Sleep-emotion-body health linkage risk assessment and intervention system and method

Abstract

The invention discloses a sleep-emotion-body health linkage risk assessment and intervention system and method, and relates to the technical field of medical informationization, and the method comprises the steps: obtaining multi-source heterogeneous data of sleep monitoring, emotion assessment and body health diagnosis dimensions, carrying out the standardization processing of the multi-source heterogeneous data of each dimension, and carrying out the analysis of the multi-source heterogeneous data of each dimension; obtaining a corresponding target data stream; extracting associated feature indexes according to the target data flow of each dimension, and generating a sleep-emotion-body health associated information triplex graph according to the associated feature indexes; constructing a risk assessment model based on the information triplet graph, the user basic information and the clinical knowledge graph, performing clinical risk judgment on the real-time multi-source heterogeneous data through the risk assessment model, outputting a clinical risk result, and generating specific warning content and intervention suggestions; the problem of intervention lag is effectively solved, the time sequence data processing efficiency is improved, meanwhile, only the abnormal association of clinical significance is reserved, and the risk identification accuracy is improved.

Description

Technical Field

[0001] This application relates to the field of medical information technology, specifically to a system and method for risk assessment and intervention linking sleep, mood, and physical health. Background Technology

[0002] Currently, the management of sleep, mood, and physical health problems in psychiatric patients typically employs several independent approaches: Using devices such as smart mattresses or wristbands, but the data is usually only displayed in the device's own app, reporting "sleep quality scores" or "low blood oxygenation events," lacking clinical interpretation. During outpatient visits or regular follow-ups, doctors or nurses conduct manual assessments using scales such as the Genomic Nutrition Scale (GDS), with the results recorded in electronic medical records. This is a static, point-based assessment. Nursing staff record the patient's vital signs and complaints of discomfort (such as pain or difficulty breathing) in nursing records, which are stored in a different system than sleep and mood data.

[0003] However, sleep data, mood scale data, and nursing record data are stored in different devices and information systems with inconsistent formats and lack of interconnectivity. This prevents the formation of a unified and holistic view of the patient's health status, limiting routine human health assessments to a single dimension. For example, a wristband may report "nighttime sleep disruption," but it cannot automatically correlate this with whether it is caused by sleep apnea (a physical problem) leading to daytime low mood (an emotional problem). It cannot achieve a deep cross-analysis of the strong correlation between the "sleep-emotion-physical" three dimensions, and it relies on manual discovery of abnormal correlations, which is inefficient and prone to omissions. Usually, the problem is only discovered when it has already significantly affected the patient's health, thus missing the golden period for early intervention.

[0004] The patent "A Big Data-Based Intelligent Diagnosis and Health Management System for Traditional Chinese Medicine," publication number CN120376162A, discloses a system comprising a multi-source data acquisition and integration unit, a big data analysis and mining unit, and an intelligent diagnosis and health management service unit. The multi-source data acquisition and integration unit's built-in functions include data source analysis and data integration. Data sources include clinical medical records, health monitoring data, ancient medical literature, and modern medical laboratory data. Data integration includes data cleaning, data standardization, and data fusion. The big data analysis and mining unit's built-in functions include disease-syndrome correlation analysis, prescription efficacy evaluation and optimization, personalized health management analysis, and drug development auxiliary analysis. The disease-syndrome correlation analysis involves feature extraction, machine learning modeling, and model evaluation and optimization. However, this solution primarily integrates multi-source data to construct a model for diagnosing patient conditions. Due to individual differences, the amount of data required is complex and massive. Improving data quality reduces data analysis efficiency, resulting in a model that cannot balance accuracy and efficiency, ultimately failing to achieve early intervention. Summary of the Invention

[0005] The purpose of this application is to address the problems of conventional medical information systems being unable to identify abnormal correlations and struggling to balance the accuracy and efficiency of multi-source data fusion analysis, leading to unreliable analysis results and delayed interventions. It proposes a sleep-emotion-physical health linkage risk assessment and intervention system and method. By acquiring multi-dimensional data on sleep, emotion, and physical health status, it integrates and correlates these data to analyze the risk of a user's physical health status and generates corresponding intervention suggestions. This achieves cross-dimensional abnormal correlation analysis, avoiding delayed interventions. Furthermore, by combining a clinical knowledge graph to construct a model, it improves the efficiency of time-series data processing while retaining only clinically significant abnormal correlations, thus enhancing the accuracy of risk identification.

[0006] To achieve the above objectives, the technical solutions adopted in the embodiments of this application are as follows: In a first aspect, embodiments of this application provide a method for risk assessment and intervention linking sleep, emotion, and physical health, the method comprising: Acquire multi-source heterogeneous data from sleep monitoring, emotion assessment, and physical health diagnosis dimensions; standardize the multi-source heterogeneous data from each dimension to obtain the corresponding target data stream. Based on the target data stream of each dimension, relevant feature indicators are extracted, and a tripartite diagram of the correlation between sleep, mood and physical health is generated based on the relevant feature indicators. A risk assessment model is constructed based on information tripartite graphs, user basic information, and clinical knowledge graphs. The model is used to assess clinical risks from real-time multi-source heterogeneous data, output clinical risk results, and generate specific warnings and intervention suggestions.

[0007] This solution integrates previously scattered and heterogeneous data into a unified health data view centered on the patient and oriented along a timeline by processing and fusing sleep, mood, and physical health data to generate a tripartite graph. This allows healthcare professionals to efficiently and comprehensively assess patient status from three strongly correlated dimensions. By constructing a risk identification model using correlation feature indicators and a clinical knowledge graph, it transforms the traditional, passive "problem-onset, problem-solving" model into a proactive monitoring, risk assessment, and early warning model. This enables timely warnings before or at the early stage of severe symptoms, providing a valuable "time window" for clinical intervention. It also addresses the problems of traditional methods relying on healthcare professionals' personal experience, which easily overlooks complex, multi-factor, non-linear, and time-lag-dependent correlations, leading to unreliable analysis results and delayed intervention. Furthermore, by analyzing time-series data from sleep, mood, and physical health, specific abnormal correlation patterns are identified. Combined with a clinical knowledge graph model, this accurately captures cross-time-period correlations and retains only clinically significant abnormal correlation features, improving the efficiency of parallel data fusion across dimensions while enhancing the accuracy of risk identification.

[0008] Preferably, the acquisition of multi-source heterogeneous data across sleep monitoring, emotion assessment, and physical health diagnosis dimensions, followed by standardization processing of the multi-source heterogeneous data for each dimension to obtain the corresponding target data stream, includes: Collect first and second health monitoring data of the user's sleep cycle, and calculate the target sleep monitoring data by fusing the first and second health monitoring data; Based on the user's electronic medical record, a state assessment scale is extracted, and the state assessment scale is parsed and anomaly annotation is performed to obtain structured target emotion assessment data. Based on user care information, key entities are identified using a trained medical named entity recognition model. The identified key entities are then standardized and mapped to symptom attribute relationships to form target physical health data.

[0009] Preferably, the step of collecting first and second health monitoring data of the user's sleep cycle, and then fusing the first and second health monitoring data to calculate the target sleep monitoring data, includes: The user's first health monitoring data is obtained based on the first monitoring device, including at least heart rate data, blood oxygen saturation data, and body movement data; Based on the second monitoring device, cardiac impact signals of specific human body areas of the user are obtained, and respiratory and heartbeat signals are obtained after signal preprocessing of the cardiac impact signals. By integrating heart rate data, blood oxygen saturation data, body movement data, respiratory signals, and heartbeat signals, heart rate indicators, respiratory rate indicators, and body movement indicators are calculated, and the sleep state type is determined based on the data of each indicator; The heart rate, respiratory rate, body movement, and sleep state types are encapsulated in a preset data frame format to obtain standardized target sleep monitoring data.

[0010] Preferably, the step of extracting a state assessment scale based on the user's electronic medical record, parsing and annotating the state assessment scale, and obtaining structured target emotion assessment data includes: Based on the first communication link, obtain the mental state assessment scale of the user in the electronic medical record system and / or psychological testing system; Based on the mental state assessment scale, extract key assessment fields including at least the assessment type, assessment name and its score, and assessment date; Outlier validation is performed on the key assessment fields, and outliers are marked as anomalies or trigger the re-collection of the mental state assessment scale. Based on the validated key assessment fields, a unified structured format is mapped to obtain the user's target emotion assessment data.

[0011] Preferably, the key assessment fields extracted based on the mental state assessment scale include at least the assessment type, assessment name and its score, and assessment date, and also include: If the mental state assessment scale is in unstructured text form, then the mental state assessment scale is semantically parsed based on natural language processing technology to obtain initial semantic information; Based on the initial semantic information, a similarity match is performed with the key evaluation information database to extract key evaluation fields.

[0012] Preferably, the step of identifying key entities based on user care information using a trained medical named entity recognition model, and then standardizing and mapping the identified key entities to symptom attribute relationships to form target physical health data includes: Nursing progress information in the hospital information system is obtained based on the second communication link. For structured nursing progress information, key entity data is extracted, including at least vital signs, medication records, laboratory test results, and nursing assessment items. For unstructured nursing progress information, a medical named entity recognition model is used to identify key entities, including at least symptom entities, sign entities, and treatment measure entities. After normalizing the identified symptom entities, sign entities, and treatment measure entities using standard medical terminology, symptom-attribute relationship pairs are constructed and entity data is mapped to obtain structured symptom attribute data. The key entity data and the symptom attribute structured data together constitute the target physical health data.

[0013] Preferably, the step of extracting correlation feature indicators based on the target data stream of each dimension, and generating a tripartite graph of information on the correlation between sleep, mood, and physical health based on the correlation feature indicators, includes: The target sleep monitoring data, the target emotion assessment data, and the target physical health data are spatiotemporally aligned. Based on spatiotemporal alignment, the relevant feature indicators of the target data streams in each dimension are extracted and visualized to generate a three-part information graph that combines sleep monitoring, emotion assessment, and vital sign monitoring.

[0014] Preferably, a risk assessment model is constructed based on an information tripartite graph, user basic information, and a clinical knowledge graph. This model assesses the clinical risk of real-time multi-source heterogeneous data, outputs clinical risk results, and generates specific warnings and intervention recommendations, including: Based on the aforementioned correlation feature indicators, temporal correlation features and clinically derived correlation features are obtained, and a clinical knowledge graph is constructed by combining the temporal correlation features and clinically derived correlation features with a clinical professional knowledge base. A basic risk identification model is trained based on the aforementioned correlation feature indicators and the aforementioned clinical knowledge graph. The basic risk identification model includes a temporal encoder, a knowledge-enhanced encoder, and a generative model. Based on the associated feature indicators labeled with risk and intervention system tags, combined with the clinical knowledge graph and user basic information, the temporal encoder, knowledge enhancement encoder and generative model are optimized and trained to obtain an optimized risk identification model; Based on the risk identification model, clinical risk assessment is performed on real-time multi-source heterogeneous data, and clinical risk results are output, along with specific warning content and intervention suggestions.

[0015] Preferably, the basic risk identification model trained based on the associated feature indicators and the clinical knowledge graph includes a temporal encoder, a knowledge-enhancing encoder, and a generative model, comprising: Each temporal correlation feature in the correlation feature index is mapped to the corresponding node of the clinical knowledge graph, and used as the first training data to train a temporal encoder for recognizing the correlation relationships of features in each dimension and aligning with clinical semantics. The triples based on the clinical knowledge graph, combined with the output of the temporal encoder, are used as the second training data to train a knowledge-enhancing encoder that integrates temporal features of each dimension with clinical knowledge rules. The triples based on the clinical knowledge graph are converted into natural language descriptions, and combined with medical text corpora corresponding to sleep, emotions, and physical conditions as third training data to train a generative model for risk assessment and intervention recommendations.

[0016] Secondly, embodiments of this application provide a sleep-emotion-physical health linked risk assessment and intervention system, including: The data monitoring module is used to acquire multi-source heterogeneous data from sleep monitoring, emotion assessment, and physical health diagnosis dimensions, and to standardize the multi-source heterogeneous data from each dimension to obtain the corresponding target data stream. The triptych generation module is used to extract correlation feature indicators from the target data stream of each dimension, and generate a triptych of information on the correlation between sleep, mood and physical health based on the correlation feature indicators. The risk identification module is used to build a risk assessment model based on the information tripartite graph, user basic information, and clinical knowledge graph. The risk assessment model is used to make clinical risk judgments on real-time multi-source heterogeneous data, output clinical risk results, and generate specific warning content and intervention suggestions.

[0017] The beneficial effects of this application are: 1. By processing and integrating data on sleep, mood, and physical health to generate a tripartite graph, the originally scattered and heterogeneous data is integrated into a unified health data view centered on the patient and with time as the axis, enabling medical staff to efficiently assess the patient's condition comprehensively and holistically from three strongly correlated dimensions. 2. By constructing a risk identification model through correlation feature indicators and clinical knowledge graphs, the traditional, passive "problem-once-it-occurs, problem-once" model is transformed into a proactive monitoring, risk assessment, and early warning model. This enables timely warnings before or at the early stage of severe symptoms, gaining a valuable "time window" for clinical intervention. It also solves the problem that traditional methods rely on the personal experience of medical staff, which easily overlooks complex correlations involving multiple factors, nonlinearity, and time lag, leading to unreliable analysis results and delayed intervention. This improves the targeting of intervention recommendations and shortens the intervention response time. 3. By analyzing time-series data from three dimensions—sleep, emotion, and physical condition—specific abnormal correlation patterns between them are identified. A risk identification model, which includes a time-series encoder, a knowledge-enhanced encoder, and a generative model, is constructed using a clinical knowledge graph. This model can accurately capture cross-time-period correlations and retain only clinically significant abnormal correlation features, thereby improving the efficiency of parallel fusion of cross-dimensional data while enhancing the accuracy of risk identification. Attached Figure Description

[0018] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0019] Figure 1 Flowchart of the sleep-emotion-physical health linkage risk assessment and intervention method provided in the embodiments of this application; Figure 2 A schematic diagram of a triptych of information provided in an embodiment of this application; Figure 3 This is a schematic diagram of a sleep-emotion-physical health linkage risk assessment and intervention system module provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely one preferred embodiment of this application and are only used to explain this application. They do not limit the scope of protection of this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] Example 1: As Figure 1 As shown, a method for risk assessment and intervention linking sleep, mood, and physical health includes steps S1-S3, wherein: S1. Acquire multi-source heterogeneous data from sleep monitoring, emotion assessment, and physical health diagnosis dimensions, standardize the multi-source heterogeneous data of each dimension, and obtain the corresponding target data stream.

[0022] As an optional implementation, step S1 includes: Collect first and second health monitoring data of the user's sleep cycle, and calculate the target sleep monitoring data by fusing the first and second health monitoring data; Based on the user's electronic medical record, a state assessment scale is extracted, and the state assessment scale is parsed and anomaly annotation is performed to obtain structured target emotion assessment data. Based on user care information, key entities are identified using a trained medical named entity recognition model. The identified key entities are then standardized and mapped to symptom attribute relationships to form target physical health data.

[0023] In this embodiment, by standardizing the data of each dimension, standardized, time-seriesd, and correlated basic data is provided for multi-source data fusion. By using a unified data format and time granularity, the originally isolated sleep, emotion, and physical data can be fused, avoiding the problem of low fusion efficiency caused by data heterogeneity in conventional systems.

[0024] As an optional implementation, the step of collecting first and second health monitoring data of the user's sleep cycle, and fusing the first and second health monitoring data to calculate the target sleep monitoring data, includes: The user's first health monitoring data is obtained based on the first monitoring device, including at least heart rate data, blood oxygen saturation data, and body movement data; Based on the second monitoring device, cardiac impact signals of specific human body areas of the user are obtained, and respiratory and heartbeat signals are obtained after signal preprocessing of the cardiac impact signals. By integrating heart rate data, blood oxygen saturation data, body movement data, respiratory signals, and heartbeat signals, heart rate indicators, respiratory rate indicators, and body movement indicators are calculated, and the sleep state type is determined based on the data of each indicator; The heart rate, respiratory rate, body movement, and sleep state types are encapsulated in a preset data frame format to obtain standardized target sleep monitoring data.

[0025] In some embodiments, the first monitoring device includes a smartwatch / band, and the first health monitoring data represents aggregated data collected by the smartwatch / band from a user over a certain period of time. This includes: heart rate data: including resting heart rate, average heart rate, heart rate variability, and heart rate curve data; blood oxygen saturation data: nighttime baseline blood oxygen saturation values, a list of nighttime blood oxygen saturation measurement records, and identified blood oxygen decline events; body movement data: time series of body movement frequency and intensity; and activity data: steps, calories burned, and activity duration.

[0026] The second monitoring device includes equipment for monitoring data from specific areas of the human body, such as a smart mattress. This involves deploying a high-frequency photoelectric sensor array in specific areas of the mattress (corresponding to the chest and neck areas). This sensor array is highly sensitive to minute body vibrations (i.e., cardiac impaction signals) caused by heartbeats, blood flow, and respiratory movements. The raw analog voltage signal output from the sensor array is amplified and filtered by a signal conditioning circuit, while suppressing low-frequency body turning noise and high-frequency environmental electronic noise. The signal is then converted into a digital signal by an analog-to-digital converter. In the controller, further digital filtering and signal processing are performed on the digital signal to separate signal components corresponding to different physiological activities, including respiratory and heartbeat signals. The respiratory signal is obtained by extracting ultra-low frequency signal components in the 0.1 Hz to 0.5 Hz frequency band, and the heartbeat signal is obtained by extracting mid-to-low frequency signal components in the 0.5 Hz to 20 Hz frequency band.

[0027] Specifically, heart rate data, blood oxygen saturation data, body movement data, respiratory signals, and heartbeat signals are integrated to calculate heart rate indicators, respiratory rate indicators, and body movement indicators. Based on these indicators, the sleep state type is determined, including: Peak detection is performed on the separated heartbeat signal, and combined with primary health monitoring data to identify continuous heartbeat intervals and calculate instantaneous and average heart rates to obtain heart rate indicators: The isolated respiratory signal is subjected to peak or trough detection, and the respiratory cycle and respiratory rate are calculated by combining the first health monitoring data to obtain the respiratory rate index; Body movement events are identified and body movement indicators are obtained by comparing the amplitude variance or energy abrupt changes between primary health monitoring data and raw cardiac impact signals or signals in specific frequency bands.

[0028] Furthermore, based on the user's body movement frequency, heart rate, respiratory rate, and their variability throughout the night, sleep states are categorized into accurate sleep onset, REM sleep, light sleep, and deep sleep. The calculated heart rate, respiratory rate, and body movement metrics are then used to determine the corresponding sleep state type within each time window. The heart rate, respiratory rate, and body movement metrics, along with the sleep state type, are then aggregated at preset time intervals to obtain the target sleep monitoring data.

[0029] In this embodiment, dual-source acquisition of sleep-related data simultaneously covers three major categories: peripheral physiology, core cardiopulmonary activity, and body movement behavior. This provides more comprehensive data support for sleep staging and apnea / hypopnea event identification, and enables uninterrupted monitoring throughout the night, improving data completeness and accuracy. Multi-source data fusion calculation effectively filters out measurement biases from single monitoring devices, further enhancing data reliability. The integrated data not only includes isolated physiological values ​​but also correlates with sleep states at corresponding time periods, enabling correlation analysis between physiological indicators and sleep states, providing a prerequisite for subsequent correlation analysis between sleep and physical abnormalities.

[0030] As an optional implementation, the step of extracting a state assessment scale based on the user's electronic medical record, parsing and annotating the state assessment scale, and obtaining structured target emotion assessment data includes: Based on the first communication link, obtain the mental state assessment scale of the user in the electronic medical record system and / or psychological testing system; Based on the mental state assessment scale, extract key assessment fields including at least the assessment type, assessment name and its score, and assessment date; Outlier validation is performed on the key assessment fields, and outliers are marked as anomalies or trigger the re-collection of the mental state assessment scale. Based on the validated key assessment fields, a unified structured format is mapped to obtain the user's target emotion assessment data.

[0031] As an optional implementation, the extraction of key assessment fields based on the mental state assessment scale includes at least the assessment type, assessment name and its score, and assessment date, and further includes: If the mental state assessment scale is in unstructured text form, then the mental state assessment scale is semantically parsed based on natural language processing technology to obtain initial semantic information; Based on the initial semantic information, a similarity match is performed with the key evaluation information database to extract key evaluation fields.

[0032] In some embodiments, the first communication link refers to a connection established with the hospital's electronic medical record system and dedicated psychological assessment system via a medical information exchange standard interface, used to access structured emotion assessment report data and unstructured assessment report text. Specifically, for structured emotion assessment report data, key assessment fields such as assessment scale name, total scale score, assessment date and time, and assessor are precisely extracted through database queries or API calls. For unstructured assessment report text, such as assessment results in nursing records or doctor's notes, natural language processing technology is used to parse it, first converting it into machine-readable initial semantic information, and then further identifying and extracting key entities such as assessment name, total score, and assessment date, thereby transforming unstructured text information into structured data.

[0033] Furthermore, the rationality of data extracted from different sources and in different formats is verified (for example, the total score of geriatric depression assessment should be between 0 and 30), outliers are marked or re-collection is triggered, the verified data is uniformly mapped to a standard format for storage, and each assessment record is attached with a timestamp accurate to the date and time to facilitate time-series correlation analysis with physical health data.

[0034] As an optional implementation, the step of identifying key entities based on user care information using a trained medical named entity recognition model, and then standardizing and mapping the identified key entities to symptom attribute relationships to form target physical health data includes: Nursing progress information in the hospital information system is obtained based on the second communication link. For structured nursing progress information, key entity data is extracted, including at least vital signs, medication records, laboratory test results, and nursing assessment items. For unstructured nursing progress information, a medical named entity recognition model is used to identify key entities, including at least symptom entities, sign entities, and treatment measure entities. After normalizing the identified symptom entities, sign entities, and treatment measure entities using standard medical terminology, symptom-attribute relationship pairs are constructed and entity data is mapped to obtain structured symptom attribute data. The key entity data and the symptom attribute structured data together constitute the target physical health data.

[0035] In this embodiment, the second communication link refers to the communication link that interacts with the nursing information system, vital signs monitoring system, laboratory information system, and medical order system through a medical information exchange standard interface; to query and extract nursing progress information, such as: vital signs: body temperature, blood pressure (systolic / diastolic), pulse, manually measured blood oxygen saturation, respiratory rate, with precise measurement timestamps attached; medication records, extracted from the medical order system, including: drug name, dosage, route of administration (oral, intravenous, etc.), administration time, and frequency; laboratory test results: key indicators such as complete blood count, electrolytes, and liver function; nursing assessment items: such as standardized scores like pain score (NRS) and fall risk assessment (Morse score).

[0036] In some embodiments, for free text in nursing records and physician progress notes, a medical named entity recognition model based on natural language processing technology is used for semantic parsing and extraction. For example, it identifies symptom entities (such as "chest pain," "dyspnea," "dizziness," "nausea," "paroxysmal nocturnal dyspnea"), sign entities (such as "bilateral lower extremity edema," "pulmonary rales"), and treatment entity entities (such as "oxygen therapy," "physical cooling") from the text. Synonymous but differently expressed symptoms are normalized to standard medical terms; for example, "shortness of breath" and "dyspnea" are normalized to "dyspnea," and associated with an industry-standard terminology set. Further extraction of symptom severity (such as "severe" in "severe headache"), body part (such as "left lower extremity" in "left lower extremity pain"), and time modifiers (such as "nighttime" in "worsening cough at night") is performed to construct "symptom-attribute" relationship pairs, forming structured data.

[0037] In some embodiments, all data from all sources, namely structured data, symptoms and signs processed by natural language processing, device data, etc., are aligned and fused using patient ID and timestamp as unique indexes; ultimately, each piece of physical health data includes patient ID, observation type (such as "symptoms", "vital signs", "medication"), specific content (such as "dyspnea", "blood pressure"), value / status, unit, and time of occurrence.

[0038] S2. Extract relevant feature indicators from the target data stream of each dimension, and generate a three-part information graph showing the relationship between sleep, mood and physical health based on the relevant feature indicators.

[0039] As an optional implementation, step S2 includes: The target sleep monitoring data, the target emotion assessment data, and the target physical health data are spatiotemporally aligned. Based on spatiotemporal alignment, the relevant feature indicators of the target data streams in each dimension are extracted and visualized to generate a three-part information graph that combines sleep monitoring, emotion assessment, and vital sign monitoring.

[0040] Specifically, the correlation characteristic indicators of the target data stream in each dimension should include at least: Sleep dimension characteristics: sleep efficiency (total sleep time / time in bed), percentage of deep sleep, sleep latency, number of nighttime awakenings, apnea-hypopnea index, and average nighttime blood oxygen saturation; Emotional dimension characteristics: total score of the Geriatric Depression Scale and its key sub-item scores, total score of the Generalized Anxiety Scale and its key sub-item scores. The scale scores were converted into a continuous numerical sequence. For non-continuous assessment points, interpolation or nearest neighbor methods were used for smoothing to form a time series. Somatic dimensional features: numerical, including average heart rate, respiratory rate, and pain score; event-based, transforming symptoms (such as "difficulty breathing" and "dizziness") and medication records (such as "taking painkillers") in the symptom attribute structured data into binary events that occur on the time axis (occurrence = 1, non-occurrence = 0).

[0041] In some examples, such as Figure 2 As shown, the associated characteristic indicators share a coordinate axis, constructing a parallel triptych: the horizontal axis represents time, and the upper region of the three parallel vertical axis regions plots the sleep stage curves, where: N1 represents the initial stage of transition from wakefulness to sleep, N2 represents the light sleep stage, N3 represents the deep sleep stage, and REM represents the rapid eye movement sleep stage, with shaded bar charts overlaid to show the intensity and distribution of apnea events and body movement events; the middle region plots the score change curves of mood scales such as GDS (Geriatric Depression) and GAD-7 (Generalized Anxiety Disorder) in the form of line charts or area charts; the lower region plots vital sign curves (such as heart rate and blood oxygen), and marks the occurrence of pain events, dyspnea symptoms, and medication time points in the form of event markers or bar charts.

[0042] In this embodiment, constructing a tripartite information graph helps to quickly locate the temporal linkage between sleep, mood, and vital signs. This allows for the direct transformation of explicit correlations into core features of the subsequent risk identification model, i.e., screening high-value features, reducing model computational complexity, decreasing data processing volume, and thus improving model training efficiency. Simultaneously, traditional raw data tables or feature matrices are difficult for clinical experts to understand, while shared-axis visualization charts can intuitively present data patterns. This allows experts to quickly point out that the interpolated mood score at a certain time period does not conform to clinical reality (e.g., the patient's anxiety score should have decreased rather than increased due to medication adjustments), thereby correcting data preprocessing biases.

[0043] S3. Based on the information triad graph, user basic information and clinical knowledge graph, a risk assessment model is constructed. The risk assessment model is used to judge the clinical risk of real-time multi-source heterogeneous data, output the clinical risk results and generate specific warning content and intervention suggestions.

[0044] As an optional implementation, step S3 includes: Based on the aforementioned correlation feature indicators, temporal correlation features and clinically derived correlation features are obtained, and a clinical knowledge graph is constructed by combining the temporal correlation features and clinically derived correlation features with a clinical professional knowledge base. A basic risk identification model is trained based on the aforementioned correlation feature indicators and the aforementioned clinical knowledge graph. The basic risk identification model includes a temporal encoder, a knowledge-enhanced encoder, and a generative model. Based on the associated feature indicators labeled with risk and intervention system tags, combined with the clinical knowledge graph and user basic information, the temporal encoder, knowledge enhancement encoder and generative model are optimized and trained to obtain an optimized risk identification model; Based on the risk identification model, clinical risk assessment is performed on real-time multi-source heterogeneous data, and clinical risk results are output, along with specific warning content and intervention suggestions.

[0045] Specifically, obtaining time-series correlation features and clinically derived correlation features based on the aforementioned correlation feature indicators includes: Data cleaning is performed for different data types. Specifically, outlier removal and missing value filling are performed on numerical data in the target sleep monitoring data and target physical health data. Logical verification and missing value processing are performed on the target emotion assessment data, including returning and supplementing outlier data, marking missing values ​​of single scales as invalid or filling them with historical averages. Deduplication is performed on event data in the target physical health data. Based on the cleaned data, basic features are extracted, including numerical features, time-series features, and clinically derived features. Clinically derived features characterize the interaction between the body, sleep, and emotions, as well as the corresponding risk status. Based on the correlation and influence of dimensional features and combined with clinical guidelines, the phenomenon types and feature thresholds of dimensions are formulated. For example, the phenomenon types of emotions are major depression and moderate depression, and the feature thresholds for major depression and moderate depression are a total score greater than 12 points and a total score greater than 8 points, respectively. Based on the basic characteristics, determine the temporal correlation and clinically derived correlation among multi-dimensional features, and then obtain the temporal correlation features and clinically derived correlation features. The temporal correlation features include temporal lag correlation features (such as the change of feature B after phenomenon a occurs in dimension A) and co-occurrence features (such as the simultaneous occurrence frequency of phenomenon a and phenomenon y in feature B). The clinically derived correlation features include risk stratification features (such as constructing a composite feature by combining the total score of patient age and emotional assessment type) and trend correlation features (such as the correlation between the slope of the decline in sleep efficiency and the slope of the increase in anxiety score).

[0046] Specifically, a clinical knowledge graph is constructed based on temporal correlation features and clinically derived correlation features, combined with a clinical professional knowledge base, including: A clinical professional knowledge base is constructed based on clinical guidelines, expert consensus, or medical terminology. By integrating the aforementioned clinical professional knowledge base, temporal correlation features, and clinical derivative correlation features, the graph relationships, relational attributes, and graph nodes are determined, and then a clinical knowledge graph is constructed based on the graph relationships, relational attributes, and graph nodes.

[0047] It should be noted that the clinical knowledge graph covers the entire chain of knowledge, including phenomena, attributes, relationships, risks, and interventions. It is used to transform scattered clinical rules, data association features, and medical terminology into a structured node-relationship-attribute graph, which ensures both clinical rationality and data adaptability, and provides interpretable, quantifiable, and iterative clinical knowledge support for risk identification models.

[0048] As an optional implementation, the basic risk identification model is trained based on the associated feature indicators and the clinical knowledge graph. The basic risk identification model includes a temporal encoder, a knowledge-enhancing encoder, and a generative model, comprising: Each temporal correlation feature in the correlation feature index is mapped to the corresponding node of the clinical knowledge graph, and used as the first training data to train a temporal encoder for recognizing the correlation relationships of features in each dimension and aligning with clinical semantics. The triples based on the clinical knowledge graph, combined with the output of the temporal encoder, are used as the second training data to train a knowledge-enhancing encoder that integrates temporal features of each dimension with clinical knowledge rules. The triples based on the clinical knowledge graph are converted into natural language descriptions, and combined with medical text corpora corresponding to sleep, emotions, and physical conditions as third training data to train a generative model for risk assessment and intervention recommendations.

[0049] In this embodiment, the temporal encoder is mainly used to transform hourly sleep, emotion, and physical raw features into vector representations with temporal dependencies, providing basic data features for subsequent knowledge fusion; the knowledge enhancement encoder is mainly used to fuse temporal features with clinical association rules in the knowledge graph, thereby solving the problem that pure data-driven models do not understand clinical logic, and providing fusion features of data patterns and clinical knowledge for the subsequent generative risk identification module; the generative model generates structured risk judgment results (including risk level information) and transforms the results into natural language warnings and intervention suggestions that doctors can directly use, realizing the effective implementation from data judgment to clinical application.

[0050] In some embodiments, the temporal encoder, knowledge-enhanced encoder, and generative model are optimized and trained based on the associated feature indicators labeled with risk and intervention system tags, combined with the clinical knowledge graph and user basic information, to obtain an optimized risk identification model, including: The hourly feature sequence output by the pre-trained temporal encoder is used as the basic temporal sequence, and the symptom attribute relationship pairs corresponding to the current user sample are extracted from the clinical knowledge graph to obtain knowledge association features. The knowledge association features are used as attention bias inputs to the attention layer of the knowledge augmentation encoder, and the basic temporal sequence is input to the knowledge augmentation encoder. The attention mechanism is used to achieve the fusion and optimization training of the temporal encoder and the knowledge augmentation encoder. The fusion feature vector output by the knowledge-enhanced encoder after optimization training is combined with the correlation feature index labeled with risk level and intervention suggestion as input to the generative model for optimization training, and the optimized generative model is obtained.

[0051] In some embodiments, the risk and intervention system labels include primary labels, secondary labels, and tertiary labels. Primary labels indicate whether the association pattern is valid, that is, whether any phenomenon in one dimension changes with any phenomenon in other dimensions. This is labeled using binary labels: 1 if yes, 0 otherwise. Secondary labels indicate the clinical risk level of the association pattern, including no risk, low risk, stroke, and high risk. Tertiary labels indicate the warning content and intervention suggestions for the corresponding risk, labeled using text. For example, high risk: low sleep efficiency accompanied by severe depression, which easily leads to nocturnal heart rate arrhythmia; intervention suggestions: adjust sleep medication and increase the frequency of psychological counseling.

[0052] Specifically, through optimized training, the encoder is ensured to output feature sequences with temporal correlation and basic clinical semantics. The feature sequences of the temporal encoder are used as input to the knowledge-enhanced encoder. At this time, the knowledge-enhanced encoder retrieves the embedding vectors and association rules of the corresponding phenomena in the knowledge graph, thereby upgrading the pure data features into fused features of data and knowledge, providing accurate judgment basis for risk identification of the generative model. After the fused features of the knowledge-enhanced encoder are input into the generative model, the structured instructions of the association judgment results and risk levels are first obtained, and then the instructions are input into the generative model together with the intervention rules in the knowledge graph to generate warnings and intervention suggestions.

[0053] In this embodiment, through model pre-training and optimization training, it is ensured that the final risk identification model fully learns the expression paradigm of clinical terminology and intervention recommendations, thereby accurately and reliably converting structured judgment results into natural language clinical recommendations. The three basic models work together to solve the problem of low efficiency in processing time-series data in conventional systems, and also make up for the business needs of lacking clinical knowledge and practical output. This provides risk identification with the ability to process efficiently, integrate accurately, and output practically, ultimately achieving reliable risk judgment while taking into account judgment efficiency, improving intervention response efficiency, and overcoming the problems of conventional systems that only output data reports and cannot synchronously and accurately provide intervention plans and have intervention lag.

[0054] In other possible embodiments, when the risk identification model identifies a novel association that has not yet been included in the preset rule base, it adds it to the optimized preset rule base to perform reverse optimization of the model. For example, the lag effect of "a significant decrease in daytime activity does not immediately affect mood, but will cause sleep disruption after 2 days" allows the model to be continuously optimized and updated, ensuring that the accuracy of the analysis and clinical relevance are continuously improved, and further continuously improving the model's risk identification accuracy and the reliability of intervention recommendations.

[0055] Example 2, as Figure 3 As shown in the embodiment of this application, a sleep-emotion-physical health linkage risk assessment and intervention system is provided, including: The data monitoring module is used to acquire multi-source heterogeneous data from sleep monitoring, emotion assessment, and physical health diagnosis dimensions, and to standardize the multi-source heterogeneous data from each dimension to obtain the corresponding target data stream. The triptych generation module is used to extract correlation feature indicators from the target data stream of each dimension, and generate a triptych of information on the correlation between sleep, mood and physical health based on the correlation feature indicators. The risk identification module is used to build a risk assessment model based on the information tripartite graph, user basic information, and clinical knowledge graph. The risk assessment model is used to make clinical risk judgments on real-time multi-source heterogeneous data, output clinical risk results, and generate specific warning content and intervention suggestions.

[0056] In this embodiment, by processing and fusing data from sleep, mood, and physical health to generate a tripartite graph, the originally scattered and heterogeneous data is integrated into a unified health data view centered on the patient and based on time. This allows medical staff to efficiently and comprehensively assess the patient's condition from three strongly correlated dimensions. By constructing a risk identification model using correlation feature indicators and a clinical knowledge graph, the traditional, passive "problem-once-it-occurs, then problem-solving" model is transformed into a proactive monitoring, risk assessment, and early warning model. This allows for timely warnings before or at the early stage of severe symptoms, providing a valuable "time window" for clinical intervention. It also addresses the problem of traditional methods relying on the personal experience of medical staff, which easily overlooks complex correlations involving multiple factors, nonlinearity, and time lags, leading to unreliable analysis results and delayed intervention. Furthermore, by analyzing time-series data from sleep, mood, and physical health, specific abnormal correlation patterns between them are identified. Combined with a model constructed using a clinical knowledge graph, this accurately captures cross-time-period correlations and retains only clinically significant abnormal correlation features, improving the efficiency of parallel fusion of cross-dimensional data while enhancing the accuracy of risk identification.

[0057] The above-described embodiments are preferred embodiments of this application and are not intended to limit the specific scope of this application. The scope of this application includes but is not limited to the specific embodiments described above. All equivalent changes made in accordance with the shape, structure, and method of this application are within the protection scope of this application.

Claims

1. A method for risk assessment and intervention linking sleep, mood, and physical health, characterized by: Includes the following steps: Acquire multi-source heterogeneous data from sleep monitoring, emotion assessment, and physical health diagnosis dimensions; standardize the multi-source heterogeneous data from each dimension to obtain the corresponding target data stream. Based on the target data stream of each dimension, relevant feature indicators are extracted, and a tripartite diagram of the correlation between sleep, mood and physical health is generated based on the relevant feature indicators. A risk assessment model is constructed based on information tripartite graphs, user basic information, and clinical knowledge graphs. The model is used to assess clinical risks from real-time multi-source heterogeneous data, output clinical risk results, and generate specific warnings and intervention suggestions.

2. The method for risk assessment and intervention linking sleep, emotion, and physical health according to claim 1, characterized in that: The process involves acquiring multi-source heterogeneous data across sleep monitoring, emotion assessment, and physical health diagnosis dimensions, standardizing the multi-source heterogeneous data for each dimension, and obtaining the corresponding target data stream, including: Collect first and second health monitoring data of the user's sleep cycle, and calculate the target sleep monitoring data by fusing the first and second health monitoring data; Based on the user's electronic medical record, a state assessment scale is extracted, and the state assessment scale is parsed and anomaly annotation is performed to obtain structured target emotion assessment data. Based on user care information, key entities are identified using a trained medical named entity recognition model. The identified key entities are then standardized and mapped to symptom attribute relationships to form target physical health data.

3. The method for risk assessment and intervention linking sleep, emotion, and physical health according to claim 2, characterized in that: The process of collecting first and second health monitoring data from the user's sleep cycle, and then fusing the first and second health monitoring data to calculate the target sleep monitoring data, includes: The user's first health monitoring data is obtained based on the first monitoring device, including at least heart rate data, blood oxygen saturation data, and body movement data; Based on the second monitoring device, cardiac impact signals of specific human body areas of the user are obtained, and respiratory and heartbeat signals are obtained after signal preprocessing of the cardiac impact signals. By integrating heart rate data, blood oxygen saturation data, body movement data, respiratory signals, and heartbeat signals, heart rate indicators, respiratory rate indicators, and body movement indicators are calculated, and the sleep state type is determined based on the data of each indicator; The heart rate, respiratory rate, body movement, and sleep state types are encapsulated in a preset data frame format to obtain standardized target sleep monitoring data.

4. The method for risk assessment and intervention linking sleep, emotion, and physical health according to claim 2, characterized in that: The process involves extracting a state assessment scale from the user's electronic medical record, parsing and annotating the state assessment scale, and obtaining structured target emotion assessment data, including: Based on the first communication link, obtain the mental state assessment scale of the user in the electronic medical record system and / or psychological testing system; Based on the mental state assessment scale, extract key assessment fields including at least the assessment type, assessment name and its score, and assessment date; Outlier validation is performed on the key assessment fields, and outliers are marked as anomalies or trigger the re-collection of the mental state assessment scale. Based on the validated key assessment fields, a unified structured format is mapped to obtain the user's target emotion assessment data.

5. The method for risk assessment and intervention linking sleep, emotion, and physical health according to claim 4, characterized in that: The extraction of key assessment fields based on the mental state assessment scale includes at least the assessment type, assessment name and its score, and assessment date, and also includes: If the mental state assessment scale is in unstructured text form, then the mental state assessment scale is semantically parsed based on natural language processing technology to obtain initial semantic information; Based on the initial semantic information, a similarity match is performed with the key evaluation information database to extract key evaluation fields.

6. The method for risk assessment and intervention linking sleep, emotion, and physical health according to claim 2, characterized in that: The process involves identifying key entities based on user care information using a trained medical named entity recognition model, standardizing the identified key entities, and mapping symptom attribute relationships to form target physical health data, including: Nursing progress information in the hospital information system is obtained based on the second communication link. For structured nursing progress information, key entity data is extracted, including at least vital signs, medication records, laboratory test results, and nursing assessment items. For unstructured nursing progress information, a medical named entity recognition model is used to identify key entities, including at least symptom entities, sign entities, and treatment measure entities. After normalizing the identified symptom entities, sign entities, and treatment measure entities using standard medical terminology, symptom-attribute relationship pairs are constructed and entity data is mapped to obtain structured symptom attribute data. The key entity data and the symptom attribute structured data together constitute the target physical health data.

7. The method for risk assessment and intervention linking sleep, emotion, and physical health according to claim 2, characterized in that: The step of extracting correlation feature indicators from the target data stream of each dimension and generating a tripartite graph of the correlation between sleep, mood, and physical health based on the correlation feature indicators includes: The target sleep monitoring data, the target emotion assessment data, and the target physical health data are spatiotemporally aligned. Based on spatiotemporal alignment, the relevant feature indicators of the target data streams in each dimension are extracted and visualized to generate a three-part information graph that combines sleep monitoring, emotion assessment, and vital sign monitoring.

8. The method for risk assessment and intervention linking sleep, emotion, and physical health according to claim 7, characterized in that: A risk assessment model is constructed based on a tripartite graph of information, basic user information, and a clinical knowledge graph. This model assesses clinical risk from real-time, multi-source, heterogeneous data, outputs clinical risk results, and generates specific warnings and intervention recommendations, including: Based on the aforementioned correlation feature indicators, temporal correlation features and clinically derived correlation features are obtained, and a clinical knowledge graph is constructed by combining the temporal correlation features and clinically derived correlation features with a clinical professional knowledge base. A basic risk identification model is trained based on the aforementioned correlation feature indicators and the aforementioned clinical knowledge graph. The basic risk identification model includes a temporal encoder, a knowledge-enhanced encoder, and a generative model. Based on the associated feature indicators labeled with risk and intervention system tags, combined with the clinical knowledge graph and user basic information, the temporal encoder, knowledge enhancement encoder and generative model are optimized and trained to obtain an optimized risk identification model; Based on the risk identification model, clinical risk assessment is performed on real-time multi-source heterogeneous data, and clinical risk results are output, along with specific warning content and intervention suggestions.

9. The method for risk assessment and intervention linking sleep, emotion, and physical health according to claim 8, characterized in that: The basic risk identification model trained based on the associated feature indicators and the clinical knowledge graph includes a temporal encoder, a knowledge-enhancing encoder, and a generative model, comprising: Each temporal correlation feature in the correlation feature index is mapped to the corresponding node of the clinical knowledge graph, and used as the first training data to train a temporal encoder for recognizing the correlation relationships of features in each dimension and aligning with clinical semantics. The triples based on the clinical knowledge graph, combined with the output of the temporal encoder, are used as the second training data to train a knowledge-enhancing encoder that integrates temporal features of each dimension with clinical knowledge rules. The triples based on the clinical knowledge graph are converted into natural language descriptions, and combined with medical text corpora corresponding to sleep, emotions, and physical conditions as third training data to train a generative model for risk assessment and intervention recommendations.

10. A sleep-emotion-physical health linkage risk assessment and intervention system, applicable to the sleep-emotion-physical health linkage risk assessment and intervention method as described in any one of claims 1-9, characterized in that: include: The data monitoring module is used to acquire multi-source heterogeneous data from sleep monitoring, emotion assessment, and physical health diagnosis dimensions, and to standardize the multi-source heterogeneous data from each dimension to obtain the corresponding target data stream. The triptych generation module is used to extract correlation feature indicators from the target data stream of each dimension, and generate a triptych of information on the correlation between sleep, mood and physical health based on the correlation feature indicators. The risk identification module is used to build a risk assessment model based on the information tripartite graph, user basic information, and clinical knowledge graph. The risk assessment model is used to make clinical risk judgments on real-time multi-source heterogeneous data, output clinical risk results, and generate specific warning content and intervention suggestions.

Images