Personalized evaluation system for depression biofeedback result based on physiological signal

Abstract

The invention relates to an individualized evaluation system for a depression biofeedback result based on a physiological signal. The individualized evaluation system comprises a signal acquisition module for acquiring first data of an intervention group and second data of a waiting group; the feature extraction module screens out physiological signals related to depression according to the first data and the second data, and the physiological signals serve as input features of a deep learning model to be trained; the deep learning model training module is used for preprocessing each physiological signal, related to depression, of the screened intervention group, inputting the preprocessed physiological signal into a one-dimensional residual neural network (ResNet-1D) model for training, obtaining an output result and evaluating the output result; when the evaluation reaches a threshold value, obtaining a trained ResNet-1D model of which the evaluation result accords with an expected range; the ResNet-1D model is used for evaluating the physiological signal of the individual. The model constructed by the system can realize prediction of association between the biofeedback therapy and individualized curative effect prediction, the prediction result is accurate, and the use experience of individuals is improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to an individualized evaluation system for depression biofeedback results based on physiological signals. Background Art

[0002] Depression is primarily manifested by persistent low mood, loss of interest, and anhedonia, accompanied by somatic symptoms such as sleep disturbances, changes in appetite or weight, fatigue, and loss of energy. Currently, medication, psychological, and physical therapies are the main treatment options. Biofeedback intervention, which combines both physical and psychological benefits, is widely used in the clinical treatment of patients with mild to moderate depression. However, the selection of biofeedback intervention protocols is currently based primarily on the clinician's experience and is highly subjective. The effectiveness of biofeedback intervention is primarily assessed through symptom scales and clinician observation and judgment, lacking effective objective evaluation indicators. This makes it impossible to predictably and individually assess the efficacy of the intervention, resulting in the inability to promptly replace ineffective protocols and prolonging the treatment cycle.

[0003] Physiological signals can be used to objectively assess a patient's physical and emotional improvement, providing an objective basis for evaluating the efficacy of biofeedback. Studies have also found that a patient's physiological signals before intervention are closely related to the effectiveness of biofeedback training. Currently, most studies are retrospective, comparing the effects before and after biofeedback intervention to analyze the characteristics of individuals who respond to biofeedback intervention. If it were possible to predict which individuals would respond to intervention before intervention, it would reduce trial and error time, shorten the intervention cycle, and facilitate personalized, precise intervention.

[0004] Biofeedback therapy is a non-drug treatment for depression, typically lasting at least four weeks. Currently, treatment options are primarily selected based on physician experience. Evaluations of biofeedback interventions are often subjective, lacking objective efficacy indicators. Most evaluations are conducted after the intervention has concluded, which can delay treatment and increase the burden of illness.

[0005] Therefore, there is an urgent need for a method that can objectively process and predict physiological signals using computers. Summary of the Invention

[0006] (1) Technical issues to be resolved

[0007] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides an individualized assessment system for depression biofeedback results based on physiological signals, which solves the defects of the prior art such as long prediction time and non-objectivity.

[0008] (2) Technical solution

[0009] In order to achieve the above objectives, the main technical solutions adopted by the present invention include:

[0010] In a first aspect, an embodiment of the present invention provides a personalized assessment system for depression biofeedback results based on physiological signals, comprising:

[0011] a signal acquisition module for collecting first data from the intervention group and second data from the waiting group; the first data from the intervention group includes: physiological signals of the intervention group before biofeedback therapy, physiological signals within a specified time after biofeedback therapy, and scores of the intervention group on a self-rating depression scale within a specified time before and after the intervention; the second data from the waiting group includes: physiological signals of the waiting group before each biofeedback therapy, and scores of the waiting group on a self-rating depression scale within a specified time; the waiting group does not undergo biofeedback therapy;

[0012] a feature extraction module, configured to screen out physiological signals related to depression based on the first data of the intervention group and the second data of the waiting group, and use them as input features of the deep learning model to be trained;

[0013] The deep learning model training module is used to pre-process the physiological signals related to depression in the selected intervention group, input the pre-processed values ​​into the deep learning model for training, obtain output results, and evaluate the output results; when the evaluation reaches a threshold, a deep learning model is obtained whose evaluation results after training are within the expected range;

[0014] The deep learning model is used to evaluate the physiological signals of an individual.

[0015] Optionally, the individualized assessment system further includes: a prediction module;

[0016] The signal acquisition module is used to acquire physiological signals of the patient to be evaluated;

[0017] The prediction module is used to input the physiological signal of the patient to be evaluated into the trained deep learning module, obtain the results output by the deep learning model, and determine whether the biofeedback therapy is suitable for the patient to be evaluated.

[0018] Optionally, the signal acquisition module is specifically used to:

[0019] The physiological signals of each member of the intervention group were collected before and after the biofeedback therapy for 12 sessions over four weeks.

[0020] Before the biofeedback therapy, the PHQ-9 scores of each member in the intervention group were collected;

[0021] The physiological signals include one or more of the following: physical relaxation index, psychological relaxation index, frontal surface electromyography, blood volume pulse, heart rate variability, heart rate, natural logarithm of the root mean square of the difference between consecutive heart beat intervals, natural logarithm of the percentage of adjacent heart beat intervals changing by more than 50 milliseconds, natural logarithm of high frequency to low frequency, and ratio of low frequency to high frequency.

[0022] Optionally, the feature extraction module is specifically used to:

[0023] The Wilcoxon rank sum test was used to compare the difference in the change in physiological signals between baseline and week 4 between the intervention group and the waiting group. By comparing the magnitude of change between the two groups, physiological signals showing significant changes in the intervention group were screened out. At the same time, whether the significantly changed physiological signals were associated with the degree of depression was also evaluated.

[0024] In the baseline data of the intervention group, Spearman correlation analysis was used to evaluate the correlation between physiological signals and the scores of the self-rating depression scale, and physiological signals related to depression were screened out as input features of the deep learning model to be trained.

[0025] Optionally, the process of using the Wilcoxon rank sum test to compare differences includes:

[0026] Let X be the target physiological signal, measured at baseline t0 and week 4 t4;

[0027] Define the change Δ: ΔX = Xt4 - Xt0;

[0028] G is the group variable: G = {1, intervention group; 0, waiting group};

[0029] Test procedure: Rank Sum Test: Compare ΔX of intervention group G=1 and waiting group G=0;

[0030] H0: median (ΔX|G=1) = median (ΔX∣G=0);

[0031] Ha: Median (ΔX|G=1)≠Median (ΔX∣G=0);

[0032] The Wilcoxon rank sum test was used to statistically analyze the p-value and screen for significant physiological signal features, i.e., p-value < 0.05; value is a given parameter;

[0033] Screening results: A set of physiological signals with significant changes was obtained: F 筛选 ={X1,X2,…,X k};

[0034] Spearman correlation analysis was used to evaluate the correlation between physiological signals and scores on the Self-Rating Depression Scale, including:

[0035] In the baseline data of the intervention group, X i Indicates that from F 筛选 The physiological signal features screened out in , Y represents the depression score; then X i Correlation coefficient between ρ and Y: XiY =Spearman(X i ,Y); Test correlation: H 01 :ρ XiY =0;H a1 :ρ XiY ≠0;

[0036] The Spearman correlation test was used to statistically analyze the p1 value and screen for significantly correlated physiological signals: p1-value < 0.05;

[0037] Screening results: Obtained a set of significantly correlated physiological signals: F 相关 ={X i1 ,X i2 ,…,X im}

[0038] F 相关 It is the input feature of the deep learning model to be trained.

[0039] Optionally, the deep learning model is a ResNet-1D model, which includes: a feature sequence input layer, an initial convolution module, a residual feature extraction module, a global feature aggregation layer, a classification output layer,

[0040] The feature sequence input layer is a one-dimensional residual neural network, which is used to receive physiological signal sequence data of shape (N, C, L), where N is the number of samples, C=1 is the number of input channels, and L is the length of the physiological feature sequence;

[0041] The physiological feature sequence data received by the feature sequence input layer is processed by the initial convolution module, the residual feature extraction module, and the global feature aggregation layer, and then the classification output layer outputs the result of whether the biofeedback therapy is suitable;

[0042] The criterion for the effectiveness of biofeedback therapy is a reduction of ≥50% in the total score of the Self-Rating Depression Questionnaire-9 (PHQ-9) after four weeks. The output is the result that biofeedback therapy is effective, i.e., the total score of the PHQ-9 is reduced by ≥50%.

[0043] The deep learning model training module adopts an innovative compound learning rate scheduling strategy, the core formula of which is the learning rate formula:

[0044]

[0045] in, is the outer exponential decay term, cos(π·(t madT) / T) is the inner cosine term, η base is the initial learning rate, η min is the minimum learning rate lower limit, k is the global attenuation coefficient, T is the basic cycle length, t is the current training step number, β is the noise mixing coefficient, ∈ t is a random noise term, mod represents the modulo operation, which means calculating the remainder after dividing two numbers.

[0046] According to the above formula, when in the initial stage of training (t<T), Exponential term e -0 =1, the noise term maintains the maximum intensity, at this time the learning rate mainly presents a complete cosine oscillation + strong random perturbation;

[0047] At the end of training (t=4T), the exponential term decays significantly, and the learning rate oscillates with weak perturbations in a lower amplitude range.

[0048] By integrating the cosine annealing learning rate scheduling strategy with the triple mechanisms of exponential decay basis, dynamic cycle adjustment and adaptive noise injection, efficient model training is achieved through the synergistic effect of the triple mechanisms.

[0049] Optionally, the deep learning model training module is also used to:

[0050] The 5-fold cross-validation method is used to evaluate the prediction results of the deep learning model. TP is the number of correctly predicted positive samples, TN is the number of correctly predicted negative samples, FP is the number of incorrectly predicted positive samples, and FN is the number of incorrectly predicted negative samples.

[0051] Precision: The proportion of samples predicted as positive that are correctly predicted.

[0052] Accuracy: The ratio of the number of correct predictions in positive and negative examples to the total number.

[0053]

[0054] Recall: The ratio of the predicted correct positive examples to the total positive examples among the actual positive examples.

[0055] FI score: an indicator that neutralizes precision and recall.

[0056] Optionally, it also includes:

[0057] The predictive model interpretability analysis module is used to apply the game theory method SHAP to perform model interpretability analysis, and uses SHAP values ​​to visualize the degree of dependence of different physiological signals and their impact on the prediction results.

[0058] Optionally, the number of members in the intervention group is 50, and the number of members in the intervention group and the waiting group is the same;

[0059] The input features include: physical relaxation index, frontal surface electromyography, psychological relaxation index and the ratio of low frequency to high frequency, i.e. LF / HF.

[0060] (3) Beneficial effects

[0061] In response to the limitations of existing technologies in the selection and clinical efficacy evaluation of biofeedback therapy for depression, physiological signals have better stability and measurement repeatability in depression-related research. In the embodiments of the present invention, objective physiological signals are combined with advanced artificial intelligence algorithms to evaluate the therapeutic effect of biofeedback on patients with depression. This not only solves the subjectivity brought by scale evaluation, but also introduces objective and intelligent evaluation methods.

[0062] The system of the present invention integrates artificial intelligence technology, simulates the expert decision-making process, can assist clinicians in making treatment decisions, and adjust treatment plans in a timely manner; it helps clinicians to objectively evaluate the effects of biofeedback treatment, achieve precise treatment and individualized efficacy prediction; thereby providing patients with more timely and effective medical services, promoting their mental health, and having high translational application value.

[0063] Furthermore, when used for individual prediction, it is possible to recommend an intervention plan to individuals who have responded to biofeedback therapy before the intervention. This invention can effectively solve the problem of traditional methods that cannot objectively evaluate and predict biofeedback therapy, and more objectively and accurately predict patients' intervention response to biofeedback therapy, providing an objective basis for individualized intervention for depression.

[0064] That is, it is possible to predict the early effects of biofeedback intervention based on the physiological signals of depression patients before intervention, which helps to achieve individualized and precise intervention and improve the quality of medical services. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 and Figure 2 They are respectively architecture diagrams of a personalized assessment system for depression biofeedback results based on physiological signals provided by an embodiment of the present invention;

[0066] Figure 3 Schematic diagram of the correlation between physiological signals and depression scale scores in patients with baseline depression;

[0067] Figure 4Schematic diagram of the confusion matrix and ROC curve of the biofeedback intervention effect prediction model built based on the ResNet-1D algorithm and four physiological signals;

[0068] Figure 5 Schematic diagram of SHAP value analysis of the constructed model after training;

[0069] Figure 6 Schematic diagram of the deep learning model training process. DETAILED DESCRIPTION

[0070] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation methods in conjunction with the accompanying drawings.

[0071] With the development of artificial intelligence, deep learning can not only find the model with the best predictive performance, but also reveal the weight and importance of different features in the prediction process, providing an intelligent evaluation method for the evaluation of the effect of biofeedback intervention.

[0072] The method of the embodiment of the present invention relates to the field of computer application technology. It uses advanced artificial intelligence algorithms to construct a prediction model for the effect of biofeedback treatment of depression based on physiological signals. It can predict the early effects of biofeedback treatment based on the physiological signals of depression patients before treatment, which helps to achieve individualized precise intervention and improve the quality of medical services.

[0073] In the embodiment of the present invention, a personalized assessment system for depression biofeedback results based on physiological signals is established. Compared with single subjective clinical assessment and retrospective analysis and comparison, the solution of the present invention can objectively evaluate the effect of biofeedback intervention, improve prediction accuracy, reduce the time of clinical trial and error, shorten the treatment cycle, help to achieve personalized and precise intervention, and has certain clinical application value.

[0074] The physiological signals and their meanings involved in the embodiments of the present invention are described as follows:

[0075] Body_Data, which represents the body relaxation index and is a comprehensive indicator that combines multiple physiological signals;

[0076] Mental_Data, which represents the psychological relaxation index and is a comprehensive indicator that combines multiple physiological signals;

[0077] EMG (Electromyography), which stands for electromyography, is used to record electrical signals generated by muscle activity;

[0078] HR (Heart Rate), stands for heart rate, which refers to the number of heartbeats per minute;

[0079] LN(sdnn), "sdnn" is the natural logarithm of the standard deviation, which represents the standard deviation of all normal heart beat intervals and is used to measure heart health and stress levels;

[0080] LN(rmssd), "rmssd" is the natural logarithm of the root mean square difference between consecutive heart beat intervals, which is used to evaluate the parasympathetic nervous system activity of the heart;

[0081] LN(pnn50), "pnn50" is the natural logarithm of the percentage of adjacent heart beat intervals that vary by more than 50 milliseconds, used to assess parasympathetic nervous system activity;

[0082] LN(tp), "tp" is the natural logarithm of total power, which represents the total energy of all frequency components in heart rate variability analysis;

[0083] LN(vlf), "vlf" is the natural logarithm of very low frequency, which is related to the regulatory function of the autonomic nervous system;

[0084] LN(lf), where "lf" is the natural logarithm of the low frequency, which is related to both sympathetic and parasympathetic nervous system activity;

[0085] LN(hf), where "hf" is the natural logarithm of high frequency, a frequency component of HRV primarily associated with parasympathetic nervous system activity

[0086] LF / HF, the ratio of low frequencies (usually in the frequency range of 0.04 to 0.15 Hz) to high frequencies (usually in the frequency range of 0.15 to 0.4 Hz), is used to assess the balance between sympathetic and parasympathetic nerves. A higher LF / HF ratio is generally interpreted as relatively high sympathetic activity, while a lower ratio may indicate a predominance of parasympathetic activity. LF / HF represents a ratio.

[0087] Example 1

[0088] The embodiment of the present invention provides an individualized assessment system for depression biofeedback results based on physiological signals, such as Figure 1 and Figure 2 As shown, the individualized assessment system of this embodiment includes:

[0089] a signal acquisition module, configured to acquire first data from the intervention group and second data from the waiting group; the first data from the intervention group includes: physiological signals of the intervention group before biofeedback therapy, physiological signals within a specified time period after biofeedback therapy (in this embodiment, the data is collected twice, once before the intervention and once after the complete intervention), and the scores of the self-rating depression scale of the intervention group within the specified time period; the second data from the waiting group includes: physiological signals of the waiting group before each biofeedback therapy, and the scores of the self-rating depression scale of the waiting group within the specified time period; the waiting group does not undergo biofeedback therapy;

[0090] a feature extraction module, configured to screen out physiological signals related to depression based on the first data of the intervention group and the second data of the waiting group, and use them as input features of the deep learning model to be trained;

[0091] The deep learning model training module is used to pre-process the screened physiological signals, input the pre-processed values ​​into the deep learning model for training, obtain the output results, and evaluate the output results; when the evaluation reaches the threshold, obtain the deep learning model whose evaluation results after training are within the expected range;

[0092] The deep learning model is used to evaluate the physiological signals of an individual.

[0093] In a specific application, the individualized assessment system of this embodiment further includes: a prediction module;

[0094] The signal acquisition module is used to acquire physiological signals of the patient to be evaluated;

[0095] The prediction module is used to input the physiological signal of the patient to be evaluated into the trained deep learning module, obtain the results output by the deep learning model, and determine whether the biofeedback therapy is suitable for the patient to be evaluated.

[0096] For example, the aforementioned signal acquisition module can be specifically used for:

[0097] The physiological signals of each member of the intervention group were collected before and after the biofeedback therapy for 12 sessions over four weeks.

[0098] The PHQ-9 score of each member of the intervention group was collected. In this embodiment, the physiological signals and depression scale were collected before the intervention and after 12 interventions, for a total of 2 times.

[0099] The physiological signals include one or more of the following: a physical relaxation index, a psychological relaxation index, a frontal surface electromyogram (EMG), a blood volume pulse, heart rate variability, heart rate, the natural logarithm of the root mean square of the difference between consecutive heartbeat intervals, the natural logarithm of the percentage of adjacent heartbeat intervals that vary by more than 50 milliseconds, the natural logarithm of the high-frequency to low-frequency ratio, and the ratio of the low-frequency to high-frequency ratio. This embodiment is not limited to the above physiological signals, and additional physiological signals may be added as needed.

[0100] In this embodiment, the number of members in the intervention group is 50, and the number of members in the intervention group and the waiting group is the same.

[0101] The aforementioned feature extraction module is specifically used for:

[0102] The Wilcoxon rank sum test was used to compare the difference in the change in physiological signals from baseline to week 4 between the intervention group and the waiting group. By comparing the magnitude of change between the two groups, physiological signals that showed significant changes in the intervention group were screened out. At the same time, whether the significantly changed physiological signals were associated with the degree of depression was also evaluated.

[0103] In the baseline data of the intervention group, Spearman correlation analysis was used to evaluate the correlation between physiological signals and the scores of the self-rating depression scale, and physiological signals related to depression were screened out as input features of the deep learning model to be trained.

[0104] For example, the process of comparing differences using the Wilcoxon rank sum test includes:

[0105] Let X be the target physiological signal, measured at baseline t0 and week 4 t4;

[0106] Define the change Δ: ΔX = Xt4 - Xt0;

[0107] G is the group variable: G = {1, intervention group; 0, waiting group};

[0108] Test procedure: Rank sum test: compare ΔX between the intervention group (G = 1) and the waiting group (G = 0);

[0109] H0: median (ΔX|G=1) = median (ΔX∣G=0);

[0110] H a :Median (ΔX|G=1)≠Median (ΔX∣G=0);

[0111] The Wilcoxon rank sum test was used to calculate the p value and screen for significant physiological signal features.

[0112] That is: p-value<0.05; value is a given parameter;

[0113] Screening results: Obtained significantly changed physiological signal set / physiological signal feature set: F 筛选 ={X1,X2,…,X k};

[0114] In addition, this embodiment also uses Spearman correlation analysis to evaluate the correlation between physiological signals and scores of the self-rating depression scale, including:

[0115] In the baseline data of the intervention group, let X i It is from F 筛选 Physiological signal features screened out, let Y be the depression score;

[0116] X i The correlation coefficient between and Y is: ρ XiY =Spearman(X i ,Y);

[0117] Test correlation: H 01 :ρ XiY =0;H a1 :ρ XiY ≠0;

[0118] The Spearman correlation test was used to calculate the p1 value and screen for significantly correlated physiological signals: p1-value < 0.05;

[0119] Screening results: Obtained a set of significantly correlated physiological signals: F 相关 ={X i1 ,X i2 ,…,X im}

[0120] F 相关 It is the input feature of the deep learning model to be trained.

[0121] For example, the input features may include: a body relaxation index, a frontal surface electromyogram, a psychological relaxation index, and a ratio of low frequency to high frequency, ie, LF / HF, where LF is low frequency and HF is high frequency.

[0122] like Figure 6 As shown, the deep learning model training module of this embodiment can be specifically used for:

[0123] The ResNet-1D model is trained. For example, the model's feature sequence input layer receives a physiological signal sequence of shape (N, C, L), where N is the number of samples, C = 1 is the number of input channels, and L is the length of the physiological feature sequence. The input data must be preprocessed by Z-score normalization. The initial convolution module includes a 7×1 convolutional layer with 64 filters and a stride of 2, followed by batch normalization and ReLU activation function, and a 3×1 maximum pooling layer with a stride of 2.

[0124] Residual feature extraction module: It contains a four-level residual structure, each level contains two residual blocks. The first level is configured with 64 filters, the second level is configured with 128 filters and uses a downsampling connection with a stride of 2, the third level is configured with 256 filters and uses a downsampling connection with a stride of 2, and the fourth level is configured with 512 filters and uses a downsampling connection with a stride of 2. Each residual block contains two 3×1 convolutional layers, each convolutional layer is followed by batch normalization and ReLU activation function, and cross-layer identity connections are used between blocks;

[0125] Global feature aggregation layer: Adaptive average pooling is used to compress the feature map into a fixed dimension;

[0126] Classification output layer: predicts the biofeedback efficacy through the fully connected layer and Sigmoid function;

[0127] During model training, the data was divided into 80% training set and 20% test set, and the following hyperparameters were optimized through grid search: the initial learning rate was selected from {0.001, 0.0005, 0.0001}, the batch size was selected from {16, 32}, and the weight decay coefficient was selected from {0, 0.0001}; the AdamW optimizer was used during training, and the cosine annealing learning rate schedule was applied. An early stopping mechanism was set to terminate training when the AUC of the validation set did not improve for five consecutive rounds, and a dropout rate of 0.2 was applied to the fully connected layer for regularization.

[0128] Among them, when applying cosine annealing learning rate scheduling, combined with exponential decay and dynamic cycle adjustment, the learning rate formula is:

[0129]

[0130] in, is the outer exponential decay term, cos(π·(t madT) / T) is the inner cosine term, η base is the initial learning rate, η min is the minimum learning rate lower limit, k is the global attenuation coefficient, T is the basic cycle length, t is the current training step number, β is the noise mixing coefficient, ∈ t is a random noise term, mod represents the modulo operation, which means calculating the remainder after dividing two numbers.

[0131] According to the above formula, when in the initial stage of training (t<T), Exponential term e -0 =1, the noise term maintains the maximum intensity, at this time the learning rate mainly presents a complete cosine oscillation + strong random perturbation;

[0132] At the end of training (t=4T), the exponential term decays significantly, and the learning rate oscillates with weak perturbations in a lower amplitude range.

[0133] In this embodiment, the cosine annealing learning rate scheduling strategy is integrated with the triple mechanisms of exponential decay basis, dynamic cycle adjustment and adaptive noise injection. Based on the standard cosine oscillation, this strategy realizes the global learning rate decrease through the exponential decay term. At the same time, random noise negatively correlated with the training progress is introduced to enhance the model's ability to escape from local extreme values, and a cycle adaptive adjustment algorithm based on the direction of the loss gradient is designed, thereby significantly optimizing the biological efficacy prediction performance of the ResNet-1D model.

[0134] It should be noted that during deep learning model training, the learning rate scheduling strategy directly impacts the model's convergence speed and ultimate performance. Traditional methods, such as fixed learning rates and step-by-step descent, suffer from slow convergence and a tendency to fall into local optima. While advanced strategies such as cosine annealing have improved, balancing global convergence and local optimization remains difficult when processing time-series data such as physiological signals. The compound learning rate scheduling method proposed in the above embodiment effectively addresses this issue.

[0135] F 相关 The deep learning model was input; the ResNet-1D algorithm was used, and the PHQ-9 total score reduction of ≥50% after four weeks was used as the criterion for the effectiveness of biofeedback therapy. The output was the result that biofeedback therapy was effective, that is, the total score of PHQ-9 was reduced by ≥50%, otherwise it was ineffective.

[0136] Deep learning model training module, also used for:

[0137] The 5-fold cross-validation method is used to evaluate the prediction results of the deep learning model. TP is the number of correctly predicted positive samples, TN is the number of correctly predicted negative samples, FP is the number of incorrectly predicted positive samples, and FN is the number of incorrectly predicted negative samples.

[0138] Precision: The proportion of samples predicted as positive that are correctly predicted.

[0139] Accuracy: The ratio of the number of correct predictions in positive and negative examples to the total number.

[0140] Recall: The ratio of the predicted correct positive examples to the total positive examples among the actual positive examples.

[0141] FI score: an indicator that neutralizes precision and recall.

[0142] Of course, the individualized assessment system of this embodiment also includes:

[0143] The predictive model interpretability analysis module is used to apply the game theory method SHAP to perform model interpretability analysis, and uses SHAP values ​​to visualize the degree of dependence of different physiological signals and their impact on the prediction results.

[0144] The system of this embodiment integrates artificial intelligence technology to simulate the expert decision-making process, which can assist clinicians in making treatment decisions and adjusting treatment plans in a timely manner; it helps clinicians to objectively evaluate the effects of biofeedback treatment, achieve precise treatment and individualized efficacy prediction; thereby providing patients with more timely and effective medical services, promoting their mental health, and having high translational application value.

[0145] Furthermore, when used for individual prediction, it is possible to recommend an intervention plan to individuals who have responded to biofeedback therapy before the intervention. This invention can effectively solve the problem of traditional methods that cannot objectively evaluate and predict biofeedback therapy, and more objectively and accurately predict patients' intervention response to biofeedback therapy, providing an objective basis for individualized intervention for depression.

[0146] Example 2

[0147] The individualized evaluation system for depression biofeedback results based on physiological signals of this embodiment may include: a physiological signal collection and data processing module, a feature extraction module, a prediction model construction and evaluation module, a prediction model interpretable analysis module, and an individualized prediction module; Figure 2 shown.

[0148] (1) Physiological signal collection and data collation module, specifically used to collect physiological signals of depression patients before and after intervention.

[0149] The current depression intervention is a standardized 4-week biofeedback therapy intervention for patients with depression, 3 times a week, for a total of 12 sessions. The specific intervention plan is shown in Table 1. The waiting group did not receive any other intervention during the same period.

[0150] For example, the depression scale PHQ-9 scale and physiological signal data of depression patients are collected before and after a 4-week biofeedback therapy intervention (once before the intervention and once after the intervention).

[0151] Physiological signals include one or more of the following: physical relaxation index, psychological relaxation index, frontal surface electromyography, blood volume pulse, heart rate variability, heart rate, natural logarithm of the root mean square of the difference between consecutive heartbeat intervals, natural logarithm of the percentage of adjacent heartbeat intervals changing by more than 50 milliseconds, natural logarithm of high frequency to low frequency, ratio of low frequency to high frequency, and other physiological signals. Preprocessing is performed, such as eliminating data that cannot be used to evaluate efficacy or interfere with efficacy evaluation; see Table 2 for details. The main outcome indicator for evaluating the effect of biofeedback intervention is the PHQ-9 scale score.

[0152] The outcome indicator here is whether the biofeedback therapy is effective, that is, the PHQ-9 total score is reduced by ≥50% as the standard for the effectiveness of biofeedback therapy. The output is the result that the biofeedback therapy is effective, that is, the PHQ-9 total score is reduced by ≥50%, otherwise it is an ineffective result.

[0153] Table 1. Biofeedback treatment options

[0154]

[0155] Table 2. Physiological signals collected in this embodiment

[0156]

[0157]

[0158] (2) Feature extraction module, specifically used for:

[0159] The Wilcoxon rank sum test was used to compare the difference in the change (△) of the physiological signals between the intervention group and the waiting group from baseline to week 4. By comparing the magnitude of change between the two groups, the physiological signals showing significant changes in the intervention group were screened out. It was further evaluated whether the physiological signals with significant changes were related to the degree of depression. In the baseline data of the intervention group, Spearman correlation analysis was used to evaluate the correlation between the physiological signals and the scores of the self-rating depression scale, and to screen out the physiological signals related to depression. Figure 3 As shown in the above analysis, it is used as the input feature of the model.

[0160] Wilcoxon rank sum test analysis of differences:

[0161] Let X be the target physiological signal, measured at baseline (t0) and week 4 (t4).

[0162] Define the change (Δ): ΔX = Xt4 - Xt0;

[0163] G is the group variable: G = {1, intervention group; 0, waiting group};

[0164] Test procedure: Rank sum test: compare ΔX between the intervention group (G = 1) and the waiting group (G = 0);

[0165] H0: median (ΔX|G=1) = median (ΔX∣G=0);

[0166] H a :Median (ΔX|G=1)≠Median (ΔX∣G=0);

[0167] The Wilcoxon rank sum test was used to statistically analyze the p-value and screen for significant physiological signal features, i.e., p-value < 0.05, where value is a given parameter.

[0168] Screening results: A set of physiological signals with significant changes was obtained: F 筛选 ={X1,X2,…,X k}.

[0169] Furthermore, Spearman correlation analysis was performed:

[0170] In the baseline data of the intervention group, let X i It is from F 筛选 Physiological signal features screened out, let Y be the depression score.

[0171] Plan X i Correlation coefficient between ρ and Y: XiY =Spearman(X i ,Y)

[0172] Test for correlation: H0:ρ XiY =0;H a :ρ XiY ≠0

[0173] Use the Spearman correlation test to obtain statistically significant p1 values ​​and filter out significantly correlated physiological signals:

[0174] P1-value<0.05, screening results: a set of significantly correlated physiological signals was obtained: F 相关 ={X i1 ,X i2 ,…,X im};F 相关 This is the characteristic of the model.

[0175] (3) Prediction model construction and evaluation module, specifically used for:

[0176] The aforementioned deep learning model is a ResNet-1D model. A ResNet-1D model is constructed, which includes: a feature sequence input layer, an initial convolution module, a residual feature extraction module, a global feature aggregation layer, and a classification output layer.

[0177] Among them, the feature sequence input layer: receives the physiological signal sequence of shape (N, C, L), where N is the number of samples, C = 1 is the number of input channels, and L is the length of the physiological feature sequence. The input data needs to be preprocessed by Z-score standardization;

[0178] Initial convolutional module: This consists of a 7×1 convolutional layer with 64 filters and a stride of 2, followed by batch normalization and ReLU activation, and a 3×1 max pooling layer with a stride of 2.

[0179] Residual feature extraction module: It contains a four-level residual structure, each level contains two residual blocks. The first level is configured with 64 filters, the second level is configured with 128 filters and uses a downsampling connection with a stride of 2, the third level is configured with 256 filters and uses a downsampling connection with a stride of 2, and the fourth level is configured with 512 filters and uses a downsampling connection with a stride of 2. Each residual block contains two 3×1 convolutional layers, each convolutional layer is followed by batch normalization and ReLU activation function, and cross-layer identity connections are used between blocks;

[0180] Global feature aggregation layer: Adaptive average pooling is used to compress the feature map into a fixed dimension;

[0181] Classification output layer: predict the biofeedback efficacy through the fully connected layer and Sigmoid function.

[0182] During model training, the data can be divided into 80% training set and 20% test set, and the following hyperparameters are optimized through grid search: the initial learning rate is selected from {0.001, 0.0005, 0.0001}, the batch size is selected from {16, 32}, and the weight decay coefficient is selected from {0, 0.0001}; the AdamW optimizer is used during training, and the cosine annealing learning rate schedule is applied. The early stopping mechanism is set to terminate the training when the AUC of the validation set does not improve for 5 consecutive rounds, and a dropout rate of 0.2 is applied to the fully connected layer for regularization.

[0183] As described in Example 1, a compound learning rate scheduling strategy was also used to optimize convergence performance when training the ResNet-1D model. This strategy combines cosine annealing, exponential decay, and dynamic noise injection to improve the model's generalization ability on physiological signal data.

[0184] The baseline values ​​of physiological signals related to depression that changed significantly before and after biofeedback treatment in the intervention group were used as input features. The ResNet-1D algorithm was used, and a ≥50% reduction in the total score of the Self-Rating Depression Questionnaire-9 (PHQ-9) after 4 weeks was used as the criterion for effective treatment. The output was whether the depression treatment was effective (a ≥50% reduction in the PHQ-9 total score) or ineffective (a <50% reduction in the PHQ-9 total score).

[0185] The 5-fold cross-validation method is used to evaluate the result type predicted by the model. TP is the number of correctly predicted positive samples, TN is the number of correctly predicted negative samples, FP is the number of incorrectly predicted positive samples, and FN is the number of incorrectly predicted negative samples.

[0186] Precision: The proportion of samples predicted as positive that are correctly predicted.

[0187] Accuracy: The ratio of the number of correct predictions in positive and negative examples to the total number.

[0188]

[0189] Recall: The ratio of the predicted correct positive examples to the total positive examples among the actual positive examples.

[0190] FI score: an indicator that neutralizes precision and recall.

[0191] ROC and AUC: ROC is the abbreviation of Receiver Operating Characteristic, and AUC is the abbreviation of Area Under ROC Curve. The larger the AUC value, the better the classification performance.

[0192] The evaluation indicators of the model include receiver operating characteristic (ROC) curve, area under the ROC curve (AUC), accuracy, precision, recall and F1 value.

[0193] In order to ensure the reliability and stability of the model, reduce the risk of overfitting, and improve the generalization ability of the model, a five-fold cross-validation was used, and the data was divided into an 80% training set and a 20% test set. The training set data was used for training, and the test set data was used to predict the tuned model.

[0194] First, the physiological signal data before intervention were preprocessed, and the physiological signals that changed significantly before and after intervention and were related to depression were used as input features. Then, the classifier variables (output of the model) were encoded, and a decrease in the PHQ-9 total score of less than 50% after 4 weeks of intervention was marked as 0 (no response to intervention); a decrease in the PHQ-9 total score of ≥50% was marked as 1 (no response to intervention). A biofeedback intervention effect prediction model based on physiological signals was established.

[0195] Based on the confusion matrix data, there are five indicators used to evaluate the generalization performance of the classifier, including the receiver operating characteristic curve (ROC), the area under the curve (AUC), accuracy, precision, recall and F1 value. Figure 4 shown.

[0196] The ROC curve is a graph plotted with the true positive rate on the y-axis and the false positive rate on the x-axis. The diagonal line corresponds to the random guessing model, and the closer the curve is to the upper left corner, the better the classification performance. The ROC curve is a qualitative metric that is simple, intuitive, and highly readable. The more convex the ROC curve is and the closer it is to the upper left corner, the greater its diagnostic value. The AUC is a quantitative metric that is more accurate and objective. A larger AUC value indicates better classification performance.

[0197] (4) Prediction model interpretable analysis module, specifically used for:

[0198] Applying the game theory method SHAP to analyze the model interpretability, and using SHAP values ​​to visualize the degree of dependence on different physiological signals and their impact on the prediction results, will help realize the practical application of the model, such as Figure 5 shown.

[0199] (5) Individualized prediction module, specifically used for:

[0200] For each patient to be evaluated, the trained deep learning model and features are used to calculate whether the individual's physiological signal values ​​are responsive to biofeedback therapy. The results predicted by the personalized prediction module are presented to doctors and patients, making intervention decisions more accurate and personalized.

[0201] The system of this embodiment uses physiological signals significantly altered before and after biofeedback intervention and associated with depression as objective indicators and employs advanced artificial intelligence algorithms to construct a biofeedback efficacy prediction model. This model provides an intelligent, individualized efficacy assessment method, helping to address the technical issues of existing technologies that prevent predictive individualized assessment of intervention effects, resulting in inadequate individualized matching of intervention plans and poor intervention effectiveness. This helps achieve individualized efficacy prediction and precise intervention. This invention is easily implemented in actual clinical practice in hospitals and is a highly efficient and intelligent biofeedback efficacy assessment method for depression.

[0202] In response to the limitations in the selection of biofeedback intervention programs for depression and the evaluation of clinical efficacy, physiological signals have better stability and measurement repeatability in depression-related research. The present invention combines objective physiological signals with advanced artificial intelligence algorithms to evaluate the intervention effect of biofeedback on patients with depression, providing an intelligent evaluation method and an individualized prediction system for the effect of biofeedback intervention. It not only solves the subjectivity brought by scale evaluation, but also introduces an objective and intelligent evaluation method. The present invention integrates artificial intelligence technology and simulates the decision-making process of experts, which can assist clinicians in making intervention decisions and adjusting intervention programs in a timely manner; it helps clinicians objectively evaluate the effect of biofeedback intervention and achieve precise intervention and individualized efficacy prediction; thereby providing patients with more timely and effective medical services, promoting their mental health, and having high translational application value.

[0203] Example 3

[0204] Step 1. Collection and data collation of physiological signals of patients with depression before and after biofeedback intervention;

[0205] Fifty patients with depression who had never received any intervention were recruited and randomly divided into an intervention group and a waiting group using a double-blind randomized controlled design. The patients' demographic data, depression scales, and physiological signals were collected. The intervention group underwent 4 weeks of biofeedback training, which included 12 units, each of which took about 30 minutes to complete. After the 4-week intervention, the PHQ-9 scale was evaluated and physiological signals were collected. The physiological signals and depression scale data before and after the intervention were collated. Figure 4 As shown, the depression of patients with depression was significantly improved before and after 4 weeks of group biofeedback intervention.

[0206] Step 2. By comparing depression scales and physiological signals before and after the intervention between the intervention and waitlist groups, we identified physiological signals that significantly changed and were associated with depression before and after the intervention, which served as feature variables for the deep learning model. After four weeks of intervention, the intervention group showed significant improvement in PHQ-9 scores. Intergroup analysis between the intervention and waitlist groups revealed significant differences in changes in PHQ-9 scores after four weeks of biofeedback intervention.

[0207] Compared with the waiting group, the physiological characteristics that showed significant differences at the fourth week (Δ) were the body relaxation index, surface electromyography, and LF / HF in the intervention group. Compared with before the intervention, the main changes in the intervention group after four weeks were the body relaxation index, surface electromyography, psychological relaxation index, LF / HF, LN (sdnn), LN (rmssd), LN (pnn50), and LN (tp). Correlation analysis results showed that before the intervention, the PHQ-9 total score in the biofeedback intervention group was correlated with the body relaxation index, surface electromyography, psychological relaxation index, and LF / HF, such as Figure 3 shown.

[0208] Step 3. Using the four features of physical relaxation index, surface electromyography, psychological relaxation index, and LF / HF as input features, the ResNet-1D algorithm was used to establish a depression biofeedback intervention effect prediction model, namely a deep learning model.

[0209] The model was evaluated using a five-fold cross-validation method combined with the compound learning rate scheduling strategy described in Example 1. The results were AUC = 0.912, accuracy = 0.865, precision = 0.863, recall = 0.880, and F1 score = 0.871.

[0210] Figure 4 A schematic diagram of the confusion matrix (left) and the ROC curve (right) of the biofeedback intervention effect prediction model constructed based on the ResNet-1D algorithm and four physiological signals are shown. This shows that the model has a relatively high accuracy and can accurately judge the effect of depression intervention.

[0211] Step 4. Prediction model interpretability analysis, SHAP value analysis shows that Figure 5 As shown in the results, body relaxation index is the most important factor affecting the effect of biofeedback intervention, followed by surface electromyography.

[0212] Step 5. Using the trained prediction model, the pre-intervention physiological signal values ​​of the new patient were input into the deep learning model to calculate whether the patient responded to biofeedback therapy. These results show that the pre-intervention physical relaxation index, surface electromyography, psychological relaxation index, and LF / HF can predict the effectiveness of biofeedback intervention in patients with depression, as shown in the table below.

[0213]

[0214] The individualized prediction system of this embodiment can effectively conduct individualized predictive evaluations, more accurately predict different patients' responses to biofeedback interventions, help replace ineffective intervention plans in real time, provide a basis for individualized interventions for patients with depression, help improve the quality of medical services, promote patients' mental health, and have certain translational application value.

[0215] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0216] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions.

[0217] It should be noted that, in the claims, any reference signs placed between brackets shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention may be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims enumerating several means, several of these means may be embodied by one and the same hardware. The use of the words first, second, third etc. is for convenience only and does not indicate any order. These words may be understood as part of the component name.

[0218] In addition, it should be noted that, in the description of this specification, the description of the terms "one embodiment", "some embodiments", "embodiment", "example", "specific example" or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples, unless they are contradictory.

[0219] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments after learning the basic creative concept. Therefore, the claims should be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0220] Obviously, those skilled in the art may make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if such modifications and variations fall within the scope of the claims and their equivalents, the present invention shall also include such modifications and variations.

Claims

1. A personalized assessment system for depression biofeedback results based on physiological signals, characterized by: include: A signal acquisition module, configured to acquire first data of the intervention group and second data of the waiting group; The first data of the intervention group includes: physiological signals of the intervention group before biofeedback therapy, physiological signals within a specified time after biofeedback therapy, and scores of the self-rating depression scale of the intervention group within a specified time before and after the intervention; the second data of the waiting group includes: physiological signals of the waiting group before each biofeedback therapy, and scores of the self-rating depression scale of the waiting group within a specified time; the waiting group does not undergo biofeedback therapy; a feature extraction module, configured to screen out physiological signals related to depression based on the first data of the intervention group and the second data of the waiting group, and use them as input features of the deep learning model to be trained; The deep learning model training module is used to pre-process the physiological signals related to depression in the selected intervention group, input the pre-processed values ​​into the deep learning model for training, obtain output results, and evaluate the output results; when the evaluation reaches a threshold, a deep learning model is obtained whose evaluation results after training are within the expected range; The deep learning model is used to evaluate the physiological signals of an individual.

2. The individualized assessment system according to claim 1, characterized in that The individualized assessment system also includes: a prediction module; The signal acquisition module is used to acquire physiological signals of the patient to be evaluated; The prediction module is used to input the physiological signal of the patient to be evaluated into the trained deep learning module, obtain the results output by the deep learning model, and determine whether the biofeedback therapy is suitable for the patient to be evaluated.

3. The individualized assessment system according to claim 1, characterized in that: The signal acquisition module is specifically used for: The physiological signals of each member of the intervention group were collected before and after the biofeedback therapy for 12 sessions over four weeks. Before the biofeedback therapy, the PHQ-9 scores of each member in the intervention group were collected; The physiological signals include one or more of the following: physical relaxation index, psychological relaxation index, frontal surface electromyography, blood volume pulse, heart rate variability, heart rate, natural logarithm of the root mean square of the difference between consecutive heart beat intervals, natural logarithm of the percentage of adjacent heart beat intervals changing by more than 50 milliseconds, natural logarithm of high frequency to low frequency, and ratio of low frequency to high frequency.

4. The individualized assessment system according to claim 3, characterized in that: The feature extraction module is specifically used for: The Wilcoxon rank sum test was used to compare the difference in the change in physiological signals between baseline and week 4 between the intervention group and the waiting group. By comparing the magnitude of change between the two groups, physiological signals showing significant changes in the intervention group were screened out. At the same time, whether the significantly changed physiological signals were associated with the degree of depression was also evaluated. In the baseline data of the intervention group, Spearman correlation analysis was used to evaluate the correlation between physiological signals and the scores of the self-rating depression scale, and physiological signals related to depression were screened out as input features of the deep learning model to be trained.

5. The individualized assessment system according to claim 4, characterized in that: The process of comparing differences using the Wilcoxon rank sum test includes: Let X be the target physiological signal, measured at baseline t0 and week 4 t4; Define the change Δ: ΔX = Xt4 - Xt0; G is the group variable: G = {1, intervention group; 0, waiting group}; Test procedure: Rank Sum Test: Compare ΔX of intervention group G=1 and waiting group G=0; H0: Median (ΔX|G=1) = Median (ΔX∣G=0); H a : Median (ΔX∣G=1) ≠ Median (ΔX∣G=0); The Wilcoxon rank sum test was used to statistically analyze the p-value and screen for significant physiological signal features, i.e., p-value < 0.05; value is a given parameter; Screening results: A set of physiological signals with significant changes was obtained: F 筛选 ={X1,X2,…,X k }; Spearman correlation analysis was used to evaluate the correlation between physiological signals and scores on the Self-Rating Depression Scale, including: In the baseline data of the intervention group, X i Indicates that from F 筛选 The physiological signal features screened out in , Y represents the depression score; then X i Correlation coefficient between ρ and Y: XiY =Spearman(X i ,Y); Test correlation: H 01 :ρ XiY =0;H a1 :ρ XiY ≠0; The Spearman correlation test was used to statistically analyze the p1 value and screen for significantly correlated physiological signals: p1-value < 0.05; Screening results: Obtained a set of significantly correlated physiological signals: F 相关 ={X i1 ,X i2 ,…,X im } F 相关 It is the input feature of the deep learning model to be trained.

6. The individualized assessment system according to claim 4, characterized in that: The deep learning model is the ResNet-1D model, which includes: feature sequence input layer, initial convolution module, residual feature extraction module, global feature aggregation layer, classification output layer, The feature sequence input layer is a one-dimensional residual neural network, which is used to receive physiological signal sequence data of shape (N, C, L), where N is the number of samples, C=1 is the number of input channels, and L is the length of the physiological feature sequence; The physiological feature sequence data received by the feature sequence input layer is processed by the initial convolution module, the residual feature extraction module, and the global feature aggregation layer, and then the classification output layer outputs the result of whether the biofeedback therapy is suitable; The effectiveness of biofeedback therapy was determined by a ≥50% reduction in the total score of the Self-Rating Depression Questionnaire-9 (PHQ-9) after four weeks. The output was the effectiveness of biofeedback therapy, i.e., a ≥50% reduction in the total score of the PHQ-9.

7. The individualized assessment system according to claim 6, characterized in that: The deep learning model training module adopts a compound learning rate scheduling strategy during the training process. The learning rate calculation formula in the compound learning rate scheduling strategy is: in, is the outer exponential decay term, cos(π·(tmadT) / T) is the inner cosine term, η base is the initial learning rate, η min is the minimum learning rate lower limit, k is the global attenuation coefficient, T is the basic cycle length, t is the current training step number, β is the noise mixing coefficient, ∈ t is a random noise term, mod represents the modulo operation, which means calculating the remainder after dividing two numbers; According to the above formula, when t<T at the beginning of training, Exponential term e -0 = 1, the noise term maintains maximum strength, and the learning rate exhibits complete cosine oscillations and strong random perturbations; At the end of training (t=4T), the exponential term decays significantly, and the learning rate oscillates with weak perturbations in a lower amplitude range. By integrating the cosine annealing learning rate scheduling strategy with the triple mechanisms of exponential decay basis, dynamic cycle adjustment and adaptive noise injection, efficient model training is achieved through the synergistic effect of the triple mechanisms.

8. The individualized assessment system according to claim 6, characterized in that: Deep learning model training module, also used for: The 5-fold cross-validation method is used to evaluate the prediction results of the deep learning model. TP is the number of correctly predicted positive samples, TN is the number of correctly predicted negative samples, FP is the number of incorrectly predicted positive samples, and FN is the number of incorrectly predicted negative samples. Precision: The proportion of samples predicted as positive that are correctly predicted. Accuracy: The ratio of the number of correct predictions in positive and negative examples to the total number. Recall: The ratio of the predicted correct positive examples to the total positive examples among the actual positive examples. FI score: an indicator that neutralizes precision and recall.

9. The individualized assessment system according to claim 8, characterized in that: Also includes: The predictive model interpretability analysis module is used to apply the game theory method SHAP to perform model interpretability analysis, and uses SHAP values ​​to visualize the degree of dependence of different physiological signals and their impact on the prediction results.

10. The individualized assessment system according to claim 4, characterized in that: The number of members in the intervention group was 50, and the number of members in the intervention group and the waiting group was the same; The input features include: physical relaxation index, frontal surface electromyography, psychological relaxation index and the ratio of low frequency to high frequency, i.e. LF / HF.

Images