect.

[0192] 1. Root node: User's current sleep state (corrected 22-dimensional feature state vector) + Stable baseline vector The risk probability output by the risk prediction module serves as the initial input for Monte Carlo tree search, encompassing individual user characteristics, baseline attributes, and risk status, providing core data support for path selection.

[0193] 2. First-level node: Intervention type (corresponding to the four functional levels of the 22-dimensional features, specifically including behavior regulation, psychology, traditional Chinese medicine, and nutrition). Among them, behavior regulation can cover environmental intervention-related content, and there is no need to set "environment" and "auxiliary support" as first-level nodes separately, to ensure consistency with the module's preset intervention types.

[0194] 3. Second-level nodes: Specific intervention methods (corresponding to each primary intervention type, such as TCM treatment and massage for TCM, nutrition for nutrition and dietary guidance, psychology for CBT-I, and behavior regulation for yoga and light therapy), to achieve a precise correspondence between intervention types and specific methods.

[0195] 4. Leaf Node: Intervention Execution Attributes (Execution Order + Execution Intensity + Duration + Expected Effect Score). The "Execution Days" has been changed to "Duration" to better align with clinical intervention descriptions. At the same time, the expected effect score is retained to support the calculation of the comprehensive scoring function and ensure consistency with the leaf node attribute requirements in the module.

[0196] The search process employs the classic 4-step Monte Carlo tree algorithm, entirely based on quantized scores:

[0197] 1. Selection

[0198] Starting from the root node, select the most promising intervention path according to the upper confidence limit formula (UCT), prioritizing branches with high user confidence, low execution difficulty, and high risk matching.

[0199] 2. Expansion

[0200] For the selected path, generate new solution nodes with different orders and intensities. For example, expand the path by psychological first and then physical, or by behavioral first and then environmental, or by intensity level, such as light, medium, and strong.

[0201] 3. Simulation

[0202] The sleep improvement effect of the preset plan is quickly simulated in a virtual space for 1 to 7 days. The output includes the risk reduction rate, sleep score improvement, user acceptance, and persistence probability, all of which are quantitative values.

[0203] 4. Backpropagation

[0204] The simulation results score is fed back to update the entire tree, and the path with the best performance and highest acceptance is marked.

[0205] Repeat the above 4 steps until the optimal path is found.

[0206] For example, searchable paths:

[0207] 1. Path A: Mind (Mindfulness) → Body (Traditional Chinese Medicine Massage) → Behavior (Day and Night) → Environment (Dim Light)

[0208] 2. Path B: Body (Nutrition) → Mind (Relaxation) → Behavior (Blue Light Withdrawal)

[0209] 3. Path C: Behavior (fixed wake-up time) → Environment (noise reduction) → Body (yoga)

[0210] 4. Path D: Traditional Chinese Medicine (calming the mind and soothing the nerves) → Yoga (breathing) → Psychological (emotional release)

[0211] When determining the optimal path, it can be calculated based on effectiveness score, belief level, execution difficulty coefficient, and persistence level, with the path scoring the highest being the optimal path. For example, a comprehensive scoring function can be constructed, combining intervention effectiveness score, execution difficulty coefficient, and user compliance index to calculate the total score for each path, and the path with the highest total score can be determined as the optimal intervention path, generating a personalized intervention plan that matches user characteristics.

[0212] In one embodiment, the best effect (40%) represents the greatest reduction in sleep risk and the greatest improvement in sleep quality; the highest user trust (30%) represents the recommended method that users trust, agree with, and are willing to implement; the most reasonable execution order (20%) represents from easy to difficult, from light to heavy, and in line with the user's psychological pattern; and the ability to persist and achieve a closed loop (10%) represents the ability to form a habit and ultimately lead to an autonomous self-closed loop.

[0213] Example 9

[0214] refer to Figure 4 This embodiment details the internal architecture and collaborative working mechanism of a sleep state assessment system based on a hierarchical dynamic state space model. The system aims to integrate the aforementioned methodological logic into an electronic device, achieving closed-loop management from bottom-level data acquisition to high-level risk warning through the interaction of seven core functional modules.

[0215] 1. Data Acquisition and Semantic Recovery Module

[0216] This module serves as the system's perception front end, executing adaptive question generation logic through interaction with the intelligent analysis terminal.

[0217] Interactive control: The module calculates the posterior distribution entropy of risk labels in real time, dynamically filters and pushes the optimal items, and obtains the highest information gain with the least interactive load.

[0218] Multimodal fusion: The module obtains multimodal response vectors that include the user's immediate score, basic demographic features, and past medical history. .

[0219] Semantic completion: using a pre-defined semantic mapping matrix The operation will be performed. Within the mathematical space, semantic completion is performed on the information lost in the interaction to generate the original semantic state vector. .

[0220] 2. Hierarchical Feature Construction Module

[0221] This module is responsible for transforming the recovered semantic vectors into structured feature maps with medical interpretability.

[0222] High-dimensional mapping: utilizing the feature mapping matrix will Mapped into a 22-dimensional space, generating feature state vectors. .

[0223] Hierarchical logical division: The module divides the 22 dimensions into a time rhythm layer, a sleep function layer, a cognitive emotion layer, and a physical regulation layer, providing multidimensional data support for subsequent classification and assessment.

[0224] 3. Baseline calibration module

[0225] This module is the core calibration unit that ensures the objectivity of system evaluation.

[0226] Individual baseline acquisition: The system obtains a stable baseline vector reflecting a user's personality traits, emotional susceptibility, and regulation ability by mining historical data or performing propensity tests. .

[0227] Dynamic weight modulation: module utilization For the preset initial weights Perform online correction; the calculation formula is as follows: .

[0228] Bias filtering: by adjusting weights Filter out subjective biases from users to ensure that subsequent evaluation conclusions reflect objective sleep function.

[0229] 4. Dynamic Evolution Modeling Module

[0230] This module uses a dynamic Bayesian network as a computing engine to track the transition trajectory of sleep states over time.

[0231] State evolution equation: Module construction equation Among them, personality modulation item This ensures that evolutionary inferences are deeply modulated by the individual's inherent psychological baseline.

[0232] Real-time progress monitoring: Module calculates latent variables of time progress. It can characterize the degree to which an individual's state approaches the ideal health baseline in real time, thereby capturing the slope of sleep recovery or early warning signals of deterioration.

[0233] 5. Time Schedule Latent Variable Monitoring Module

[0234] A time-based latent variable calculation algorithm is used to stabilize the baseline vector. Probability of past risks and historical time progress latent variables As the core input, through formula Real-time calculations are performed, where μ is the system's preset base recovery rate. η is the weight vector (used to correct the impact of individual baseline characteristics on recovery speed), and η is the risk feedback coefficient (quantifying the hindering effect of past risks on current recovery progress). This is a random noise term; the clip function is used to... Strictly limited to the [0,1] range, combined with dynamic threshold judgment logic, the system analyzes in real time whether the user's sleep state is in the recovery zone, risk zone, or warning zone, and outputs standardized progress monitoring results.

[0235] 6. Risk Prediction and Imagery Profiling Module: This module is responsible for transforming abstract mathematical states into intuitive assessment conclusions.

[0236] Multidimensional risk inference: The module will and The data is then fed into a machine learning model to predict the current probability of sleep risk.

[0237] Cultural imagery matching: By calculating the cosine similarity between the current integrated state vector and the prototype vector, imagery tags with Chinese traditional cultural connotations are matched for the user.

[0238] 7. Intervention Path Optimization Module

[0239] The Monte Carlo tree search algorithm is used with the corrected 22-dimensional feature state vector. Stable baseline vector Using the risk probability output from the risk prediction module as the core input, a Monte Carlo tree is constructed as its root node. Based on the four functional levels corresponding to the 22-dimensional features, intervention types—nutrition, traditional Chinese medicine, psychology, and behavior regulation—are designated as first-level nodes, with specific intervention methods for each type designated as second-level nodes. The intervention execution order, intensity, frequency, and duration are designated as leaf node attributes. The UCT upper confidence interval algorithm is used to perform iterative search, sequentially completing path selection, node expansion, effect simulation, and value backtracking. A comprehensive scoring function is constructed by combining intervention effect scores, execution difficulty coefficients, and user compliance. The optimal intervention path with the highest score is selected, generating a personalized intervention plan that precisely matches the user's 22-dimensional features and each functional level.

[0240] 8. Structured Report Generation Module

[0241] A report structure generation algorithm is adopted to integrate the core data output from each module and divide it into four core dimensions (numerical layer, type layer, inference layer, and intervention layer) according to a preset report framework. The numerical layer integrates the scores of each dimension of 22 features, the comprehensive scores of the four functional levels, and the latent variable of time progress. Numerical data; the type layer integrates imagery tags, sleep state types, and risk levels; the inference layer integrates risk probability, rehabilitation progress assessment, and degeneration warning prompts; the intervention layer integrates the optimal intervention path and personalized intervention plan; through a standardized format algorithm, the system automatically typesets and generates structured reports, ensuring that the reports are logically clear, data is accurate, and key points are highlighted, and can be directly used for clinical reference or user self-monitoring.

Claims

1. A sleep assessment and intervention method using a hierarchical dynamic state-space model, characterized in that, Includes the following steps: An adaptive dynamic question generation program is initiated through an intelligent analysis terminal to collect multimodal data and calculate risk labels in real time to characterize the degree of sleep disorder in users. The posterior distribution entropy value is used to select the best items and push them to the user. The multimodal data includes physiological monitoring data, environmental perception data, and lifestyle data. By obtaining real-time user responses and combining them with basic demographic characteristics and past medical history, a high-dimensional multi-source heterogeneous vector is constructed. And using a pre-defined semantic mapping matrix The multi-source heterogeneous vectors are restored to their original semantic state vectors. ; Using the feature mapping matrix The original semantic state vector Transform into feature state vector and the feature state vector It is divided into functional layers: the circadian rhythm layer, the sleep function layer, the cognitive-emotional layer, and the physical regulation layer. Establish the feature state vectors respectively And the mapping relationship between each functional level and cross-domain intervention features, wherein the cross-domain intervention features include one or more corresponding intervention units across the fields of nutrition, traditional Chinese medicine, and psychology; Obtain a stable baseline vector encompassing user personality traits, emotional susceptibility, and regulation ability. Using the stable baseline vector For the feature state vector The initial weights of each feature dimension are modulated to generate corrected weights. ; Combined with the feature state vector The stable baseline vector and historical feature state vector The historical state trajectory is represented by a dynamic Bayesian network to construct a state evolution equation, which describes the migration trajectory of sleep state over time, and introduces a personality modulation term to ensure that the inference process is deeply modulated by the individual's inherent psychological baseline. Introducing latent variables of time schedule It characterizes the degree to which an individual's state approaches the ideal health baseline and captures sleep recovery slope or deterioration warning signals in real time; The feature state vector With the stable baseline vector After concatenation, the data is input into a machine learning model to predict the risk probability, calculate the cosine similarity between the user state and the prototype vector, and match the image label. Based on the unified semantic space mapping relationship, the Monte Carlo tree search algorithm is used to simulate the implementation effect of different intervention schemes under different execution orders and intensities, select the optimal intervention path, and generate personalized intervention schemes. Automatically outputs a structured evaluation report containing a numerical layer, a type layer, an inference layer, and the intervention matching results; The optimal intervention path is executed, user feedback data after intervention is collected, and the data is sent back to the multimodal data collection step to drive dynamic iterative optimization of the entire process.

2. The sleep assessment and intervention method based on a hierarchical dynamic state-space model according to claim 1, characterized in that, The adaptive dynamic question generation program includes: The system iterates through the fusion question bank, calculates the expected information gain for unanswered fusion questions, performs real-time dynamic sorting, and prioritizes pushing the top-ranked question to the user. The risk label probability distribution is updated based on real-time user feedback, and the entropy value of the overall posterior distribution is reassessed. : ； in, Represents a risk label. The probability distribution representing the risk label; When entropy value Drop to preset threshold The following conditions will trigger the automatic termination logic.

3. The sleep assessment and intervention method based on a hierarchical dynamic state-space model according to claim 1, characterized in that, The original semantic state vector The recovery process includes: Establish multi-source heterogeneous vectors It includes multimodal information such as real-time scores, age, gender, sleep habits, past health records, physiological monitoring data, environmental perception data, and lifestyle data; Through formula It achieves semantic recovery and performs semantic completion on clinical information lost in interaction within a mathematical space.

4. The sleep assessment and intervention method based on a hierarchical dynamic state-space model according to claim 1, characterized in that, The intervention program includes one or more intervention methods. The Monte Carlo tree search algorithm is used to simulate the effects of different intervention schemes under different execution orders and intensities, and to select the optimal intervention path, including: The comprehensive sleep state, composed of the feature state vector, the stable baseline vector, and the risk probability, is used as the root node of the Monte Carlo tree. The intervention types corresponding to the 22-dimensional four-layer feature space, namely nutrition, traditional Chinese medicine, psychology, and behavior regulation, are taken as first-level nodes. The specific intervention methods under each intervention type are taken as second-level nodes. The execution order, intervention intensity, execution frequency, and duration are taken as leaf node attributes. The Monte Carlo tree search iteration is performed using the UCT upper confidence interval algorithm to sequentially complete path selection, node expansion, effect simulation, and value backtracking. The UCT rule is used to select the potential optimal branch, expand the insufficiently explored nodes, and perform multiple rounds of simulation with the improvement of 22-dimensional features as the optimization objective. The simulation results are then used to update the node value in reverse. A comprehensive scoring function is constructed, which combines intervention effect score, execution difficulty coefficient and user compliance index to calculate the total score of each path. The path with the highest total score is determined as the optimal intervention path, and a personalized intervention plan matching user characteristics is generated.

5. The sleep assessment and intervention method based on a hierarchical dynamic state-space model according to claim 1, characterized in that, The feature state vector includes 22 feature dimensions, and the sleep state assessment intervention method further includes: A unified 22-dimensional four-layer feature space is pre-constructed as a public quantitative coordinate system for sleep assessment and cross-domain intervention; The multi-source heterogeneous data, clinical scale items, and intervention program items are uniformly classified into the 22-dimensional four-layer feature space through semantic mapping and clustering algorithms; The representation content, fusion data types, and corresponding intervention areas of the four functional levels are as follows: The circadian rhythm layer comprises five dimensions to characterize: diurnal pattern tendency, rhythm consistency, phase synchronization, light regulation balance, and social jet lag; the fused data comes from corresponding work-rest records, light, temperature and humidity, and noise environmental perception data; the corresponding outputs include intervention suggestions for circadian rhythm, light environment, and work-rest adjustment. The sleep function layer includes eight dimensions to characterize: subjective sleep quality, sleep latency, sleep maintenance stability, sleep efficiency, sleep duration adequacy, daytime recovery level, sleep regulation ability, and sleepiness tendency; the fused data comes from physiological monitoring data of corresponding sleep cycles, heart rate variability, blood oxygen saturation, and body movement; the corresponding outputs include intervention suggestions for sleep behavior, physiological regulation, and deep sleep optimization. The cognitive-emotional layer comprises four dimensions to characterize: sleep belief bias, sense of sleep control, sleep-related anxiety, and emotion regulation ability; the integrated data comes from psychological scales, stress feedback, emotional state, and cognitive evaluation data; the corresponding outputs include cognitive adjustment, emotional guidance, psychological relaxation, and CBT-I type intervention suggestions. The constitution regulation layer includes five dimensions to characterize: innate constitution type, dynamic imbalance pattern, sleep constitution characteristics, dream integration ability, and physical and mental regulation ability; the integrated data comes from diet, exercise, lifestyle, TCM constitution, and physical and mental sensitivity data; the corresponding outputs include TCM conditioning, nutritional diet, exercise recovery, and constitution improvement intervention suggestions.

6. The sleep assessment and intervention method based on a hierarchical dynamic state-space model according to claim 1, characterized in that, The feature state vector includes 22 feature dimensions, and the corrected weights The calculation and calibration process includes: Initial weights Preset and allocation: Obtain the initial weight vectors corresponding one-to-one with the 22 feature dimensions. The initial weights are quantified and allocated according to the preset weight ratios of the functional levels to which the 22 feature dimensions belong. The preset weight ratios for each level are: 0.25 for the circadian rhythm layer, 0.35 for the sleep function layer, 0.20 for the cognitive emotion layer, and 0.20 for the physical regulation layer. Baseline feature extraction: Obtaining stable baseline vectors ,in Represents the user's personality traits. Represents emotional susceptibility. It represents self-regulation ability; Weighted dynamic modulation calculation: using the formula Generate corrected weights, where , , The modulation coefficients preset by the system; Subjective bias filtering: through the aforementioned corrected weights For the feature state vector The original feature values ​​of each dimension at the current moment are weighted to identify and filter extreme or moderate scoring biases caused by the user's specific answering tendencies, so that the final evaluation conclusion returns to the objective sleep function level.

7. The sleep assessment and intervention method based on a hierarchical dynamic state-space model according to claim 1, characterized in that, The state evolution equation is expressed as: in, The feature state vector at the current moment. This is the historical feature state vector from the previous moment; This is the vector of observations obtained at the current moment through the adaptive dynamic question generation program; This is the state transition matrix, used to describe the spontaneous evolution inertia of the sleep state without external disturbance; The observation contribution weight matrix is ​​used to determine the degree to which the current observation contributes to the state update; This is a personality modulation matrix used to ensure that the evolution of states is deeply modulated by the individual's inherent psychological baseline; The stable baseline vector; This is a random noise term used to absorb system measurement errors or sudden environmental disturbances.

8. The sleep assessment and intervention method based on a hierarchical dynamic state-space model according to claim 7, characterized in that, The time progress latent variable The calculation process includes: in, Based on the recovery rate; This is a weight vector used to incorporate individual baseline features. Correction to recovery speed; For risk feedback coefficient, This represents the probability of past risks, used to describe the hindering effect of risks on current progress. This represents random noise.

9. The sleep assessment and intervention method based on a hierarchical dynamic state-space model according to claim 1, characterized in that, The image tag matching mechanism includes: The feature state vector With the stable baseline vector Concatenate to construct an enhanced input vector; Obtain the preset prototype vector set , where each prototype vector The standardized state center vector is obtained in advance through feature extraction and cluster analysis of historical clinical sample data; Iterate through the set of prototype vectors and calculate the relationship between the enhanced input vector and each prototype vector. Cosine similarity between : Compare the cosine similarity The values ​​of the items are used to determine the highest similarity, and the item with the largest value is defined as the highest similarity. Based on the highest similarity, the user's current feature state vector Map to the preset discrete state space identifier with the highest similarity; The identifier is visually represented using imagery tags with semantic relevance, which are used to represent the specific pathological or psychological prototype clustering region in which the user is located in a 22-dimensional feature space.

10. A sleep assessment and intervention system based on a hierarchical dynamic state-space model, used to implement the sleep state assessment and intervention method based on a hierarchical dynamic state-space model as described in any one of claims 1-9, characterized in that, include: The data acquisition and semantic recovery module initiates an adaptive dynamic question generation program through an intelligent analysis terminal to collect multimodal data and calculate risk labels in real time to characterize the degree of sleep disorders in users. The posterior distribution entropy value is obtained, and the optimal questions are selected and pushed to the user. The multimodal data includes physiological monitoring data, environmental perception data, and lifestyle data. The user's real-time response feedback is obtained, and combined with basic demographic characteristics and past medical history, a high-dimensional multi-source heterogeneous vector is constructed. And using a pre-defined semantic mapping matrix The multi-source heterogeneous vectors are restored to their original semantic state vectors. ; Hierarchical feature building module, used to utilize feature mapping matrix The original semantic state vector Transform into feature state vector and the feature state vector It is divided into functional layers: time rhythm layer, sleep function layer, cognitive emotion layer and physical regulation layer; it is also used to establish the mapping relationship between the feature state vector and each functional layer and cross-domain intervention features, wherein the cross-domain intervention features include one or more corresponding intervention units from the fields of nutrition, traditional Chinese medicine and psychology. The baseline calibration module is used to acquire a stable baseline vector that encompasses user personality traits, emotional susceptibility, and regulatory capacity. Using the stable baseline vector to the feature state vector The initial weights of each feature dimension are modulated to generate corrected weights. ; The dynamic evolution modeling module, combined with the feature state vector The stable baseline vector and historical feature state vector The historical state trajectory is represented by a dynamic Bayesian network to construct a state evolution equation, which describes the migration trajectory of sleep state over time, and introduces a personality modulation term to ensure that the inference process is deeply modulated by the individual's inherent psychological baseline. The intervention path optimization module is used to simulate the implementation effects of different intervention schemes under different execution orders and intensities using the Monte Carlo tree search algorithm, to select the optimal intervention path, and generate personalized intervention schemes. The risk prediction and report generation module is used to generate the feature state vector. With the stable baseline vector After concatenation, the data is input into a machine learning model to predict the risk probability, calculate the cosine similarity between the user state and the prototype vector, and match the image label; receive the optimal intervention path and personalized intervention plan output by the intervention path optimization module, output a complete intervention plan including the intervention matching results; and automatically output a structured evaluation report including the numerical layer, type layer, inference layer, the intervention matching results, and the optimal intervention path.

11. An electronic device comprising at least one processor and a memory communicatively connected to said at least one processor, characterized in that, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method of any one of claims 1-9.

Images