A user physiological state prediction method and system based on multi-source health data

Abstract

This application proposes a method and system for predicting user physiological states based on multi-source health data, belonging to the field of health data processing technology. It acquires heterogeneous health data from multiple sources, semantically aligns time-series numerical physiological data with textual health records, and fuses them across modalities to generate a structured health data sequence. Multi-scale deep feature extraction is performed on this sequence, and a causal correlation graph is constructed by combining textual semantic correction of node causal strength, resulting in a comprehensive feature matrix. Long-term behavioral feature sequences are extracted, and the behavioral baseline drift of resting heart rate is quantified based on causal paths. Context-weighted correction is performed through textual semantics, and the drift is removed to obtain accurate pathological offsets, thereby identifying abnormal events and constructing an anomaly assessment baseline. Resting heart rate regression prediction and physiological anomaly warning are achieved in parallel through a multi-task sequence encoding network. By fusing multi-modal data and causal inference, physiological interference is effectively removed, improving the accuracy of user physiological state prediction and anomaly warning.

Description

Technical Field

[0001] This application relates to the field of health data processing technology, and in particular to a method and system for predicting user physiological state based on multi-source health data. Background Technology

[0002] With the widespread adoption of wearable devices, users can collect physiological and behavioral data such as resting heart rate, steps taken, and sleep duration over a long period of time, providing a data foundation for daily monitoring and early warning of health status.

[0003] In practical applications, a prominent problem that troubles users is the frequent false alarms. For example, when a user switches from a sedentary lifestyle to regular running, their resting heart rate gradually decreases from 75 beats per minute to around 60 beats per minute, which is a normal physiological adaptation due to improved cardiopulmonary function. However, one day, due to early signs of a cold, the user's resting heart rate rises again from 60 beats per minute to 68 beats per minute. For existing health monitoring systems, a value of 68 beats per minute is significantly deviated from the recent baseline, often triggering an abnormal warning. In reality, this may simply be a temporary fluctuation caused by a cold, requiring no emergency intervention.

[0004] The root cause of these false alarms lies in the fact that current technologies typically use historical averages within a fixed time window as a health baseline, and use the deviation between the current observation and the historical average as the basis for anomaly detection. This method cannot distinguish between normal physiological baseline drift caused by long-term behavioral pattern changes and pathological shifts caused by disease factors. When a user begins regular exercise, their historical average slowly declines due to the incorporation of data from the exercise adaptation period. Once the resting heart rate rises due to slight discomfort, this rise is considered abnormal, generating a false alarm.

[0005] Therefore, the key issue in improving the accuracy of health monitoring system warnings is how to accurately isolate normal baseline drift caused by long-term behavioral pattern changes from the daily fluctuations of physiological indicators and only issue warnings for true pathological deviations. Summary of the Invention

[0006] This application provides a method and system for predicting user physiological state based on multi-source health data, in order to overcome the shortcomings of the prior art.

[0007] In a first aspect, this application provides a method for predicting a user's physiological state based on multi-source health data, comprising the following steps: Acquire users' multi-source heterogeneous health data, perform semantic alignment and cross-modal fusion on time-series numerical physiological data and text health records, and generate semantically enhanced structured health data sequences; Multi-scale deep feature extraction is performed on the structured health data sequence, and a causal relationship graph with physiological indicators as nodes is constructed simultaneously. Based on the causal relationship graph, the spatiotemporal features of the nodes are propagated and aggregated through a graph neural network to obtain a comprehensive feature matrix with enhanced causal structure. The long-term behavioral feature sequence within a preset time window is extracted from the comprehensive feature matrix. The causal path from the behavioral node to the resting heart rate node in the causal relationship graph is used to quantify the drift of the long-term behavioral pattern on the resting heart rate baseline. The semantic features of the text health record are used to perform context-weighted correction on the drift. From the total fluctuation between the current observed value of resting heart rate and the historical baseline value, the corrected drift amount is removed to obtain the pathological deviation amount after removing the behavioral baseline drift; when the pathological deviation amount exceeds a preset threshold, a pathological deviation event is determined to have occurred, and an abnormal assessment baseline is constructed based on the cumulative degree of the pathological deviation amount. The comprehensive feature matrix and the anomaly assessment baseline are input into a multi-task sequence coding network. The shared encoder extracts the temporal representation, and a dedicated task branch performs future resting heart rate regression prediction and abnormal event warning in parallel based on the deviation between the temporal representation and the anomaly assessment baseline, outputting the physiological state prediction result.

[0008] According to the technical solution provided in this application, the step of acquiring multi-source heterogeneous health data of users, performing semantic alignment and cross-modal fusion on the time-series numerical physiological data and text health records to generate a semantically enhanced structured health data sequence includes the following steps: The system collects time-series numerical physiological data, including heart rate, blood oxygen saturation, steps taken, and accelerometer signals, using wearable devices. It identifies resting time periods based on the accelerometer signals and extracts the corresponding heart rate values ​​within those periods as resting heart rate samples. Activity intensity data is calculated from the accelerometer signals. Sleep duration data is collected using the wearable device's built-in sleep monitoring module. Finally, a health management application on a mobile terminal collects symptom descriptions and emotion tags input by the user based on preset question templates, which are used as the text health record. The text health record is input into a pre-trained language model to obtain a text semantic embedding vector; the time-series numerical physiological data is encoded through a one-dimensional convolutional neural network to obtain a numerical feature embedding vector. The text semantic embedding vector is aligned with the numerical feature embedding vector based on the timestamp, and an attention fusion layer is used to generate a semantically enhanced structured health data sequence.

[0009] According to the technical solution provided in this application, the step of performing multi-scale deep feature extraction on the structured health data sequence and simultaneously constructing a causal relationship graph with physiological indicators as nodes; based on the causal relationship graph, obtaining a comprehensive feature matrix with enhanced causal structure by propagating and aggregating the spatiotemporal features of the nodes through a graph neural network, includes the following steps: A temporal convolutional network containing large, medium and small convolutional kernels is used to extract local features of the structured health data sequence at multiple time scales in parallel, and then spliced ​​to obtain multi-scale deep features; Using heart rate, blood oxygen saturation, exercise steps, activity intensity, and sleep duration as nodes, a direct linear non-Gaussian causal discovery algorithm is used to learn the causal direction and initial strength between nodes from historical data. The text semantic embedding vector corresponding to the text health record is input into the causal strength adjustment network, and the correction coefficients of each causal edge are output. The initial strength is weighted and adjusted to obtain the corrected causal strength, and the causal relationship graph is constructed. The causal relationship graph and the multi-scale deep features are input into a graph attention network. The graph attention layer propagates node features along the causal edges and aggregates neighbor node information, outputting the spatiotemporal representation of each node to form the comprehensive feature matrix.

[0010] According to the technical solution provided in this application, the step of extracting long-term behavioral feature sequences within a preset time window from the comprehensive feature matrix, quantifying the drift of long-term behavioral patterns on the resting heart rate baseline using the causal path from behavioral nodes to resting heart rate nodes in the causal relationship graph, and using the semantic features of the text health record to perform context-weighted correction on the drift includes the following steps: Using the number of steps and activity intensity as behavioral nodes, the daily average number of steps and daily average activity intensity for multiple consecutive days are extracted from the comprehensive feature matrix to form the long-term behavioral feature sequence; In the causal relationship graph, the causal path from the behavior node to the resting heart rate node is extracted; the long-term behavior feature sequence is input into a pre-trained causal effect prediction network, which takes the change in behavior feature as input and the change in the resting heart rate baseline as output, and obtains the initial drift amount through forward propagation; The text semantic embedding vector corresponding to the text health record is input into a gating network, which is a single-layer fully connected network followed by a sigmoid activation function to generate weighting coefficients corresponding to each time point of the initial drift. The initial drift is multiplied element by element by the weighting coefficients to obtain the corrected drift.

[0011] According to the technical solution provided in this application, the step of removing the drift from the total fluctuation between the current observed value of resting heart rate and the historical baseline value to obtain the pathological offset after removing behavioral baseline drift includes the following steps: The difference between the observed resting heart rate at the current time point and the historical baseline value is calculated as the total fluctuation; the historical baseline value is the average resting heart rate under similar behavioral patterns during the same period within a preset number of days in the past. Subtract the drift from the total fluctuation to obtain the residual value. Take the absolute value of the residual value as the pathological offset and determine the offset direction based on the sign of the residual value.

[0012] According to the technical solution provided in this application, the step of determining that a pathological deviation event has occurred when the pathological deviation exceeds a preset threshold, and constructing an abnormality assessment baseline based on the cumulative degree of the pathological deviation, includes the following steps: A pathological offset threshold is set, which is adaptively updated based on the preset upper quantile of the user's historical pathological offset sequence; when the pathological offset at multiple consecutive sampling time points exceeds the threshold, a pathological offset event is determined to have occurred. The pathological offset values ​​at each sampling time point during the pathological offset event are weighted and summed to obtain the cumulative offset value, which is then used as the anomaly assessment baseline.

[0013] According to the technical solution provided in this application, the step of inputting the comprehensive feature matrix and the anomaly assessment baseline into a multi-task sequence coding network, extracting temporal representations by a shared encoder, and having a dedicated task branch perform parallel future resting heart rate regression prediction and abnormal event warning based on the deviation between the temporal representations and the anomaly assessment baseline, and outputting physiological state prediction results, includes the following steps: The shared encoder uses a bidirectional long short-term memory network to encode the comprehensive feature matrix, thereby obtaining a shared temporal representation vector; The first task branch concatenates the shared temporal representation vector with the anomaly assessment baseline and inputs it into the fully connected regression layer, outputting the predicted resting heart rate value at a future preset time point; The second task branch calculates the Euclidean distance between the shared temporal representation vector and the anomaly assessment baseline. First, the anomaly assessment baseline is copied and expanded to the same dimension as the shared temporal representation vector, and both are L2 normalized. Then, the Euclidean distance between the normalized vectors is calculated. The dynamic threshold is adaptively updated based on the preset lower quantile of the user's historical Euclidean distance sequence. When the Euclidean distance is lower than the dynamic threshold, an anomaly warning signal is generated. The third task branch inputs the shared temporal representation vector into the fully connected regression layer and outputs the predicted blood oxygen saturation value at a future preset time point. The predicted resting heart rate, the predicted blood oxygen saturation, and the abnormal warning signal are collectively output as the physiological state prediction result.

[0014] According to the technical solution provided in this application, the determination of the historical baseline value includes the following steps: The long-term behavioral feature sequence is subjected to trend detection to determine whether the user has entered a behavioral pattern transition period. The trend detection adopts the Mann-Kendall trend test. When the absolute value of the test statistic exceeds the critical value corresponding to the preset significance level, it is determined that the user has entered a behavioral pattern transition period. When the duration of the behavioral pattern transition exceeds a preset physiological adaptation threshold, the user is determined to be in a physiological adaptation state; the physiological adaptation threshold is set to 21 days based on population statistics. Under the physiological adaptation state, the historical baseline value is frozen, and the steady-state baseline before the behavioral pattern transition period is used as the historical baseline value. After the behavioral pattern transition period ends, when the coefficient of variation of the resting heart rate is lower than the preset coefficient of variation threshold within a 7-day sliding window, it is determined that the variability of the resting heart rate has returned to steady state conditions, the freeze is lifted, and the historical baseline value is recalculated based on the new steady state data after the behavioral pattern transition period.

[0015] According to the technical solution provided in this application, the output of the causal effect prediction network further undergoes the following verification and correction steps: When a pathological offset event is determined to have occurred based on the pathological offset amount, the behavioral feature sequence and resting heart rate observation sequence within the preset backtracking window before the event occurs, as well as the semantic features of the text health record for the corresponding time period, are extracted. Using the starting time point of the backtracking window as the intervention point, and utilizing the causal relationship diagram and structural causal model, a counterfactual resting heart rate trajectory is generated under the assumption that the behavioral pattern is maintained before the intervention point. Calculate the cumulative deviation between the resting heart rate observation sequence and the counterfactual resting heart rate trajectory, and compare the cumulative deviation with the drift amount to obtain the drift amount estimation residual; In response to the drift estimation residual exceeding the preset tolerance, an instant correction sample is constructed using the change in behavioral features within the backtracking window and the cumulative deviation, and the parameters of the causal effect prediction network are updated step by step. The updated causal effect prediction network parameters are synchronized to the drift quantization step for drift estimation at subsequent time points.

[0016] Secondly, this application proposes a user physiological state prediction system based on multi-source health data, used to execute the user physiological state prediction method based on multi-source health data as described above, including: The acquisition module is configured to acquire users' multi-source heterogeneous health data, perform semantic alignment and cross-modal fusion on the time-series numerical physiological data and text health records, and generate semantically enhanced structured health data sequences. The matrix enhancement module is configured to perform multi-scale deep feature extraction on the structured health data sequence and simultaneously construct a causal relationship graph with physiological indicators as nodes; based on the causal relationship graph, the spatiotemporal features of the nodes are propagated and aggregated through a graph neural network to obtain a comprehensive feature matrix with enhanced causal structure. The extraction module is configured to extract long-term behavioral feature sequences within a preset time window from the comprehensive feature matrix, quantify the drift of long-term behavioral patterns on the resting heart rate baseline using the causal path from behavioral nodes to resting heart rate nodes in the causal relationship graph, and apply context-weighted correction to the drift using the semantic features of the text health records. The baseline construction module is configured to remove the corrected drift from the total fluctuation between the current observed value of resting heart rate and the historical baseline value to obtain the pathological offset after removing behavioral baseline drift; when the pathological offset exceeds a preset threshold, a pathological offset event is determined to have occurred, and an abnormality assessment baseline is constructed based on the cumulative degree of the pathological offset. The prediction module is configured to input the comprehensive feature matrix and the anomaly assessment baseline into a multi-task sequence coding network, where a shared encoder extracts temporal representations, and a dedicated task branch performs future resting heart rate regression prediction and abnormal event warning in parallel based on the degree of deviation between the temporal representations and the anomaly assessment baseline, and outputs physiological state prediction results.

[0017] Compared with the prior art, the beneficial effects of this application are as follows: I. Effectively reduce false alarm rates and improve the specificity of anomaly detection. By constructing a causal relationship graph with physiological indicators as nodes and quantifying the drift of long-term behavioral patterns on the resting heart rate baseline from behavioral nodes, the causal decoupling between normal physiological baseline drift caused by behavioral changes and pathological shifts caused by disease factors is achieved. After removing the drift from the total fluctuation, the resulting pathological shift can truly reflect abnormal changes in the user's internal health status, significantly reducing false alarms triggered by behavioral changes such as users starting regular exercise or adjusting their sleep schedules, thus improving the accuracy of health monitoring and user trust.

[0018] II. Multimodal semantic information assists in drift correction, improving decoupling accuracy. In addition to using time-series numerical data to model the causal relationship between behavior and physiological indicators, this approach also incorporates textual semantic features such as symptoms and emotions actively recorded by users to perform context-weighted correction of drift. When users complain of subjective feelings such as fatigue or low mood, textual semantics can help determine the nature of the current physiological fluctuations, making the drift estimation more consistent with the real situation and further reducing the risk of misjudging pathological shifts.

[0019] III. Cross-modal fusion and causal graph neural network collaboration enhance feature representation and prediction accuracy. By temporally aligning and attentionally fusing textual semantic embedding with numerical features, a semantically enhanced structured data sequence is generated. This sequence is then combined with a causal graph neural network to aggregate spatiotemporal features. The resulting comprehensive feature matrix contains multi-scale temporal information and embeds the causal structure between indicators, providing a more discriminative feature representation for subsequent multi-task prediction and effectively improving the accuracy of future resting heart rate regression prediction.

[0020] IV. Continuous Accumulation of Anomaly Assessment Baseline Enables Multi-Task Joint Early Warning. An anomaly assessment baseline is dynamically constructed based on the stripped pathological shifts and fed as an auxiliary input into a multi-task sequence coding network, simultaneously driving two tasks: future physiological indicator regression prediction and anomaly event early warning. The anomaly assessment baseline is continuously updated according to the accumulation of pathological shift events, enabling the early warning branch to make judgments based on a comprehensive understanding of historical pathological accumulation status. This approach is more robust than single-threshold judgments, further reducing the probability of missed and false alarms. Attached Figure Description

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

[0022] Figure 1 A flowchart illustrating a user physiological state prediction method based on multi-source health data provided in this application embodiment; Figure 2 This is a schematic diagram of a user physiological state prediction system based on multi-source health data, provided in an embodiment of this application.

[0023] The diagram is labeled as follows: 1. Acquisition module; 2. Matrix enhancement module; 3. Extraction module; 4. Baseline construction module; 5. Prediction module. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. In the description of this application, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0025] Example 1 Firstly, this application provides a method for predicting a user's physiological state based on multi-source health data, such as... Figure 1 As shown, it includes the following steps: S1. Obtain multi-source heterogeneous health data from users, perform semantic alignment and cross-modal fusion on time-series numerical physiological data and text health records, and generate a semantically enhanced structured health data sequence. S2. Perform multi-scale deep feature extraction on the structured health data sequence and simultaneously construct a causal relationship graph with physiological indicators as nodes; based on the causal relationship graph, propagate and aggregate the spatiotemporal features of the nodes through a graph neural network to obtain a comprehensive feature matrix with enhanced causal structure. S3. Extract long-term behavioral feature sequences within a preset time window from the comprehensive feature matrix, and quantify the drift of long-term behavioral patterns on the resting heart rate baseline using the causal path from behavioral nodes to resting heart rate nodes in the causal relationship graph; use the semantic features of the text health record to perform context-weighted correction on the drift. S4. From the total fluctuation between the current observed value of resting heart rate and the historical baseline value, remove the corrected drift amount to obtain the pathological deviation amount after removing the behavioral baseline drift; when the pathological deviation amount exceeds a preset threshold, determine that a pathological deviation event has occurred, and construct an abnormal assessment baseline based on the cumulative degree of the pathological deviation amount. S5. Input the comprehensive feature matrix and the anomaly assessment baseline into a multi-task sequence coding network. The shared encoder extracts the temporal representation, and the dedicated task branch performs future resting heart rate regression prediction and abnormal event warning in parallel based on the deviation between the temporal representation and the anomaly assessment baseline, and outputs the physiological state prediction result.

[0026] Specifically, multi-source heterogeneous health data comprises two parts. The first part is time-series numerical physiological data continuously recorded by smart bracelets or watches. Specific fields include heart rate and blood oxygen saturation obtained from photoplethysmography (PPG) sensors, triaxial acceleration signals recorded by accelerometers, and step counts obtained from the acceleration signals. The second part is text-based health records actively filled out by users through health management applications on their mobile phones. Each record includes answers to several standard questions, such as whether symptoms like dizziness, chest tightness, or fatigue occurred that day, and a mood tag selected from a preset list for the day, such as calm, anxious, or tired. The records are timestamped to the minute. These two types of data are heterogeneous in the sense of being multi-source and multi-format. Processing them first requires bridging the modal gap between them to achieve semantic alignment and cross-modal fusion. The specific approach to semantic alignment involves feeding each health record into a finely tuned bidirectional language model, such as BioBERT, on a medical corpus. The model's last layer extracts the output vector corresponding to the specific markers of the entire sentence, resulting in a fixed-length text semantic embedding vector. Simultaneously, numerical sequences such as heart rate, blood oxygen, steps, and activity intensity are resampled every minute and fed into a one-dimensional convolutional neural network consisting of three layers of convolutions and max pooling. The network slides the convolution kernels along the time axis, automatically capturing local fluctuation patterns, and ultimately outputting a numerical feature embedding vector with the same dimension as the text embedding. Cross-modal fusion pairs numerical and text vectors within the same time period based on timestamps. Specifically, it uses a day as the basic alignment unit, pairing the average of all numerical vectors for the day with the text embedding for that day. During fusion, a multi-head attention mechanism is used, treating the numerical vector as the query and the text vector as both key and value. The contribution weight of text semantics to each dimension of the numerical vector is calculated, and residual connections are used to inject weighted text context information into the numerical vector. Finally, layer normalization is applied to obtain a semantically enhanced structured health data sequence. Each time step in the sequence not only carries the original physiological measurements, but also explicitly encodes higher-order semantics such as the user's subjective symptoms and emotional state.

[0027] For the above sequences, the next step is to perform multi-scale deep feature extraction. A temporal convolutional network with three parallel branches is constructed, with kernel lengths of 3, 5, and 7 respectively, corresponding to capturing short-term instantaneous changes, physiological responses within tens of minutes, and long-term rhythms with a daily cycle. For longer kernels, dilated convolution is used to expand the receptive field without increasing the number of parameters. Each branch consists of two convolutional layers, batch normalization, and a GELU activation function. Ultimately, three feature maps at different scales are obtained at each time step, which are concatenated along the channel dimension to form multi-scale deep features, allowing for the simultaneous preservation of fine-grained beat-by-beat heart rate changes and coarse-grained circadian rhythm information. Simultaneously with feature extraction, a causal relationship graph is constructed with physiological indicators as nodes. The node set includes five items: resting heart rate, blood oxygen saturation, steps taken, activity intensity, and sleep duration. Resting heart rate was taken from the median heart rate of the wearer during resting periods. Activity intensity was obtained by integrating the sum of accelerations per minute from the accelerometer signal. Sleep duration was the daily sum output by the wearable device's sleep monitoring algorithm. To reconstruct the causal relationships between these nodes from historical observation data, a direct linear non-Gaussian causal discovery algorithm was employed. Its core operation involves constructing a matrix from the long-term daily series of variables for each node, iteratively performing a "regression-residual independence test" step to gradually determine the causal order of the nodes, and estimating the direction and initial strength of the causal edge between each pair of nodes. The causal graph then needs to be adjusted by textual semantics to adapt to changes in the strength of physiological associations under different physical and mental situations. Specifically, a two-layer fully connected causal strength adjustment network is introduced, accepting textual semantic embedding vectors as input and outputting a vector with a dimension equal to the total number of causal edges. This vector is compressed to between 0 and 1 using a sigmoid function, serving as the correction coefficient for each edge. Multiplying the initial strength by the corresponding correction coefficient yields the corrected causal strength, forming the final causal graph. The graph contains a directed path from step count and activity intensity to resting heart rate.

[0028] Next, the causal relationship graph and multi-scale deep features are fed into a graph attention network. Each layer in the graph attention network calculates the attention weights of neighboring features only along the direction of the causal edge. The weight calculation considers not only the similarity between node features but also an additional multiplication by the modified causal strength, thereby strengthening the information flow of strong causal paths. After stacking two layers of graph attention, each physiological node, based on its original multi-scale features, absorbs the features of its causal parent node, forming a comprehensive feature matrix with enhanced causal structure. This matrix contains both temporal evolution information and embedded causal dependencies.

[0029] Next comes the crucial step of stripping away behavioral baseline drift. From the comprehensive feature matrix, features corresponding to two behavioral nodes—steps and activity intensity over the past 21 days—are extracted and averaged daily to form a long-term behavioral feature sequence, used to characterize the user's recent stable behavioral patterns. Using the causal path from the behavioral node to the resting heart rate node in the causal relationship graph, the drift caused by changes in long-term behavioral patterns on the resting heart rate baseline is quantified. Specifically, the long-term behavioral feature sequence is subtracted from its respective historical baseline value to obtain the change in behavioral features. Then, based on the product of the strengths of each causal edge on the causal path and a pre-fitted linear transfer coefficient between nodes, an initial drift, expressed in minutes of heart rate change, is calculated. This is then corrected using contextual weighting: the semantic embedding vector of the current day's text is fed into a gated network consisting of a fully connected layer and a sigmoid activation function. The output is a weighted coefficient corresponding to a specific time point; a coefficient close to 1 indicates that the influence of behavior on heart rate is amplified by the context, while a coefficient close to 0 indicates suppression. The initial drift is then multiplied point by point by the weighted coefficient to obtain the corrected drift. Next, a historical baseline value for resting heart rate is determined. From the past 30 days, all resting heart rate observations with behavioral patterns (daily steps and activity intensity) within the same quantile range as the current time point are selected, and the median is used as the baseline. The difference between the current resting heart rate observation and the baseline value is the total fluctuation. Subtracting the corrected drift from the total fluctuation, the remaining absolute value is the pathological shift, with its sign indicating whether the shift is increasing or decreasing. An adaptive threshold is set, updated every 3 days based on the 95th quantile of the pathological shift sequence over the past 90 days. When the pathological shift at 3 consecutive sampling points exceeds the threshold, a pathological shift event is identified. The pathological shifts at each point during the event are summed with exponentially decaying weights to obtain a cumulative shift value, which serves as the anomaly assessment baseline for this event, used to quantify the severity of the anomaly.

[0030] In the final prediction stage, the comprehensive feature matrix and the anomaly assessment baseline are input into a multi-task sequence coding network. The shared encoder is a two-layer bidirectional long short-term memory network with a hidden layer dimension of 128. It reads the comprehensive feature matrix step by step and outputs a shared temporal representation vector sequence, taking the vector at the last time step as the condensed representation of the entire sequence. The first dedicated task branch concatenates this representation vector with the anomaly assessment baseline, then passes it through a fully connected regression layer to output the predicted resting heart rate value 24 hours later. The second dedicated task branch is responsible for anomaly warning. It first replicates and expands the anomaly assessment baseline to the same dimension as the representation vector, then performs L2 normalization on both the representation vector and the expanded baseline vector, and calculates the Euclidean distance between them. The dynamic threshold is determined by the 95th percentile of the distance sequence over the past 60 days; when the distance exceeds the threshold, a warning signal is triggered. The third branch has a similar structure to the first branch, except that the regression target is changed to the predicted blood oxygen saturation value 24 hours later. Finally, the predicted resting heart rate value, predicted blood oxygen value, and warning flag are output as the physiological state prediction result.

[0031] The principle behind this technical solution is that changes in daily habits, such as consistent exercise, cause a slow drift in the resting heart rate baseline. This change is a benign physiological adaptation and should not be considered abnormal. By constructing a causal graph and quantifying the causal effect of behavior on resting heart rate, this drift can be accurately separated from the total heart rate fluctuation. The remaining pathological offset truly reflects heart rate abnormalities caused by factors such as infection, stress, or underlying diseases. Integrating textual semantics gives the model the ability to perceive subjective feelings and contextual differences, making drift estimation and abnormality threshold determination more closely aligned with the individual's current state. A multi-task shared representation network allows related tasks to mutually reinforce each other, improving the consistency of heart rate prediction, blood oxygenation prediction, and abnormality warning while reducing computational overhead. Therefore, the entire method exhibits high detection sensitivity and a low exercise-related false alarm rate.

[0032] In a preferred embodiment, the step of acquiring a user's multi-source heterogeneous health data, performing semantic alignment and cross-modal fusion on the time-series numerical physiological data and text health records to generate a semantically enhanced structured health data sequence includes the following steps: The system collects time-series numerical physiological data, including heart rate, blood oxygen saturation, steps taken, and accelerometer signals, using wearable devices. It identifies resting time periods based on the accelerometer signals and extracts the corresponding heart rate values ​​within those periods as resting heart rate samples. Activity intensity data is calculated from the accelerometer signals. Sleep duration data is collected using the wearable device's built-in sleep monitoring module. Finally, a health management application on a mobile terminal collects symptom descriptions and emotion tags input by the user based on preset question templates, which are used as the text health record. The text health record is input into a pre-trained language model to obtain a text semantic embedding vector; the time-series numerical physiological data is encoded through a one-dimensional convolutional neural network to obtain a numerical feature embedding vector. The text semantic embedding vector is aligned with the numerical feature embedding vector based on the timestamp, and an attention fusion layer is used to generate a semantically enhanced structured health data sequence.

[0033] Specifically, taking a smart bracelet integrating a photoplethysmography (PPG) sensor and a triaxial accelerometer as an example, the wearable device records heart rate, blood oxygen saturation, and triaxial accelerometer signals at a frequency of once per minute, and accumulates the steps per minute using a built-in step counting algorithm. To obtain a pure resting heart rate sample, it is necessary to identify resting periods from the continuous acceleration signal. In practice, a 5-minute sliding window is used, sliding forward once every 10 seconds, calculating the resultant acceleration sequence of the triaxial acceleration within the window, and then calculating the mean absolute deviation of the resultant acceleration. If the deviation value is lower than a preset low activity threshold of 0.02g, and the number of steps within the window is zero, then the window is marked as a resting period, and the median of the heart rate values ​​within the window is extracted as a resting heart rate sample point. The method for calculating the mean absolute deviation of the resultant acceleration is as follows: first, take the Euclidean norm of the raw acceleration signal every 10 seconds to obtain the magnitude of the resultant acceleration, subtract the gravitational acceleration of 1g, take the absolute value, and then calculate the average value over 5 minutes. Activity intensity data is calculated by integrating the absolute value of the sum of acceleration signals per minute after removing gravity, resulting in an activity count. This count is proportional to the energy consumed during physical activity and can be directly used as an indicator of activity intensity. Sleep duration data comes from the band's built-in sleep monitoring module. This module analyzes low-frequency fluctuations in heart rate and loss of body movement patterns, dividing the nighttime period into deep sleep, light sleep, and REM sleep stages, and summing the duration of each stage to obtain the total sleep duration. Text health record collection relies on a health management application on the mobile device. The application interface presents a set of preset question templates, such as "Did you feel chest tightness today?" and "Did you feel unusually tired today?", with yes or no buttons under each question. The emotion section provides a list of eight emotion words, including calm, tension, joy, and sadness, for single selection. After the user submits, the application concatenates the selected symptom and emotion word into a short sentence, such as "chest tightness, fatigue, tension," and stores it in the local database along with the submission timestamp, forming a text health record.

[0034] After obtaining the data, the language model used in the semantic alignment stage was PubMedBERT, pre-trained on biomedical literature and clinical records. This model employs a Transformer architecture, capable of encoding input short sentences into semantically rich 768-dimensional vectors. In practice, the concatenated text generated by the user daily was input into the model, and the hidden layer vector corresponding to the classification label position in the model's output was taken as the text semantic embedding vector. For numerical feature embedding, the structure of the one-dimensional convolutional neural network was designed as follows: the first convolutional layer has a kernel size of 5, a stride of 1, and outputs 64 channels, followed by a max-pooling kernel of size 2; the second convolutional layer has a kernel size of 5 and outputs 128 channels, also followed by pooling; the third convolutional layer has a kernel size of 3 and outputs 128 channels, which, after global average pooling, is followed by a fully connected layer to map the dimensions to 768 dimensions, matching them with the text embedding dimensions to obtain the numerical feature embedding vector. The timestamp alignment process employs a daily-level alignment strategy: the element-wise average of all numerical feature embedding vectors collected from midnight to midnight is taken to obtain a daily average numerical embedding vector representing the overall physiological state of the day. This vector is then paired with text embedding vectors generated from user-submitted text records from the previous night to the current day, as symptom and emotion records often reflect the overall feeling within the past 24 hours. For cross-modal fusion, the attention fusion layer uses scaled dot product multi-head attention with 8 heads. The daily average numerical embedding vector is used as the query, and the text semantic embedding vector as the key and value. After calculating the attention weights, the value vectors are weighted and summed to obtain a context vector that fuses textual information. This context vector is residually concatenated with the original numerical embedding vector and then normalized to obtain a time-step feature of the semantically enhanced structured health data sequence. The daily sequences are then combined to form a physiological data sequence carrying subjective semantic labels on the timeline.

[0035] In a preferred embodiment, the step of performing multi-scale deep feature extraction on the structured health data sequence and simultaneously constructing a causal relationship graph with physiological indicators as nodes; based on the causal relationship graph, obtaining a comprehensive feature matrix with enhanced causal structure by propagating and aggregating the spatiotemporal features of the nodes through a graph neural network, includes the following steps: A temporal convolutional network containing large, medium and small convolutional kernels is used to extract local features of the structured health data sequence at multiple time scales in parallel, and then spliced ​​to obtain multi-scale deep features; Using heart rate, blood oxygen saturation, exercise steps, activity intensity, and sleep duration as nodes, a direct linear non-Gaussian causal discovery algorithm is used to learn the causal direction and initial strength between nodes from historical data. The text semantic embedding vector corresponding to the text health record is input into the causal strength adjustment network, and the correction coefficients of each causal edge are output. The initial strength is weighted and adjusted to obtain the corrected causal strength, and the causal relationship graph is constructed. The causal relationship graph and the multi-scale deep features are input into a graph attention network. The graph attention layer propagates node features along the causal edges and aggregates neighbor node information, outputting the spatiotemporal representation of each node to form the comprehensive feature matrix.

[0036] Specifically, firstly, the semantically enhanced structured health data sequence is structured with daily time steps, and each time step has a feature dimension of 768. Multi-scale deep feature extraction employs a temporal convolutional network containing large, medium, and small kernels with lengths of 3, 5, and 7, corresponding to small, medium, and large scales, respectively. To avoid insufficient receptive fields in deep networks, dilation factors of 2 and 4 are introduced in the medium-scale and large-scale branches, respectively. This allows the small-scale branch to capture instantaneous heart rate variability and transient blood oxygen fluctuations within a day, the medium-scale branch to capture the slow adjustments in physiological state over approximately one week, and the large-scale branch to perceive long-term trends over up to one month. Each branch consists of two layers of dilated convolution, weight normalization, and a ReLU activation function, with a uniform output channel count of 64. Therefore, three 64-dimensional feature vectors are obtained each day, which are concatenated to form a 192-dimensional multi-scale deep feature vector. The multi-scale deep feature vector across the entire sequence is a matrix of time steps multiplied by 192.

[0037] The synchronously constructed causal relationship graph uses five variables—heart rate, blood oxygen saturation, steps, activity intensity, and sleep duration—as nodes. Specifically, the heart rate node uses the median of all resting heart rate samples for the day to represent a resting physiological state; steps and activity intensity are taken as the total daily steps and average daily activity count, respectively; blood oxygen saturation is taken as the average value during the night's sleep to reduce motion artifacts; and sleep duration is the total sleep time of the previous night. A direct linear non-Gaussian causal discovery algorithm is used to learn the causal direction and initial strength between nodes from historical data. The specific process is as follows: Daily values ​​of the above five variables for at least 60 days are collected from the user, forming a 60x5 data matrix. The algorithm first standardizes all variables, and then determines the causal order according to the exogeneity priority principle through iterative linear regression and residual independence tests based on non-Gaussian distributions. In each iteration, linear regression is performed on the remaining variables as response variables and the other variables as explanatory variables. The residuals are calculated, and the independence of the residuals from the explanatory variables is tested using a kernel method. The response variable with the strongest independence is selected as the current most exogenous node, removed from the variable set, and its causal edges pointing to other nodes and regression coefficients are recorded as initial strengths. This process is repeated until all nodes are sorted, ultimately outputting a directed acyclic graph and the initial strength value for each edge. Subsequently, textual semantic embedding vectors are used to personalize the causal strength. A causal strength adjustment network is constructed, containing two fully connected layers. The first layer has 64 neurons connected to ReLU, and the second layer directly outputs a vector whose length is equal to the number of causal edges in the graph. The textual semantic embedding vector for the day is input into this network, and the output vector is activated by a sigmoid function to obtain the correction coefficient for each edge. Multiplying the initial strength by the correction coefficient yields the corrected causal strength incorporating contextual information, thus completing the construction of the causal relationship graph.

[0038] The aforementioned causal relationship graph is combined with multi-scale deep features and input into a graph attention network for the propagation and aggregation of spatiotemporal features of nodes. Each layer of the graph attention network is limited to operations along the direction of the causal edge, maintaining the causal rationality of information flow. For each time step, the 192-dimensional multi-scale deep features are divided into sub-vectors corresponding to five nodes, serving as the initial features of the nodes. Within the graph attention layer, for each causal edge pointing from node j to node i, the attention coefficient is calculated using the following formula: first, the features of node i and node j are multiplied by a shared linear transformation matrix, concatenated, and then productd through a learnable weight vector. This is followed by LeakyReLU activation, and the resulting value is multiplied by the corrected intensity corresponding to the causal edge. Softmax normalization is then performed across all neighboring nodes pointing to i. The updated features of node i are the weighted sum of the transformed features of all causal parent nodes j, which are then activated by ELU. Stacking two layers of the graph attention network allows each node to not only aggregate information from its direct parent nodes but also indirectly receive the influence of second-order causal neighbors. For example, the resting heart rate node can integrate features transmitted from the number of steps and activity intensity via the autonomic nervous system, as well as the indirect effects of sleep duration through metabolic pathways. By concatenating the updated features of all nodes at each time step, a comprehensive feature matrix with enhanced causal structure is obtained.

[0039] In a preferred embodiment, the step of extracting long-term behavioral feature sequences within a preset time window from the comprehensive feature matrix, quantifying the drift of long-term behavioral patterns on the resting heart rate baseline using the causal path from behavioral nodes to resting heart rate nodes in the causal relationship graph, and applying context-weighted correction to the drift using the semantic features of the text health record includes the following steps: Using the number of steps and activity intensity as behavioral nodes, the daily average number of steps and daily average activity intensity for multiple consecutive days are extracted from the comprehensive feature matrix to form the long-term behavioral feature sequence; In the causal relationship graph, the causal path from the behavior node to the resting heart rate node is extracted; the long-term behavior feature sequence is input into a pre-trained causal effect prediction network, which takes the change in behavior feature as input and the change in the resting heart rate baseline as output, and obtains the initial drift amount through forward propagation; The text semantic embedding vector corresponding to the text health record is input into a gating network, which is a single-layer fully connected network followed by a sigmoid activation function to generate weighting coefficients corresponding to each time point of the initial drift. The initial drift is multiplied element by element by the weighting coefficients to obtain the corrected drift.

[0040] Specifically, the long-term behavioral feature sequence refers to the time sequence of features reflecting stable behavioral habits, composed of two behavioral nodes: step count and activity intensity, within a continuous multi-day time window. The preset time window is typically 21 days, as physiologically, approximately three weeks of behavioral habits are sufficient to produce an observable baseline impact on autonomic nervous system regulation. During extraction, feature vectors corresponding to the step count and activity intensity nodes are extracted one by one from the causal structure-enhanced comprehensive feature matrix, step by step. When constructing the causal graph, the input features for these two nodes are the daily average step count and daily average activity intensity after minimum-maximum value normalization. After propagation through the graph attention network, their vectors still retain the ability to encode the original behavioral quantities. Therefore, by recording the minimum and maximum values ​​saved during the normalization phase and performing inverse normalization on the first dimension of the extracted feature vector, the actual daily average step count and daily average activity intensity can be restored. Concatenating the two values ​​for each day within the window constitutes the long-term behavioral feature sequence, which is a matrix with a length equal to the number of days in the window and a dimension of 2.

[0041] Next, a causal relationship graph is used to quantify the drift. A causal path refers to a pathway in a causal directed graph that starts from a behavior node and hops along directed edges to reach a resting heart rate node. All such paths need to be enumerated in the graph using a depth-first search, recording the nodes and edges traversed by each path. The impact of behavioral pattern changes on resting heart rate propagates along these paths; therefore, the estimation of drift needs to integrate the effects of all paths. The causal effect prediction network is a pre-trained multilayer perceptron for quantifying this cumulative causal effect. Before network training, samples are constructed on historical user data: a relatively stable reference period is selected, and its average daily steps and average activity intensity are calculated as the behavioral baseline; then, during periods of significant behavioral change, the daily change in behavioral characteristics relative to the behavioral baseline is calculated as input, and the difference between the observed resting heart rate on that day and the historical resting heart rate baseline is calculated as the output target. The historical resting heart rate baseline is the average resting heart rate during the same period within the stable behavioral period. Pairing multiple sets of such input changes with baseline change outputs, a fully connected network with two hidden layers and 64 neurons each is trained using mean squared error loss, resulting in a causal effect prediction network. In actual quantification, the daily average of the current long-term behavioral feature sequence is subtracted from the behavioral baseline to obtain the behavioral feature change, which is then fed into the network for forward propagation. The network output is the initial drift, with units identical to heart rate, representing the expected deviation of the resting heart rate baseline caused by long-term behavioral changes.

[0042] Context-weighted correction is achieved using a gating network. The gating network is designed as a single-layer fully connected network followed by a sigmoid activation function. The input is the text semantic embedding vector corresponding to each day, and the output is a scalar between 0 and 1, i.e., the weighting coefficient. The text semantic embedding vector is provided by the language model in the previous steps, encoding the symptoms and emotional state of the day. A coefficient close to 1 indicates that the impact of behavior on resting heart rate is relatively fully expressed in the current physical and mental context; for example, when physically fatigued and emotionally stressed, the heart rate response from the same amount of exercise may be greater. A coefficient close to 0 indicates that the impact of behavior is suppressed; for example, on a recovery day with sufficient deep sleep and a peaceful mood. Multiplying the initial daily drift by the corresponding weighting coefficient yields the corrected drift. This correction ensures that the behavioral baseline drift is no longer a rigid average response, but rather dynamically scales with the user's subjective state, thus significantly improving the accuracy of individualized estimation.

[0043] In a preferred embodiment, removing the drift from the total fluctuation between the current observed value of resting heart rate and the historical baseline value to obtain the pathological offset after removing behavioral baseline drift includes the following steps: The difference between the observed resting heart rate at the current time point and the historical baseline value is calculated as the total fluctuation; the historical baseline value is the average resting heart rate under similar behavioral patterns during the same period within a preset number of days in the past. Subtract the drift from the total fluctuation to obtain the residual value. Take the absolute value of the residual value as the pathological offset and determine the offset direction based on the sign of the residual value.

[0044] Specifically, the resting heart rate observation refers to the heart rate value measured during a resting period determined by accelerometer signals, which can be the median heart rate of the most recent resting window. The historical baseline value is a personalized reference standard used for comparison. When determining the historical baseline value, a time window is first defined, typically the past 30 days. Then, filtering is performed based on two conditions: contemporaneous time and similar behavioral patterns. Contemporaneous time means dividing a 24-hour day into six 4-hour periods; the time period in which the current resting heart rate observation occurs must be consistent with the historical time period. The similar behavioral pattern condition is achieved by comparing the quantile intervals of behavioral characteristics: the 10th and 90th quantiles of the total steps and average activity intensity for each day over the past 30 days are calculated, dividing the range into low, medium, and high intervals. Only when the total steps and average activity intensity of a historical day fall within the same interval as the current day are they considered similar to the current behavioral pattern. The resting heart rate observations for the same time period on historical days that meet both conditions are selected, and their arithmetic mean is used as the historical baseline value. For example, if the current timeframe is 2 PM and both step count and activity intensity are in the high range, then the baseline is the average resting heart rate of all participants over the past 30 days who simultaneously met the criteria of 2 PM and high activity levels. If the number of eligible samples is less than 3 days, the quantile range is widened to the full range to ensure baseline stability.

[0045] Total fluctuation is calculated by subtracting the historical baseline value from the current resting heart rate observation; the difference can be positive or negative. When removing the corrected drift from the total fluctuation, the drift value comes from the corrected drift value at the corresponding time point in the previously obtained sequence. This drift value can be positive or negative, indicating whether behavioral changes have caused the resting heart rate baseline to rise or fall. The remaining value after subtraction indicates the direction of the pathological shift: a positive value means the heart rate is still higher than expected after adjusting for behavioral effects, while a negative value means it is lower than expected. Taking the absolute value of the remaining value yields the pathological shift, used to measure the degree of abnormality without indicating direction. For example, if a user starts running regularly, and the estimated long-term behavioral drift is -5 beats per minute (meaning the resting heart rate baseline is expected to decrease by 5 beats per minute), but the actual observed resting heart rate is 2 beats per minute higher than the historical baseline, then the total fluctuation is 2. Subtracting -5 gives 7, and the remaining value is positive 7. The pathological shift is 7 beats per minute, indicating that a potential pathological factor may have caused the abnormally high heart rate. This stripping calculation is performed once a day, forming a time series of pathological offsets.

[0046] The principle behind this process is that changes in resting heart rate are driven by both long-term behavioral adaptation and short-term pathological conditions. Baseline drift caused by behavioral adaptation is slow and gradual, while pathological factors such as inflammation, infection, and psychological stress can lead to faster and more significant deviations. Directly using total fluctuation as an abnormal indicator can easily misinterpret a benign decrease in heart rate due to improved exercise habits as an alarm for bradycardia. By accurately subtracting the estimated drift, the remaining pathological offset becomes purer, with its value and direction directly corresponding to the intensity of pathological interference. This separation significantly improves the specificity of abnormality detection, reduces false alarms, and allows subsequent warning logic to focus on genuine health threats. Furthermore, the time-segmented and behavior-mode-segmented baseline calculation method fully considers the normal influence of the body's diurnal rhythm and activity level on resting heart rate, making the baseline more scientific and reasonable.

[0047] In a preferred embodiment, determining that a pathological deviation event has occurred when the pathological deviation exceeds a preset threshold, and constructing an abnormality assessment baseline based on the cumulative degree of the pathological deviation, includes the following steps: A pathological offset threshold is set, which is adaptively updated based on the preset upper quantile of the user's historical pathological offset sequence; when the pathological offset at multiple consecutive sampling time points exceeds the threshold, a pathological offset event is determined to have occurred. The pathological offset values ​​at each sampling time point during the pathological offset event are weighted and summed to obtain the cumulative offset value, which is then used as the anomaly assessment baseline.

[0048] Specifically, the pathological offset threshold is a dynamic boundary for judging whether a single-point offset is sufficiently abnormal. To avoid the inability of a fixed threshold to adapt to individual differences and slow physiological changes, an adaptive strategy is adopted. The threshold is updated based on the preset upper quantile of the user's historical pathological offset sequence. During operation, a 90-day sliding window of historical pathological offsets is maintained, with newly calculated pathological offsets added and the oldest values ​​removed daily. Every 3 days, the 95th percentile of all values ​​within the window is recalculated as the new threshold. The 95th percentile means that only 5% of historical offsets will exceed it, thus ensuring minimal false triggers when the user's physiological state is stable. This threshold is automatically adjusted according to the user's own fluctuations; for example, after seasonal changes or improvements in physical condition, the amplitude of pathological offsets generally decreases, and the threshold will also be adjusted downwards.

[0049] Determining the occurrence of a pathological shift event requires avoiding false alarms due to single noise or measurement errors; therefore, a continuous confirmation mechanism is introduced. A continuous sampling point count parameter is set, representing multiple consecutive sampling time points. "Multiple" means at least two, for example, 2, 3, 4, 5, or more, which can be selected by those skilled in the art based on actual monitoring accuracy and sensitivity requirements. Preferably, this continuous sampling point count parameter is set to 3, meaning that the pathological shift values ​​at 3 consecutive sampling time points must be strictly greater than the current threshold for a pathological shift event to be considered to have occurred. The sampling time point refers to the time when the pathological shift value is calculated daily, usually taken at the same time each day (e.g., the resting period in the early morning). If the condition is not met on a certain day, the continuous count is reset to zero. This requirement of continuous exceeding the threshold filters out isolated fluctuations, ensuring that only persistent pathological states are captured.

[0050] Once an event is confirmed, its cumulative severity needs to be quantified, i.e., an anomaly assessment baseline needs to be established. For each sampling time point covered during the pathological shift event, the pathological shift is multiplied by a time decay weight and then summed. The decay weight is designed in exponential form, with the weight equal to exp(negative time interval divided by a time constant), and the time constant is set to 2 days, so that the shift contribution is greater closer to the current time. The weighted sum yields the cumulative shift value, which comprehensively reflects the intensity and duration of the event. This cumulative shift value is recorded as the anomaly assessment baseline for this event and used as a reference in subsequent tasks. If the pathological shift continues to exceed the threshold during the event's duration, the cumulative value is continuously updated. The baseline is fixed after the event ends until it is reset upon the triggering of the next new event.

[0051] In a preferred embodiment, the step of inputting the integrated feature matrix and the anomaly assessment baseline into a multi-task sequence coding network, extracting temporal representations by a shared encoder, and having a dedicated task branch perform parallel future resting heart rate regression prediction and abnormal event warning based on the deviation between the temporal representations and the anomaly assessment baseline, and outputting physiological state prediction results, includes the following steps: The shared encoder uses a bidirectional long short-term memory network to encode the comprehensive feature matrix, thereby obtaining a shared temporal representation vector; The first task branch concatenates the shared temporal representation vector with the anomaly assessment baseline and inputs it into the fully connected regression layer, outputting the predicted resting heart rate value at a future preset time point; The second task branch calculates the Euclidean distance between the shared temporal representation vector and the anomaly assessment baseline. First, the anomaly assessment baseline is copied and expanded to the same dimension as the shared temporal representation vector, and both are L2 normalized. Then, the Euclidean distance between the normalized vectors is calculated. The dynamic threshold is adaptively updated based on the preset lower quantile of the user's historical Euclidean distance sequence. When the Euclidean distance is lower than the dynamic threshold, an anomaly warning signal is generated. The third task branch inputs the shared temporal representation vector into the fully connected regression layer and outputs the predicted blood oxygen saturation value at a future preset time point. The predicted resting heart rate, the predicted blood oxygen saturation, and the abnormal warning signal are collectively output as the physiological state prediction result.

[0052] Specifically, the shared encoder employs a bidirectional long short-term memory network, or BiLSTM for short. The network's structure consists of two stacked LSTM layers: a forward LSTM that reads the comprehensive feature matrix from past to present along the time step, and a backward LSTM that reads the same matrix from future to past along the time step. Each LSTM layer has 128 hidden units, and the time step length corresponds to the number of rows in the comprehensive feature matrix, i.e., the number of historical days included in the modeling, typically the past 60 days. At each time step, both the forward and backward LSTMs output a 128-dimensional hidden state vector. These are concatenated element-wise to obtain a 256-dimensional bidirectional hidden state. The 256-dimensional bidirectional hidden state vector from the last time step is taken as the shared temporal representation vector for the entire input sequence, encoding the evolutionary trajectory of the user's physiological state under causal constraints over a past period.

[0053] The first task branch is responsible for future resting heart rate regression prediction. This branch concatenates the shared temporal representation vector with the anomaly assessment baseline. The anomaly assessment baseline is a scalar derived from the cumulative offset value, scaled to between 0 and 1 using max-min normalization to avoid significant dimensionality differences when directly concatenating with the high-dimensional representation vector. The concatenated vector forms a 257-dimensional vector, which is fed into a fully connected regression layer. The fully connected regression layer consists of two hidden layers: the first layer has 64 neurons using the ReLU activation function, and the second layer has 32 neurons, also using ReLU. Finally, a single-neuron linear output layer directly outputs the predicted resting heart rate value at a preset future time point. The preset time point is assumed to be 24 hours in the future, and the corresponding training target is the actual observed resting heart rate value at that moment. During training, the mean squared error loss function is used, and the parameters of the shared encoder and this branch are updated simultaneously through backpropagation. The input to this branch includes an anomaly assessment baseline, which allows the regression model to explicitly know whether the current user is in an abnormal state with a high degree of cumulative deviation. This allows the model to assign different regression tendencies to periods with large pathological deviations when predicting future heart rate, making the predicted values ​​closer to the actual pathological evolution trend.

[0054] The second task branch is responsible for anomaly event early warning. The core calculation in this branch is the Euclidean distance between the shared temporal representation vector and the anomaly assessment baseline. Since the anomaly assessment baseline is a scalar while the shared temporal representation vector is 256-dimensional, direct distance calculation requires dimensionality matching. In practice, the anomaly assessment baseline is first copied and expanded 256 times, generating a 256-dimensional constant vector where each element equals the same normalized baseline value. Then, L2 normalization is performed on both the shared temporal representation vector and the expanded baseline vector. L2 normalization involves calculating the Euclidean norm of each vector and then dividing each element by that norm, resulting in a normalized vector with a length of 1. This step eliminates the influence of overall amplitude changes in the representation vector on the distance calculation, ensuring the distance only reflects directional differences. Finally, the Euclidean distance between the two normalized vectors is calculated, which is the square root of the sum of the squares of the elements of their difference. This distance characterizes the similarity between the current comprehensive physiological state representation and the high-risk state pattern represented by the anomaly accumulation baseline; a smaller distance indicates a closer resemblance to the state characteristics of historical anomalies.

[0055] A dynamic threshold is used to determine if the Euclidean distance is small enough to trigger an alert. The dynamic threshold is updated by maintaining a 60-day historical Euclidean distance sequence sliding window. The distance for each day is calculated and added to the window, while the oldest value is removed. Every 7 days, the lower 5% quantile of all distance values ​​within the window is used as the new threshold. The lower 5% quantile is chosen because the distance value should be significantly smaller than normal when an anomaly occurs; that is, the closer to the baseline vector direction, the more dangerous the situation. When the Euclidean distance on a given day falls below this dynamic threshold, the second task branch generates an anomaly alert signal. The signal can be a Boolean flag or include a distance value as a confidence reference. This alert logic is independent of but complementary to the pathological deviation event determination. Pathological deviation events focus on the magnitude and persistence of univariate deviations, while this branch makes a judgment from the perspective of overall deviation represented by a high-dimensional comprehensive characterization.

[0056] The third task branch is responsible for predicting future blood oxygen saturation. This branch has a similar structure to the first branch, but it does not concatenate the anomaly assessment baseline. It directly inputs the shared temporal representation vector into a separate fully connected regression layer. This layer has the same structure as the first branch: two hidden layers followed by a single neuron output. The output target is the predicted blood oxygen saturation value at a predetermined future time point, typically 24 hours later. During training, mean squared error loss is also used, updating only the parameters of the shared encoder and this branch. The blood oxygen prediction of the third branch is set up separately because the pattern of blood oxygen saturation change is related to, but not synchronous with, resting heart rate. For example, in the early stages of respiratory abnormalities, blood oxygen may decrease first, while heart rate increases compensatorily later. Multi-task parallel learning helps the shared representation capture more comprehensive physiological coupling information.

[0057] Finally, the predicted resting heart rate output from the first branch, the predicted blood oxygen saturation output from the third branch, and the abnormal warning signal generated by the second branch are combined as the physiological state prediction result, which is presented to the user through the mobile application interface or uploaded to the health management platform.

[0058] In a preferred embodiment, determining the historical baseline value includes the following steps: The long-term behavioral feature sequence is subjected to trend detection to determine whether the user has entered a behavioral pattern transition period. The trend detection adopts the Mann-Kendall trend test. When the absolute value of the test statistic exceeds the critical value corresponding to the preset significance level, it is determined that the user has entered a behavioral pattern transition period. When the duration of the behavioral pattern transition exceeds a preset physiological adaptation threshold, the user is determined to be in a physiological adaptation state; the physiological adaptation threshold is set to 21 days based on population statistics. Under the physiological adaptation state, the historical baseline value is frozen, and the steady-state baseline before the behavioral pattern transition period is used as the historical baseline value. After the behavioral pattern transition period ends, when the coefficient of variation of the resting heart rate is lower than the preset coefficient of variation threshold within a 7-day sliding window, it is determined that the variability of the resting heart rate has returned to steady state conditions, the freeze is lifted, and the historical baseline value is recalculated based on the new steady state data after the behavioral pattern transition period.

[0059] Specifically, the Mann-Kendall trend test is used to detect trends in long-term behavioral characteristic sequences. The Mann-Kendall test is a non-parametric statistical method that does not require data to follow a normal distribution, making it well-suited for handling physiological behavioral sequences with seasonal fluctuations and outliers. In practice, the average daily steps and average daily activity intensity for at least 14 consecutive days are extracted from the comprehensive feature matrix, forming two sequences. The Mann-Kendall test is performed independently on each sequence, and its test statistic Z is calculated. The specific steps are: first, calculate the sign of all subsequent values ​​minus the preceding values ​​in the sequence, count the number of positive and negative signs to obtain Kendall's S statistic; then, calculate the standardized Z value based on the sequence length and the variance of S. When the absolute value of Z exceeds the critical value corresponding to the preset significance level (usually 1.96 corresponding to a significance level of 0.05), the sequence is considered to have a significant monotonically increasing or decreasing trend, indicating that the user has entered a behavioral pattern transition period. A behavioral pattern transition period means that the user is consciously changing their exercise habits, such as starting a running program leading to a continuous increase in steps and activity intensity, or a continuous decrease in activity due to injury or illness.

[0060] When a continuous behavioral pattern transition is detected, and its duration exceeds a preset physiological adaptation threshold, the user is considered to be in a state of physiological adaptation. The physiological adaptation threshold, based on statistical understanding of the autonomic nervous system adaptation time window in exercise physiology research, is set at 21 days. The general principle behind this duration is that after approximately three weeks of regular aerobic exercise, a noticeable increase in parasympathetic tone begins to appear at rest, and the resting heart rate tends to slowly decrease and gradually stabilize at a new, lower level. If the behavioral change persists for more than 21 days, it indicates that the behavioral change has been persistent enough, and the body is likely undergoing this physiological readaptation process.

[0061] After determining that the body is in a state of physiological adaptation, measures are taken to freeze historical baseline values. Freezing means ceasing to update the historical data window used for baseline calculations, no longer including new resting heart rate observations after the start of the behavioral pattern transition period, and directly using the steady-state baseline before the behavioral pattern transition period as the current historical baseline value. Specifically, the most recent behavioral stability period of at least 21 days before the start of the behavioral pattern transition period is identified, and the average resting heart rate under similar behavioral pattern conditions during the same period is taken as the frozen baseline, which remains unchanged throughout the entire physiological adaptation period. The purpose of this is to prevent the baseline from being continuously updated to catch up during the slow decline of the behavior-driven baseline, which could cause the decline that is originally part of benign adaptation to be absorbed by daily baseline updates, thus masking the ability to detect pathological shifts.

[0062] After the behavioral pattern transition period ends, it's necessary to determine if the resting heart rate has stabilized again before unfreezing and introducing a new baseline. The condition for resting heart rate variability to return to steady state is that the coefficient of variation (COP) of resting heart rate remains below a preset COP threshold for a continuous 7-day sliding window. The COP is calculated by dividing the standard deviation of daily resting heart rate observations by the mean within the 7-day sliding window, then multiplying by 100 to obtain the percentage. The COP threshold is set based on the COP distribution of the user's historical stable periods, typically the 75th percentile or an empirical value such as 5%. When the COP remains below this threshold for a continuous 7-day window, it indicates that the daily fluctuations in resting heart rate have narrowed to a range comparable to the previous steady state, and physiological adaptation is essentially complete; at this point, the freeze is lifted. After lifting, the calculation of historical baseline values ​​resumes normal updates. At least 21 days of newly collected steady-state data after the behavioral pattern transition period are included in the window, and historical baseline values ​​are recalculated according to the previous rules for similar behavioral patterns and time periods to reflect the user's true resting heart rate baseline level under the new exercise habits.

[0063] In a preferred embodiment, the output of the causal effect prediction network further undergoes the following verification and correction steps: When a pathological offset event is determined to have occurred based on the pathological offset amount, the behavioral feature sequence and resting heart rate observation sequence within the preset backtracking window before the event occurs, as well as the semantic features of the text health record for the corresponding time period, are extracted. Using the starting time point of the backtracking window as the intervention point, and utilizing the causal relationship diagram and structural causal model, a counterfactual resting heart rate trajectory is generated under the assumption that the behavioral pattern is maintained before the intervention point. Calculate the cumulative deviation between the resting heart rate observation sequence and the counterfactual resting heart rate trajectory, and compare the cumulative deviation with the drift amount to obtain the drift amount estimation residual; In response to the drift estimation residual exceeding the preset tolerance, an instant correction sample is constructed using the change in behavioral features within the backtracking window and the cumulative deviation, and the parameters of the causal effect prediction network are updated step by step. The updated causal effect prediction network parameters are synchronized to the drift quantization step for drift estimation at subsequent time points.

[0064] First, this verification process is triggered when a pathological offset event is determined based on the pathological offset. Data within a preset backtracking window prior to the occurrence of the event is extracted, with the backtracking window length set to 14 days prior to the first time the pathological offset event exceeded the threshold. Within these 14 days, behavioral feature sequences, namely daily average steps and daily average activity intensity, as well as the corresponding resting heart rate observation sequence, and the semantic features of each day's textual health records, are extracted as the basis for subsequent counterfactual generation and comparison.

[0065] Next, using the starting time point of this backtracking window as the intervention point, and leveraging the pre-constructed causal relationship graph and the framework of the structural causal model, a counterfactual resting heart rate trajectory is generated assuming the behavioral pattern remains unchanged from its pre-intervention state. The structural causal model consists of functional relationships along each causal edge in the graph. For a causal path from a behavioral node to a resting heart rate node, the functional relationship can be approximated by a trained causal effect prediction network. The specific method for generating the counterfactual trajectory is as follows: First, determine the behavioral characteristic values ​​at the intervention point, namely the number of steps and activity intensity on the last day before the behavioral pattern change, and use these as constant behavioral inputs in the counterfactual scenario. Then, subtract the constant behavioral input from the behavioral baseline at the intervention point daily to obtain the behavioral change, which remains zero or experiences minimal natural fluctuations in the counterfactual scenario. Replace the text semantic embedding vectors corresponding to the daily constant behavioral inputs with the mean semantic embeddings before the intervention point, and input them into a gating network to obtain constant weighting coefficients. These constant inputs are then fed into the causal effect prediction network and the weighting step, outputting the daily counterfactual drift. Using the observed resting heart rate on the day of the intervention as the initial value, a counterfactual drift is added daily to obtain a counterfactual resting heart rate trajectory under the assumption that the behavior has not changed. This trajectory represents how the user's resting heart rate would evolve if the user continued to maintain the old behavioral habits.

[0066] Then, the daily cumulative deviation between the actual resting heart rate observation sequence and the counterfactual resting heart rate trajectory is calculated. The cumulative deviation is obtained by subtracting the counterfactual value from the observed value, summing the absolute values ​​of the daily differences, and obtaining a scalar. Simultaneously, the original corrected drift calculated daily by the causal effect prediction network within the backtracking window is also accumulated to obtain a drift estimate. The cumulative deviation is compared with the drift estimate, and the absolute value of the drift estimate is subtracted from the cumulative deviation to obtain the drift estimate residual. This residual reflects the magnitude of the estimation error of the causal effect prediction network during this pathological shift event. A larger residual indicates that the network's estimation of behavioral effects is less accurate, and a significant portion of the actual resting heart rate changes is incorrectly attributed to behavioral baseline drift, or that behavioral effects are significantly overestimated or underestimated.

[0067] Determine if the drift estimation residuals exceed a preset tolerance. The preset tolerance is set as the 90th quantile of the drift estimation residual sequence during historical healthy periods, or a fixed empirical value, such as twice per minute per day. If the residuals exceed the tolerance, it indicates that the causal effect prediction network performed poorly in this event and requires immediate correction. The correction method utilizes the change in behavioral characteristics within the backtracking window and the cumulative bias obtained above to construct an immediate correction sample. The input of the sample is the change in behavioral characteristics on the start day of the backtracking window relative to the baseline behavior at the intervention point, and the output target is set as the cumulative bias. This sample is used as training data to perform a single-step parameter update on the causal effect prediction network. The single-step update uses a small learning rate, such as stochastic gradient descent of 0.001, and only performs one forward and backpropagation update on the weights of the fully connected layers of the network, without involving other modules such as gating networks. This lightweight online update strategy can quickly write the relationship between newly observed behavior and resting heart rate effects into the network without waiting for sufficient samples to accumulate before retraining.

[0068] Finally, the updated causal effect prediction network parameters are immediately synchronized to the drift quantification step for drift estimation at subsequent time points. Synchronization means that the causal effect prediction network weights used in the next initial drift calculation are the newly revised versions, ensuring that future drift estimates reflect the latest learned individualized causal effects. The entire verification and correction process is executed once after each pathological shift event, forming a closed-loop feedback mechanism.

[0069] Example 2 Secondly, this application proposes a user physiological state prediction system based on multi-source health data, used to execute the user physiological state prediction method based on multi-source health data as described above, such as... Figure 2 As shown, it includes: The acquisition module 1 is configured to acquire multi-source heterogeneous health data from users, perform semantic alignment and cross-modal fusion on time-series numerical physiological data and text health records, and generate a semantically enhanced structured health data sequence. Matrix enhancement module 2 is configured to perform multi-scale deep feature extraction on the structured health data sequence and simultaneously construct a causal relationship graph with physiological indicators as nodes; based on the causal relationship graph, the spatiotemporal features of the nodes are propagated and aggregated through a graph neural network to obtain a comprehensive feature matrix with enhanced causal structure. Extraction module 3 is configured to extract long-term behavioral feature sequences within a preset time window from the comprehensive feature matrix, quantify the drift of long-term behavioral patterns on the resting heart rate baseline using the causal path from behavioral nodes to resting heart rate nodes in the causal relationship graph, and perform context-weighted correction on the drift using the semantic features of the text health record. The baseline construction module 4 is configured to remove the corrected drift from the total fluctuation between the current observed value of resting heart rate and the historical baseline value to obtain the pathological offset after removing the behavioral baseline drift; when the pathological offset exceeds a preset threshold, a pathological offset event is determined to have occurred, and an abnormal assessment baseline is constructed based on the cumulative degree of the pathological offset. Prediction module 5 is configured to input the comprehensive feature matrix and the anomaly assessment baseline into a multi-task sequence coding network, where a shared encoder extracts temporal representations, and a dedicated task branch performs future resting heart rate regression prediction and abnormal event warning in parallel based on the degree of deviation between the temporal representations and the anomaly assessment baseline, and outputs physiological state prediction results.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting user physiological state based on multi-source health data, characterized in that, Includes the following steps: Acquire users' multi-source heterogeneous health data, perform semantic alignment and cross-modal fusion on time-series numerical physiological data and text health records, and generate semantically enhanced structured health data sequences; Multi-scale deep feature extraction is performed on the structured health data sequence, and a causal relationship graph with physiological indicators as nodes is constructed simultaneously. Based on the causal relationship graph, the spatiotemporal features of the nodes are propagated and aggregated through a graph neural network to obtain a comprehensive feature matrix with enhanced causal structure. The long-term behavioral feature sequence within a preset time window is extracted from the comprehensive feature matrix, and the causal path from the behavioral node to the resting heart rate node in the causal relationship graph is used to quantify the drift of the long-term behavioral pattern on the resting heart rate baseline. The drift amount is corrected using context-weighted methods based on the semantic features of the text health records. From the total fluctuation between the current observed value of resting heart rate and the historical baseline value, the corrected drift amount is removed to obtain the pathological deviation amount after removing the behavioral baseline drift; when the pathological deviation amount exceeds a preset threshold, a pathological deviation event is determined to have occurred, and an abnormal assessment baseline is constructed based on the cumulative degree of the pathological deviation amount. The comprehensive feature matrix and the anomaly assessment baseline are input into a multi-task sequence coding network. The shared encoder extracts the temporal representation, and a dedicated task branch performs future resting heart rate regression prediction and abnormal event warning in parallel based on the deviation between the temporal representation and the anomaly assessment baseline, outputting the physiological state prediction result.

2. The method for predicting user physiological state based on multi-source health data according to claim 1, characterized in that, The process of acquiring users' multi-source heterogeneous health data, performing semantic alignment and cross-modal fusion on the time-series numerical physiological data and text health records to generate semantically enhanced structured health data sequences includes the following steps: The system collects time-series numerical physiological data, including heart rate, blood oxygen saturation, steps taken, and accelerometer signals, using wearable devices. It identifies resting time periods based on the accelerometer signals and extracts the corresponding heart rate values ​​within those periods as resting heart rate samples. Activity intensity data is calculated from the accelerometer signals. Sleep duration data is collected using the wearable device's built-in sleep monitoring module. Finally, a health management application on a mobile terminal collects symptom descriptions and emotion tags input by the user based on preset question templates, which are used as the text health record. The text health record is input into a pre-trained language model to obtain a text semantic embedding vector; the time-series numerical physiological data is encoded through a one-dimensional convolutional neural network to obtain a numerical feature embedding vector. The text semantic embedding vector is aligned with the numerical feature embedding vector based on the timestamp, and an attention fusion layer is used to generate a semantically enhanced structured health data sequence.

3. The method for predicting user physiological state based on multi-source health data according to claim 1, characterized in that, The process involves multi-scale deep feature extraction of the structured health data sequence and simultaneous construction of a causal relationship graph with physiological indicators as nodes. Based on the causal relationship graph, a comprehensive feature matrix with enhanced causal structure is obtained by propagating and aggregating the spatiotemporal features of the nodes through a graph neural network. This includes the following steps: A temporal convolutional network containing large, medium and small convolutional kernels is used to extract local features of the structured health data sequence at multiple time scales in parallel, and then spliced ​​to obtain multi-scale deep features; Using heart rate, blood oxygen saturation, exercise steps, activity intensity, and sleep duration as nodes, a direct linear non-Gaussian causal discovery algorithm is used to learn the causal direction and initial strength between nodes from historical data. The text semantic embedding vector corresponding to the text health record is input into the causal strength adjustment network, and the correction coefficients of each causal edge are output. The initial strength is weighted and adjusted to obtain the corrected causal strength, and the causal relationship graph is constructed. The causal relationship graph and the multi-scale deep features are input into a graph attention network. The graph attention layer propagates node features along the causal edges and aggregates neighbor node information, outputting the spatiotemporal representation of each node to form the comprehensive feature matrix.

4. The method for predicting user physiological state based on multi-source health data according to claim 1, characterized in that, The step of extracting long-term behavioral feature sequences within a preset time window from the comprehensive feature matrix, quantifying the drift of long-term behavioral patterns on the resting heart rate baseline using the causal path from behavioral nodes to resting heart rate nodes in the causal relationship graph, and applying context-weighted correction to the drift using the semantic features of the text health records includes the following steps: Using the number of steps and activity intensity as behavioral nodes, the daily average number of steps and daily average activity intensity for multiple consecutive days are extracted from the comprehensive feature matrix to form the long-term behavioral feature sequence; In the causal relationship graph, the causal path from the behavior node to the resting heart rate node is extracted; the long-term behavior feature sequence is input into a pre-trained causal effect prediction network, which takes the change in behavior feature as input and the change in the resting heart rate baseline as output, and obtains the initial drift amount through forward propagation; The text semantic embedding vector corresponding to the text health record is input into a gating network, which is a single-layer fully connected network followed by a sigmoid activation function to generate weighting coefficients corresponding to each time point of the initial drift. The initial drift is multiplied element by element by the weighting coefficients to obtain the corrected drift.

5. The method for predicting user physiological state based on multi-source health data according to claim 1, characterized in that, The process of removing the drift from the total fluctuation between the current observed value of resting heart rate and the historical baseline value to obtain the pathological offset after removing behavioral baseline drift includes the following steps: The difference between the observed resting heart rate at the current time point and the historical baseline value is calculated as the total fluctuation; the historical baseline value is the average resting heart rate under similar behavioral patterns during the same period within a preset number of days in the past. Subtract the drift from the total fluctuation to obtain the residual value. Take the absolute value of the residual value as the pathological offset and determine the offset direction based on the sign of the residual value.

6. The method for predicting user physiological state based on multi-source health data according to claim 1, characterized in that, When the pathological deviation exceeds a preset threshold, a pathological deviation event is determined to have occurred, and an abnormality assessment baseline is constructed based on the cumulative degree of the pathological deviation, including the following steps: A pathological offset threshold is set, which is adaptively updated based on the preset upper quantile of the user's historical pathological offset sequence; when the pathological offset at multiple consecutive sampling time points exceeds the threshold, a pathological offset event is determined to have occurred. The pathological offset values ​​at each sampling time point during the pathological offset event are weighted and summed to obtain the cumulative offset value, which is then used as the anomaly assessment baseline.

7. The method for predicting user physiological state based on multi-source health data according to claim 1, characterized in that, The process involves inputting the integrated feature matrix and the anomaly assessment baseline into a multi-task sequence coding network, where a shared encoder extracts temporal representations. A dedicated task branch then performs parallel future resting heart rate regression prediction and abnormal event warning based on the deviation between the temporal representations and the anomaly assessment baseline, outputting physiological state prediction results. This includes the following steps: The shared encoder uses a bidirectional long short-term memory network to encode the comprehensive feature matrix, thereby obtaining a shared temporal representation vector; The first task branch concatenates the shared temporal representation vector with the anomaly assessment baseline and inputs it into the fully connected regression layer, outputting the predicted resting heart rate value at a future preset time point; The second task branch calculates the Euclidean distance between the shared temporal representation vector and the anomaly assessment baseline. First, the anomaly assessment baseline is copied and expanded to the same dimension as the shared temporal representation vector, and both are L2 normalized. Then, the Euclidean distance between the normalized vectors is calculated. The dynamic threshold is adaptively updated based on the preset lower quantile of the user's historical Euclidean distance sequence. When the Euclidean distance is lower than the dynamic threshold, an anomaly warning signal is generated. The third task branch inputs the shared temporal representation vector into the fully connected regression layer and outputs the predicted blood oxygen saturation value at a future preset time point. The predicted resting heart rate, the predicted blood oxygen saturation, and the abnormal warning signal are collectively output as the physiological state prediction result.

8. The method for predicting user physiological state based on multi-source health data according to claim 5, characterized in that, The determination of the historical baseline value includes the following steps: The long-term behavioral feature sequence is subjected to trend detection to determine whether the user has entered a behavioral pattern transition period. The trend detection adopts the Mann-Kendall trend test. When the absolute value of the test statistic exceeds the critical value corresponding to the preset significance level, it is determined that the user has entered a behavioral pattern transition period. When the duration of the behavioral pattern transition exceeds a preset physiological adaptation threshold, the user is determined to be in a physiological adaptation state; the physiological adaptation threshold is set to 21 days based on population statistics. Under the physiological adaptation state, the historical baseline value is frozen, and the steady-state baseline before the behavioral pattern transition period is used as the historical baseline value. After the behavioral pattern transition period ends, when the coefficient of variation of the resting heart rate is lower than the preset coefficient of variation threshold within a 7-day sliding window, it is determined that the variability of the resting heart rate has returned to steady state conditions, the freeze is lifted, and the historical baseline value is recalculated based on the new steady state data after the behavioral pattern transition period.

9. A method for predicting user physiological state based on multi-source health data according to claim 4, characterized in that, The output of the causal effect prediction network also undergoes the following verification and correction steps: When a pathological offset event is determined to have occurred based on the pathological offset amount, the behavioral feature sequence and resting heart rate observation sequence within the preset backtracking window before the event occurs, as well as the semantic features of the text health record for the corresponding time period, are extracted. Using the starting time point of the backtracking window as the intervention point, and utilizing the causal relationship diagram and structural causal model, a counterfactual resting heart rate trajectory is generated under the assumption that the behavioral pattern is maintained before the intervention point. Calculate the cumulative deviation between the resting heart rate observation sequence and the counterfactual resting heart rate trajectory, and compare the cumulative deviation with the drift amount to obtain the drift amount estimation residual; In response to the drift estimation residual exceeding the preset tolerance, an instant correction sample is constructed using the change in behavioral features within the backtracking window and the cumulative deviation, and the parameters of the causal effect prediction network are updated step by step. The updated causal effect prediction network parameters are synchronized to the drift quantization step for drift estimation at subsequent time points.

10. A user physiological state prediction system based on multi-source health data, used to execute the user physiological state prediction method based on multi-source health data as described in any one of claims 1 to 9, characterized in that, include: The acquisition module (1) is configured to acquire multi-source heterogeneous health data of users, perform semantic alignment and cross-modal fusion on time-series numerical physiological data and text health records, and generate a semantically enhanced structured health data sequence. The matrix enhancement module (2) is configured to perform multi-scale deep feature extraction on the structured health data sequence and simultaneously construct a causal relationship graph with physiological indicators as nodes; based on the causal relationship graph, the spatiotemporal features of the nodes are propagated and aggregated through a graph neural network to obtain a comprehensive feature matrix with enhanced causal structure. Extraction module (3) is configured to extract long-term behavioral feature sequences within a preset time window from the comprehensive feature matrix, and quantify the drift of long-term behavioral patterns on the resting heart rate baseline by using the causal path from behavioral nodes to resting heart rate nodes in the causal relationship graph. The drift amount is corrected using context-weighted methods based on the semantic features of the text health records. The baseline construction module (4) is configured to remove the corrected drift from the total fluctuation between the current observed value of resting heart rate and the historical baseline value to obtain the pathological deviation after stripping the behavioral baseline drift; when the pathological deviation exceeds a preset threshold, a pathological deviation event is determined to have occurred, and an abnormal assessment baseline is constructed based on the cumulative degree of the pathological deviation. The prediction module (5) is configured to input the comprehensive feature matrix and the anomaly assessment baseline into a multi-task sequence coding network, extract temporal representations by a shared encoder, and execute future resting heart rate regression prediction and abnormal event warning in parallel by a dedicated task branch based on the degree of deviation between the temporal representations and the anomaly assessment baseline, and output physiological state prediction results.

Images