Depression disorder risk assessment method incorporating heart rate variability and behavioral data

Abstract

The application discloses a depression disorder risk assessment method combining heart rate variability and behavior data, relates to the field of digital mental health, and comprises the following steps: continuously collecting heart rate variability and multidimensional behavior data of a user through a wearable device, fusing the data into a unified feature vector after cleaning, feature extraction and standardization; constructing a depression risk assessment model based on a random forest through weight training and feature screening, optimizing and explaining feature importance by using cross-validation and SHAP value analysis, dividing a risk level according to a risk probability output by the model, and providing an abnormal prompt.The application has the advantages that HRV data and multidimensional behavior data are collected by combining an ECG device and a smart watch, the data are preprocessed and feature screened, and a depression disorder four-level risk assessment is realized based on a random forest model, so that data reliability, assessment accuracy and clinical operability are achieved.

Description

Technical Field

[0001] This invention relates to the field of digital mental health, and in particular to a method for assessing the risk of depressive disorders by combining heart rate variability and behavioral data. Background Technology

[0002] Depressive disorders are a major global public health problem, and early identification and risk assessment are crucial for improving prognosis. Traditional diagnosis relies primarily on subjective clinical interviews, which are subject to delays and biases. In recent years, the widespread adoption of wearable devices and mobile health technologies has made it possible to continuously and objectively collect multimodal physiological and behavioral data. Among these, heart rate variability, as a core indicator reflecting autonomic nervous system function, has been confirmed by numerous studies to be closely related to mood regulation disorders and abnormal stress responses in patients with depressive disorders. Simultaneously, behavioral data such as an individual's daily activity patterns, social frequency, and sleep rhythms can indirectly reflect their mood state and motivation levels.

[0003] Most current market-based methods for assessing the risk of depressive disorders rely on a single data source, or only collect HRV (Heart Rate Value) physiological data without behavioral support, or rely solely on scale scores and behavioral data while ignoring the correlation with physiological indicators. These methods fail to comprehensively reflect the complex pathogenesis of depressive disorders. Some methods use limited data collection equipment, lack dual validation with ECG and PPG data, resulting in significant data noise and bias. Furthermore, their preprocessing procedures are simplistic, failing to effectively remove outliers and properly impute missing values. Feature extraction lacks systematicity, failing to standardize and weight features, and redundant features are not effectively removed. Model construction often employs simple algorithms without cross-validation to optimize parameters, leading to insufficient assessment accuracy. In addition, most methods lack interpretable analysis, have coarse risk level classifications, fail to identify key influencing factors, and lack anomaly alerts, resulting in poor practicality and clinical applicability, failing to meet the needs for early, accurate warnings and clinical auxiliary assessments. Summary of the Invention

[0004] To improve existing methods, this paper proposes a risk assessment method for depressive disorders that combines heart rate variability (HRV) and behavioral data. This method uses ECG devices and smartwatches to collect HRV data and multi-dimensional behavioral data. After scientific preprocessing and feature screening, it uses a random forest model to achieve a four-level risk assessment of depressive disorders, which combines data reliability, assessment accuracy, and clinical operability.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: Methods for assessing the risk of depressive disorders that combine heart rate variability and behavioral data include: The raw heart rate variability data and multi-dimensional behavioral data of the subjects to be evaluated are collected in real time using ECG devices and smartwatches. All collected data are associated with the unique identifier of the subjects to be evaluated and the collection timestamp. Denoising, outlier removal, and missing value imputation were performed on HRV data and behavioral data respectively. Interference was eliminated through filtering and thresholding, data bias was calibrated, and long-term missing data were marked. Three types of features—time domain, frequency domain, and nonlinearity—are extracted from the preprocessed HRV data to obtain the physiological meaning of each feature. All features are then standardized to unify their dimensions. Five categories of behavioral features were extracted from the preprocessed behavioral data: sleep, exercise, social interaction, mobile phone use, and scale scores. Key indicators for each type of behavior were statistically analyzed and standardized to be consistent with the HRV feature dimensions. The weights of each feature are determined by training with clinical samples. The HRV features and behavioral features are weighted and summed to obtain a unified evaluation feature vector. Missing features are interpolated and labeled. A combination of analysis of variance and recursive feature elimination was used to remove redundant and irrelevant features and screen out the core features that have a significant impact on depression assessment. Based on the selected core features, a risk assessment model for depressive disorders based on the random forest algorithm was constructed. The training set and the test set were divided, the parameters were optimized through cross-validation, and the SHAP value analysis method was used to obtain the degree of influence of each core feature on the risk assessment of depressive disorders. The core feature vector of the subject to be assessed is input into the trained depressive disorder risk assessment model, which outputs the risk prediction probability, classifies it into four risk levels based on clinical criteria, and outputs abnormality prompts.

[0006] Preferably, the step of collecting raw heart rate variability data and multi-dimensional behavioral data of the subject under evaluation in real time through ECG devices and smartwatches, with all collected data associated with the unique identifier of the subject under evaluation and the collection timestamp, specifically includes: The ECG device and smartwatch were used to collect raw HRV data in collaboration, and the collection timestamp was recorded synchronously to perform dual verification and complementarity of ECG and PPG data. The raw data of the behavior of the subject to be assessed was collected by filling in the terminal, wearable device and standardized scale. The standardized scale used was the PHQ-9 depression screening scale. The subject to be assessed filled in the scale at a fixed time every day, and the daily score and the response to each item were collected. All collected data is associated with the unique identifier of the object to be evaluated and the collection timestamp.

[0007] Preferably, the steps of denoising, outlier removal, and missing value imputation for HRV data and behavioral data, eliminating interference through filtering and thresholding, calibrating data bias, and marking long-term missing data specifically include: HRV raw data preprocessing includes ECG data processing and PPG data processing; The ECG data was filtered using a low-pass method to remove high-frequency noise, and an adaptive filtering method was used to eliminate motion artifacts from interfering with the ECG signal. The peak value of the R wave in the ECG was detected by the Pan-Tompkins algorithm, and the RR interval sequence between adjacent R waves was extracted. Abnormal RR intervals caused by ectopic beats and noise were removed. For missing RR interval data, cubic spline interpolation was used to fill in the gaps. PPG data was subjected to adaptive filtering to eliminate motion artifacts and ambient light interference. The pulse interval sequence was extracted and calibrated with the RR interval sequence extracted from ECG data. PPI data with a deviation greater than 50ms were removed. During the preprocessing of raw behavioral data, outliers are removed from continuous data using the 3σ principle, while incomplete or logically contradictory entries are removed from discrete data. Missing values ​​are filled using the nearest neighbor mean imputation method.

[0008] Preferably, the step of extracting time-domain, frequency-domain, and nonlinear features from the preprocessed HRV data, obtaining the physiological meaning of each feature, and standardizing all features to unify their dimensions specifically includes: Temporal feature extraction includes the standard deviation of all RR intervals, the root mean square of the difference between adjacent RR intervals, and the proportion of adjacent RR intervals with a difference greater than 50ms. Frequency domain feature extraction uses the Fourier transform method to convert the RR interval sequence into the frequency domain, divide it into low-frequency intervals and high-frequency intervals, and calculate the power values ​​and LF / HF ratios of the two intervals respectively. Nonlinear feature extraction includes sample entropy and SD1 / SD2. Sample entropy is obtained by measuring the complexity of the RR interval sequence. SD1 / SD2 is obtained by plotting a Poincaré scatter plot and extracting the standard deviation of the short axis and the long axis in the scatter plot. All HRV features are standardized and uniformly mapped to the [0,1] interval.

[0009] Preferably, the step of classifying and extracting five categories of behavioral features from the preprocessed behavioral data—sleep, exercise, social interaction, mobile phone use, and scale scores—and statistically analyzing key indicators for each type of behavior, and then standardizing them to match the HRV feature dimensions, specifically includes: Sleep feature extraction includes average daily sleep duration, sleep latency, percentage of deep sleep, and number of sleep awakenings, which are obtained through statistical analysis of sleep stage data collected by wearable devices. Motion feature extraction includes daily average exercise duration, exercise frequency, and exercise duration, which are obtained through statistical calculation of motion data collected by an accelerometer. Social feature extraction includes the average number of social interactions per day, social duration, and number of social contacts, which are obtained through statistics of communication and social data collected from the terminal. Mobile phone usage features include average daily usage time, percentage of time spent using entertainment apps, percentage of time spent using information apps, and unlocking frequency, which are obtained through statistical analysis of application usage data collected from the terminal. The scale scoring features were extracted, including the daily total score of PHQ-9, the mean score of core symptom items, and the score of sleep-related items, which were calculated from the scale data filled in by the subjects to be assessed. All extracted behavioral features are standardized and uniformly mapped to the [0,1] interval, consistent with the HRV feature dimension.

[0010] Preferably, the step of determining the weights of each feature through clinical sample training, weighted summing of HRV features and behavioral features to obtain a unified evaluation feature vector, and interpolating and labeling missing features specifically includes: The weights of each feature were determined by training with clinical samples and using feature importance analysis. Among the HRV features, the weights of HF value, RMSSD, and SDNN were 0.25, 0.20, and 0.15, respectively. Among the behavioral features, the weights of sleep duration, PHQ-9 total score, social interaction frequency, exercise frequency, and mobile entertainment APP usage percentage were 0.05. Multiply the standardized value of each feature by its corresponding weight and sum them to obtain the fused feature vector of a single object to be evaluated; For features with missing data, the population mean of the feature is multiplied by the corresponding weight and included in the fused feature vector, and a missing data marker is added.

[0011] Preferably, the method of combining analysis of variance with recursive feature elimination to remove redundant and irrelevant features and screen out the core features that have a significant impact on depression assessment specifically includes: Feature selection was performed using a combination of analysis of variance and recursive feature elimination. The correlation between each fusion feature and the diagnostic label of depressive disorder was calculated by analysis of variance, and features with a correlation coefficient greater than 0.3 were retained. A recursive feature elimination method was adopted, based on the random forest model, to gradually eliminate features whose contribution to the model prediction was less than 0.05, and to select the core features.

[0012] Preferably, the step of constructing a depressive disorder risk assessment model based on the selected core features using the random forest algorithm, dividing the model into training and test sets, optimizing parameters through cross-validation, and using the SHAP value analysis method to obtain the degree of influence of each core feature on the depressive disorder risk assessment specifically includes: A risk assessment model for depressive disorder was constructed using the random forest algorithm. The model consisted of 150 decision trees with a depth of 8 layers. The core feature data after screening is divided into training set and test set in a 7:3 ratio, where training set is used for model training and test set is used for model performance verification. During training, core feature data from clinically diagnosed patients with depressive disorders and healthy individuals were used as training samples. The model parameters were optimized through 5-fold cross-validation, and the number, depth, and splitting threshold of decision trees were adjusted. After training, the SHAP value analysis method was used to obtain the degree of influence of each core feature on the risk assessment of depressive disorders.

[0013] Preferably, the step of inputting the core feature vector of the subject to be assessed into the trained depressive disorder risk assessment model, outputting the risk prediction probability, classifying it into four risk levels based on clinical criteria, and outputting abnormal indications specifically includes: The core feature vectors of the subjects to be evaluated are input into the depression disorder risk assessment model, and the model outputs the predicted probability of depression disorder risk for the subjects to be evaluated. Based on the predicted probability combined with clinical diagnostic criteria, the risk of depressive disorders is divided into four levels: a predicted probability of 0-0.2 indicates low risk with no obvious depressive symptoms; a predicted probability of 0.2-0.5 indicates medium risk with possible mild depressive symptoms; a predicted probability of 0.5-0.8 indicates high risk with possible moderate depressive symptoms; and a predicted probability of 0.8-1.0 indicates very high risk with possible severe depressive symptoms. Output anomaly alerts for each core feature and label the factors that lead to the risk level.

[0014] Compared with the prior art, the advantages of the present invention are: This system achieves deep integration of physiological indicators and behavioral data, balancing scientific rigor, comprehensiveness, and practicality. It collaboratively collects raw HRV data and multi-dimensional behavioral data via ECG devices and smartwatches, using timestamps and unique identifiers to ensure data traceability and dual verification to enhance data reliability. The preprocessing stage employs targeted methods to reduce noise, remove outliers, and fill in missing values, laying a high-quality data foundation for subsequent analysis. Feature extraction covers three core HRV features and five behavioral features, standardized to achieve dimensional uniformity, and feature weights determined using clinical samples. After fusion, core features are selected using scientific methods to avoid redundant interference. A model is built based on the random forest algorithm, with parameters optimized through cross-validation. The influence of features is analyzed using SHAP values ​​to improve the interpretability of the assessment. Simultaneously, it classifies four risk levels, labels key influencing factors, and outputs abnormal alerts, aligning with clinical diagnostic criteria while offering convenient practicality, enabling more accurate and comprehensive early assessment and warning of depressive disorder risk. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the method proposed in this invention; Figure 2 This is a schematic diagram of the multi-source data synchronous acquisition proposed in this invention; Figure 3 This is a schematic diagram of the data preprocessing proposed in this invention; Figure 4 This is a schematic diagram of HRV feature extraction proposed in this invention; Figure 5 This is a schematic diagram of the behavioral data feature extraction proposed in this invention; Figure 6 This is a schematic diagram of the feature fusion proposed in this invention; Figure 7 This is a schematic diagram of the feature filtering proposed in this invention; Figure 8 This is a schematic diagram illustrating the construction and training of the risk assessment model proposed in this invention; Figure 9 This is a schematic diagram illustrating the risk prediction and classification proposed in this invention. Detailed Implementation

[0016] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0017] See Figure 1 As shown, the method for assessing the risk of depressive disorders by combining heart rate variability and behavioral data includes: Step 1: Collect raw heart rate variability data and multi-dimensional behavioral data of the subject under evaluation in real time using ECG equipment and smartwatch. All collected data are associated with the unique identifier of the subject under evaluation and the collection timestamp. Step 2: Denoise reduction, outlier removal, and missing value imputation are performed on HRV data and behavioral data respectively. Interference is eliminated through filtering and thresholding, data bias is calibrated, and long-term missing data is marked. Step 3: Extract time-domain, frequency-domain, and nonlinear features from the preprocessed HRV data, obtain the physiological meaning of each feature, and standardize all features to unify their dimensions. Step 4: Extract five categories of behavioral features from the preprocessed behavioral data: sleep, exercise, social interaction, mobile phone use, and scale scores. Statistically analyze the key indicators of each type of behavior and standardize them to match the HRV feature dimensions. Step 5: Determine the weights of each feature through training with clinical samples, sum the HRV features and behavioral features with weights to obtain a unified evaluation feature vector, and perform interpolation and labeling on missing features; Step Six: Using a combination of analysis of variance and recursive feature elimination, redundant and irrelevant features are removed to identify the core features that have a significant impact on depression assessment. Step 7: Construct a risk assessment model for depressive disorders based on the selected core features using the random forest algorithm, divide the training set and the test set, optimize the parameters through cross-validation, and use the SHAP value analysis method to obtain the degree of influence of each core feature on the risk assessment of depressive disorders. Step 8: Input the core feature vector of the subject to be assessed into the trained depressive disorder risk assessment model, output the risk prediction probability, classify it into four risk levels based on clinical criteria, and output abnormality prompts.

[0018] See Figure 2 As shown, raw heart rate variability data and multi-dimensional behavioral data of the subjects under evaluation are collected in real time using ECG devices and smartwatches. All collected data are associated with the unique identifier of the subject under evaluation and the collection timestamp, specifically including: The ECG device and smartwatch were used to collect raw HRV data in collaboration, and the collection timestamp was recorded synchronously to perform dual verification and complementarity of ECG and PPG data. The raw data of the behavior of the subject to be assessed was collected by filling in the terminal, wearable device and standardized scale. The standardized scale used was the PHQ-9 depression screening scale. The subject to be assessed filled in the scale at a fixed time every day, and the daily score and the response to each item were collected. All collected data is associated with the unique identifier of the object to be evaluated and the collection timestamp.

[0019] Specifically, a patch-type long-term ECG device is used in conjunction with a smartwatch to collect raw HRV data. The patch-type long-term ECG device continuously collects ECG signals for 7-14 days and synchronously records the timestamp of each frame of signal collection. The raw behavioral data was collected collaboratively through three authorized methods: first, data was collected from the smart terminal backend authorized by the subject to be evaluated, including total daily mobile phone usage time, usage time and percentage of various types of apps, mobile phone unlocking frequency, communication records, and social app interaction data; second, data was collected from wearable device sensors, using the built-in three-axis accelerometer, heart rate sensor, and blood oxygen sensor to collect daily exercise duration, exercise intensity, and sleep data; and third, data was collected from standardized questionnaires, using the PHQ-9 depression screening scale to collect the daily total score and specific responses to each of the nine items.

[0020] See Figure 3 As shown, noise reduction, outlier removal, and missing value imputation were performed on HRV data and behavioral data, respectively. Interference was eliminated through filtering and thresholding, data bias was calibrated, and long-term missing data were marked. Specifically, these included: HRV raw data preprocessing includes ECG data processing and PPG data processing; The ECG data was filtered using a low-pass method to remove high-frequency noise, and an adaptive filtering method was used to eliminate motion artifacts from interfering with the ECG signal. The peak value of the R wave in the ECG was detected by the Pan-Tompkins algorithm, and the RR interval sequence between adjacent R waves was extracted. Abnormal RR intervals caused by ectopic beats and noise were removed. For missing RR interval data, cubic spline interpolation was used to fill in the gaps. PPG data was subjected to adaptive filtering to eliminate motion artifacts and ambient light interference. The pulse interval sequence was extracted and calibrated with the RR interval sequence extracted from ECG data. PPI data with a deviation greater than 50ms were removed. During the preprocessing of raw behavioral data, outliers are removed from continuous data using the 3σ principle, while incomplete or logically contradictory entries are removed from discrete data. Missing values ​​are filled using the nearest neighbor mean imputation method.

[0021] Specifically, ECG data denoising employs a two-step filtering method. A Butterworth-type 5Hz low-pass filter is used to filter out high-frequency electromagnetic interference above 5Hz. The filter order is set to 4th order. An adaptive filtering method is used to eliminate motion artifact interference. The acceleration signal collected by the wearable device is used as a reference signal. The filter coefficient is dynamically adjusted using a minimum mean square error algorithm to compensate for ECG signal distortion caused by physical activity in real time. R-wave peak detection uses the Pan-Tompkins algorithm. Differential and squaring operations are performed on the filtered ECG signal to enhance R-wave characteristics. A dynamic threshold is set to accurately identify the R-wave peak and extract the RR interval sequence between adjacent R waves. Abnormal RR interval removal uses a ±20% mean threshold method. The mean of all RR intervals is calculated, and RR intervals exceeding ±20% of the mean are judged as abnormal and removed. After removal, a 4Hz cubic spline interpolation method is used to fill in the missing RR intervals. Using three valid RR intervals before and after the missing data as nodes, a continuous RR interval sequence is generated through interpolation. PPG data preprocessing focuses on eliminating motion artifacts and ambient light interference: Adaptive filtering combined with infrared filtering is used. The ambient light intensity is detected by the device's built-in ambient light sensor, and filtering parameters are dynamically adjusted to filter out ambient light interference with the pulse wave signal. Motion artifact elimination employs acceleration-assisted correction, synchronously analyzing the acceleration signal and pulse wave signal, discarding pulse wave data collected when the acceleration is greater than 0.5g, and then extracting the pulse interval sequence. The PPI sequence and the RR interval sequence extracted from the ECG data are timestamp aligned and calibrated. The difference between the two at each time point is calculated, discarding PPI data with a deviation greater than 50ms, and retaining valid data with a deviation within 50ms.

[0022] See Figure 4As shown, time-domain, frequency-domain, and nonlinear features are extracted from the preprocessed HRV data to obtain the physiological meaning of each feature. All features are standardized to unify dimensions, specifically including: Temporal feature extraction includes the standard deviation of all RR intervals, the root mean square of the difference between adjacent RR intervals, and the proportion of adjacent RR intervals with a difference greater than 50ms. Frequency domain feature extraction uses the Fourier transform method to convert the RR interval sequence into the frequency domain, divide it into low-frequency intervals and high-frequency intervals, and calculate the power values ​​and LF / HF ratios of the two intervals respectively. Nonlinear feature extraction includes sample entropy and SD1 / SD2. Sample entropy is obtained by measuring the complexity of the RR interval sequence. SD1 / SD2 is obtained by plotting a Poincaré scatter plot and extracting the standard deviation of the short axis and the long axis in the scatter plot. All HRV features are standardized and uniformly mapped to the [0,1] interval.

[0023] Specifically, the temporal feature extraction focuses on the temporal distribution characteristics of HRV. It uses a preprocessed and calibrated RR interval sequence to extract three core temporal features: SDNN, RMSSD, and pNN50. SDNN is obtained by statistically analyzing the dispersion of all valid RR intervals. First, the mean of all RR intervals is calculated. Then, the deviation of each RR interval from the mean is calculated, the deviations are squared, and the average is calculated. Finally, the square root of the average is taken to obtain the SDNN value. RMSSD is achieved by calculating the root mean square of the differences between adjacent RR intervals. First, the difference between every two adjacent RR intervals is calculated. Then, all differences are squared, the average of the squared values ​​is calculated, and the square root is taken. pNN50 is obtained by statistically analyzing the proportion of outlier differences. First, the number of adjacent RR interval differences greater than 50ms is selected, and then this number is divided by the total number of RR intervals to obtain the proportion. Frequency domain feature extraction employs Fourier transform to convert the time-domain RR interval sequence into a frequency-domain signal. Zero-padding is applied to the RR interval sequence to ensure that the sequence length is an integer power of 2. Fast Fourier transform is then used to convert the RR interval sequence into a frequency-power spectral density distribution, dividing it into two core frequency intervals. The power values ​​for each interval are calculated separately. The total power for each interval is obtained by integrating the power spectral density within each interval. The LF / HF ratio is obtained by dividing the LF power value by the HF power value. LF reflects the combined activity of the sympathetic and parasympathetic nervous systems, while HF reflects parasympathetic activity. An abnormally high LF / HF ratio indicates sympathetic dominance. Nonlinear feature extraction focuses on the complexity and dynamic changes of the RR interval sequence. For sample entropy extraction, an embedding dimension of 2 and a similarity tolerance of 0.2 times the standard deviation of the RR interval sequence are first set. The RR interval sequence is then divided into multiple vectors of length 2. The similarity between each vector and other vectors is calculated, and the proportion of similar vectors is statistically analyzed. The sample entropy value is obtained through logarithmic operation; a lower value indicates stronger regularity in the RR interval sequence, suggesting a decrease in the flexibility of heart rate regulation. For SD1 / SD2 extraction, a Poincaré scatter plot is first drawn, with each RR interval as the horizontal axis and the next adjacent RR interval as the vertical axis, plotting all RR interval data points on a coordinate system. Then, the least squares method is used to fit an elliptical distribution of the scatter points, extracting the standard deviation of the minor axis as SD1 and the standard deviation of the major axis as SD2. The ratio of SD1 to SD2 is calculated to reflect the dynamic regulation characteristics of HRV. Where SD1 is the standard deviation of the minor axis of the ellipse, SD2 is the standard deviation of the major axis of the ellipse, SD(RR) is the standard deviation of the RR interval series, and △RR = RR i+1 -RR i is the difference between adjacent RR intervals, and SD(△RR) is the standard deviation of the difference sequence.

[0024] See Figure 5 As shown, five behavioral features—sleep, exercise, social interaction, mobile phone use, and scale scores—were extracted from the preprocessed behavioral data. Key indicators for each type of behavior were statistically analyzed and standardized to match the HRV feature dimensions. Specifically, these included: Sleep feature extraction includes average daily sleep duration, sleep latency, percentage of deep sleep, and number of sleep awakenings, which are obtained through statistical analysis of sleep stage data collected by wearable devices. Motion feature extraction includes daily average exercise duration, exercise frequency, and exercise duration, which are obtained through statistical calculation of motion data collected by an accelerometer. Social feature extraction includes the average number of social interactions per day, social duration, and number of social contacts, which are obtained through statistics of communication and social data collected from the terminal. Mobile phone usage features include average daily usage time, percentage of time spent using entertainment apps, percentage of time spent using information apps, and unlocking frequency, which are obtained through statistical analysis of application usage data collected from the terminal. The scale scoring features were extracted, including the daily total score of PHQ-9, the mean score of core symptom items, and the score of sleep-related items, which were calculated from the scale data filled in by the subjects to be assessed. All extracted behavioral features are standardized and uniformly mapped to the [0,1] interval, consistent with the HRV feature dimension.

[0025] Specifically, the scale scoring features were extracted based on the preprocessed PHQ-9 scale data: the daily total score was directly extracted from the total score filled in by the subject on that day; the mean scores of the core symptom items were extracted from the scores of item 1 (depressed mood) and item 2 (loss of interest), and the average of the two was calculated; the sleep-related item scores were extracted from item 3 (sleep disorder), and the self-blame-related item scores were extracted from item 6 (self-blame); all scale-related features retained their original scores.

[0026] See Figure 6 As shown, the weights of each feature are determined through training with clinical samples. The HRV features and behavioral features are weighted and summed to obtain a unified evaluation feature vector. Missing features are then interpolated and labeled, specifically including: The weights of each feature were determined by training with clinical samples and using feature importance analysis. Among the HRV features, the weights of HF value, RMSSD, and SDNN were 0.25, 0.20, and 0.15, respectively. Among the behavioral features, the weights of sleep duration, PHQ-9 total score, social interaction frequency, exercise frequency, and mobile entertainment APP usage percentage were 0.05. Multiply the standardized value of each feature by its corresponding weight and sum them to obtain the fused feature vector of a single object to be evaluated; For features with missing data, the population mean of the feature is multiplied by the corresponding weight and included in the fused feature vector, and a missing data marker is added.

[0027] Specifically, See Figure 7 As shown, by combining analysis of variance with recursive feature elimination, redundant and irrelevant features were removed, and the core features that significantly influenced depression assessment were selected, including: Feature selection was performed using a combination of analysis of variance and recursive feature elimination. The correlation between each fusion feature and the diagnostic label of depressive disorder was calculated by analysis of variance, and features with a correlation coefficient greater than 0.3 were retained. A recursive feature elimination method was adopted, based on the random forest model, to gradually eliminate features whose contribution to the model prediction was less than 0.05, and to select the core features.

[0028] See Figure 8 As shown, a risk assessment model for depressive disorders based on the selected core features was constructed using the random forest algorithm. The model was divided into training and test sets, and parameters were optimized through cross-validation. The SHAP value analysis method was used to obtain the degree of influence of each core feature on the risk assessment of depressive disorders, specifically including: A risk assessment model for depressive disorder was constructed using the random forest algorithm. The model consisted of 150 decision trees with a depth of 8 layers. The core feature data after screening is divided into training set and test set in a 7:3 ratio, where training set is used for model training and test set is used for model performance verification. During training, core feature data from clinically diagnosed patients with depressive disorders and healthy individuals were used as training samples. The model parameters were optimized through 5-fold cross-validation, and the number, depth, and splitting threshold of decision trees were adjusted. After training, the SHAP value analysis method was used to obtain the degree of influence of each core feature on the risk assessment of depressive disorders.

[0029] Specifically, 500 clinically diagnosed patients with depressive disorders and 500 healthy individuals were selected as training samples. Diagnosed patients needed to meet the DSM-5 diagnostic criteria for depressive disorders, excluding cases with co-existing cardiovascular diseases, schizophrenia, or other illnesses. Healthy individuals needed to have no history of mental illness, no serious physical illness, and a PHQ-9 score ≤ 4. HRV and behavioral characteristics were extracted from all samples. Random forest feature importance analysis was used to determine the weight of each feature. All sample features were input into the random forest model. The reduction in the Gini coefficient of each feature was calculated to measure its discriminative contribution to the depressive disorder diagnostic label; the higher the contribution, the greater the weight. Based on clinical experience, the weight allocation was adjusted, and the final weights of the core features were determined as follows: Among the HRV features, the weight of HF value was 0.25, the weight of RMSSD was 0.20, and the weight of SDNN was 0.15; among the behavioral features, the weight of sleep duration was 0.12, the weight of PHQ-9 total score was 0.10, the weight of social interaction frequency was 0.08, the weight of exercise frequency was 0.05, and the weight of mobile entertainment APP usage ratio was 0.05. The total weight of all features was 1.0. For a single object to be evaluated, the standardized HRV features and behavioral features are called, and the standardized value of each feature is multiplied by its corresponding weight to obtain the weighted score of each feature. To avoid numerical deviation of the fusion vector due to excessively high weighted scores for some features, a min-max normalization method is used to process the fused feature vector a second time. The value of the fused vector is subtracted from the minimum value of the fused vector of all objects to be evaluated, and then divided by the difference between the maximum and minimum values ​​of the fused vector of all objects to be evaluated, to ensure that the value range of the final fused feature vector is strictly within the range of [0,1].

[0030] See Figure 9 As shown, the core feature vector of the subject to be assessed is input into a trained depressive disorder risk assessment model, which outputs the risk prediction probability. This probability is then categorized into four risk levels based on clinical criteria, and abnormal alerts are output, specifically including: The core feature vectors of the subjects to be evaluated are input into the depression disorder risk assessment model, and the model outputs the predicted probability of depression disorder risk for the subjects to be evaluated. Based on the predicted probability combined with clinical diagnostic criteria, the risk of depressive disorders is divided into four levels: a predicted probability of 0-0.2 indicates low risk with no obvious depressive symptoms; a predicted probability of 0.2-0.5 indicates medium risk with possible mild depressive symptoms; a predicted probability of 0.5-0.8 indicates high risk with possible moderate depressive symptoms; and a predicted probability of 0.8-1.0 indicates very high risk with possible severe depressive symptoms. Output anomaly alerts for each core feature and label the factors that lead to the risk level.

[0031] Specifically, analysis of variance (ANOVA) was used for initial screening. The core purpose was to eliminate features irrelevant to the diagnosis of depressive disorder, obtain fusion feature vectors for all subjects to be evaluated, and associate them with corresponding clinical diagnostic labels to ensure that each fusion feature vector has a clear diagnostic label. For each fusion feature, the mean and variance of the feature were calculated for the diagnosed group and the healthy group, respectively. The F-statistic for each feature was obtained using the ANOVA formula, and the corresponding correlation coefficient was calculated based on the F-statistic. The correlation coefficient threshold was set at 0.3, and features with correlation coefficients greater than 0.3 were retained, while features with correlation coefficients ≤ 0.3 were eliminated. After initial screening, a recursive feature elimination method is used for secondary screening to further remove redundant features. A feature evaluation model is built based on the random forest model, initially set with 100 decision trees and a depth of 6 layers to avoid overfitting and affecting the accuracy of feature evaluation. The feature data and diagnostic labels after initial screening are input into the model for the first training. After training, the contribution of each feature to the model prediction result is measured by calculating the reduction of the Gini coefficient. A contribution threshold of 0.05 is set, and features with a contribution value lower than 0.05 are marked as redundant features and removed. Then, the random forest model is retrained based on the remaining features, and the process of "training-calculating contribution value-removing low contribution features" is repeated, removing 1-2 low contribution features each time, until the number of remaining features reaches 10-15, and the contribution value of all remaining features is ≥0.05, at which point the screening stops.

[0032] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0033] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0034] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method of depression disorder risk assessment combining heart rate variability and behavioral data, characterized by, include: The raw heart rate variability data and multi-dimensional behavioral data of the subjects to be evaluated are collected in real time using ECG devices and smartwatches. All collected data are associated with the unique identifier of the subjects to be evaluated and the collection timestamp. Denoising, outlier removal, and missing value imputation were performed on HRV data and behavioral data respectively. Interference was eliminated through filtering and thresholding, data bias was calibrated, and long-term missing data were marked. Three types of features—time domain, frequency domain, and nonlinearity—are extracted from the preprocessed HRV data to obtain the physiological meaning of each feature. All features are then standardized to unify their dimensions. Five categories of behavioral features were extracted from the preprocessed behavioral data: sleep, exercise, social interaction, mobile phone use, and scale scores. Key indicators for each type of behavior were statistically analyzed and standardized to be consistent with the HRV feature dimensions. The weights of each feature are determined by training with clinical samples. The HRV features and behavioral features are weighted and summed to obtain a unified evaluation feature vector. Missing features are interpolated and labeled. A combination of analysis of variance and recursive feature elimination was used to remove redundant and irrelevant features and screen out the core features that have a significant impact on depression assessment. Based on the selected core features, a risk assessment model for depressive disorders based on the random forest algorithm was constructed. The training set and the test set were divided, the parameters were optimized through cross-validation, and the SHAP value analysis method was used to obtain the degree of influence of each core feature on the risk assessment of depressive disorders. The core feature vector of the subject to be assessed is input into the trained depressive disorder risk assessment model, which outputs the risk prediction probability, classifies it into four risk levels based on clinical criteria, and outputs abnormality prompts.

2. The method of claim 1, wherein the method is used to assess the risk of depressive disorder by combining heart rate variability and behavioral data. The process of collecting raw heart rate variability data and multi-dimensional behavioral data of the subject under evaluation in real time through ECG devices and smartwatches, with all collected data linked to the unique identifier of the subject under evaluation and the collection timestamp, specifically includes: The ECG device and smartwatch were used to collect raw HRV data in collaboration, and the collection timestamp was recorded synchronously to perform dual verification and complementarity of ECG and PPG data. The raw data of the behavior of the subject to be assessed was collected by filling in the terminal, wearable device and standardized scale. The standardized scale used was the PHQ-9 depression screening scale. The subject to be assessed filled in the scale at a fixed time every day, and the daily score and the response to each item were collected. All collected data is associated with the unique identifier of the object to be evaluated and the collection timestamp.

3. The method for assessing the risk of depressive disorders by combining heart rate variability and behavioral data according to claim 1, characterized in that, The specific steps of denoising, outlier removal, and missing value imputation for HRV and behavioral data, eliminating interference through filtering and thresholding, calibrating data bias, and marking long-term missing data include: HRV raw data preprocessing includes ECG data processing and PPG data processing; The ECG data was filtered using a low-pass method to remove high-frequency noise, and an adaptive filtering method was used to eliminate motion artifacts from interfering with the ECG signal. The peak value of the R wave in the ECG was detected by the Pan-Tompkins algorithm, and the RR interval sequence between adjacent R waves was extracted. Abnormal RR intervals caused by ectopic beats and noise were removed. For missing RR interval data, cubic spline interpolation was used to fill in the gaps. PPG data was subjected to adaptive filtering to eliminate motion artifacts and ambient light interference. The pulse interval sequence was extracted and calibrated with the RR interval sequence extracted from ECG data. PPI data with a deviation greater than 50ms were removed. During the preprocessing of raw behavioral data, outliers are removed from continuous data using the 3σ principle, while incomplete or logically contradictory entries are removed from discrete data. Missing values ​​are filled using the nearest neighbor mean imputation method.

4. The method for assessing the risk of depressive disorders by combining heart rate variability and behavioral data according to claim 1, characterized in that, The process of extracting time-domain, frequency-domain, and nonlinear features from preprocessed HRV data, obtaining the physiological meaning of each feature, and standardizing all features to unify their dimensions specifically includes: Temporal feature extraction includes the standard deviation of all RR intervals, the root mean square of the difference between adjacent RR intervals, and the proportion of adjacent RR intervals with a difference greater than 50ms. Frequency domain feature extraction uses the Fourier transform method to convert the RR interval sequence into the frequency domain, divide it into low-frequency intervals and high-frequency intervals, and calculate the power values ​​and LF / HF ratios of the two intervals respectively. Nonlinear feature extraction includes sample entropy and SD1 / SD2. Sample entropy is obtained by measuring the complexity of the RR interval sequence. SD1 / SD2 is obtained by plotting a Poincaré scatter plot and extracting the standard deviation of the short axis and the long axis in the scatter plot. All HRV features are standardized and uniformly mapped to the [0,1] interval.

5. The method for assessing the risk of depressive disorders by combining heart rate variability and behavioral data according to claim 1, characterized in that, The process involves extracting five categories of behavioral features from the preprocessed behavioral data: sleep, exercise, social interaction, mobile phone use, and scale scores. Key indicators for each behavioral category are statistically analyzed and standardized to align with the HRV feature dimensions. include: Sleep feature extraction includes average daily sleep duration, sleep latency, percentage of deep sleep, and number of sleep awakenings, which are obtained through statistical analysis of sleep stage data collected by wearable devices. Motion feature extraction includes daily average exercise duration, exercise frequency, and exercise duration, which are obtained through statistical calculation of motion data collected by an accelerometer. Social feature extraction includes the average number of social interactions per day, social duration, and number of social contacts, which are obtained through statistics of communication and social data collected from the terminal. Mobile phone usage features include average daily usage time, percentage of time spent using entertainment apps, percentage of time spent using information apps, and unlocking frequency, which are obtained through statistical analysis of application usage data collected from the terminal. The scale scoring features were extracted, including the daily total score of PHQ-9, the mean score of core symptom items, and the score of sleep-related items, which were calculated from the scale data filled in by the subjects to be assessed. All extracted behavioral features are standardized and uniformly mapped to the [0,1] interval, consistent with the HRV feature dimension.

6. The method for assessing the risk of depressive disorders by combining heart rate variability and behavioral data according to claim 1, characterized in that, The process of determining the weights of each feature through training with clinical samples, weighted summing of HRV features and behavioral features to obtain a unified evaluation feature vector, and interpolating and labeling missing features specifically includes: The weights of each feature were determined by training with clinical samples and using feature importance analysis. Among the HRV features, the weights of HF value, RMSSD, and SDNN were 0.25, 0.20, and 0.15, respectively. Among the behavioral features, the weights of sleep duration, PHQ-9 total score, social interaction frequency, exercise frequency, and mobile entertainment APP usage percentage were 0.

05. Multiply the standardized value of each feature by its corresponding weight and sum them to obtain the fused feature vector of a single object to be evaluated; For features with missing data, the population mean of the feature is multiplied by the corresponding weight and included in the fused feature vector, and a missing data marker is added.

7. The method for assessing the risk of depressive disorders by combining heart rate variability and behavioral data according to claim 1, characterized in that, The method employing a combination of analysis of variance and recursive feature elimination to remove redundant and irrelevant features and screen out the core features that have a significant impact on depression assessment specifically includes: Feature selection was performed using a combination of analysis of variance and recursive feature elimination. The correlation between each fusion feature and the diagnostic label of depressive disorder was calculated by analysis of variance, and features with a correlation coefficient greater than 0.3 were retained. A recursive feature elimination method was adopted, based on the random forest model, to gradually eliminate features whose contribution to the model prediction was less than 0.05, and to select the core features.

8. The method for assessing the risk of depressive disorders by combining heart rate variability and behavioral data according to claim 1, characterized in that, The process of constructing a risk assessment model for depressive disorders based on the selected core features, dividing the model into training and test sets, optimizing parameters through cross-validation, and using SHAP value analysis to obtain the degree of influence of each core feature on the risk assessment of depressive disorders specifically includes: A risk assessment model for depressive disorder was constructed using the random forest algorithm. The model consisted of 150 decision trees with a depth of 8 layers. The core feature data after screening is divided into training set and test set in a 7:3 ratio, where training set is used for model training and test set is used for model performance verification. During training, core feature data from clinically diagnosed patients with depressive disorders and healthy individuals were used as training samples. The model parameters were optimized through 5-fold cross-validation, and the number, depth, and splitting threshold of decision trees were adjusted. After training, the SHAP value analysis method was used to obtain the degree of influence of each core feature on the risk assessment of depressive disorders.

9. The method for assessing the risk of depressive disorders by combining heart rate variability and behavioral data according to claim 1, characterized in that, The process involves inputting the core feature vector of the subject to be assessed into a pre-trained depressive disorder risk assessment model, outputting a risk prediction probability, classifying it into four risk levels based on clinical criteria, and outputting abnormality alerts, specifically including: The core feature vectors of the subjects to be assessed are input into the depressive disorder risk assessment model. The model outputs the predicted probability of the depressive disorder risk of the subjects to be assessed. Based on the predicted probability and clinical diagnostic criteria, the depressive disorder risk is divided into four levels: a predicted probability of 0-0.2 indicates low risk with no obvious depressive symptoms; a predicted probability of 0.2-0.5 indicates medium risk with possible mild depressive symptoms; a predicted probability of 0.5-0.8 indicates high risk with possible moderate depressive symptoms; and a predicted probability of 0.8-1.0 indicates very high risk with possible severe depressive symptoms. Anomalies of each core feature are output, and the factors that lead to the risk level are marked.

Images