Mental health state trend analysis method based on AI model

Abstract

The invention relates to the technical field of psychological health analysis, in particular to a psychological health state trend analysis method based on an AI model, which comprises the following steps: recording psychological health level evaluation results of the same tested object at each evaluation time point one by one to form a psychological health state sequence; a Bi-LSTM network is combined with an attention mechanism to form a Bi-LSTM-Attention network architecture, training is carried out based on the mental health state sequence, and a mental health trend prediction model is obtained; and analyzing the long-term trend and the short-term volatility of the mental health of the tested object through a least square method and standard deviation calculation. According to the method, a Bi-LSTM network is combined with an attention mechanism to form a Bi-LSTM-Attention network architecture, and the long-term result and the short-term result of mental health level evaluation of the tested object are predicted, so that the long-term trend and the short-term fluctuation of mental health of the tested object are analyzed; and the long-term trend and short-term fluctuation of the psychological condition of the patient can be mastered in advance and continuously concerned.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mental health analysis, and particularly relates to a mental health state trend analysis method based on an AI model. BACKGROUND

[0002] Tumor patients are prone to have mental health problems such as anxiety, depression, and emotional disorders, and have a huge demand for psychological assessment. The traditional psychological assessment of patients in the tumor ward mainly relies on nurses using the PHQ-9 / GAD-7 scale, which has certain subjective limitations. At present, in order to reduce the subjective limitations of the scale, objective data such as emotional data and physiological data are combined with the self-evaluation of the scale to form multi-modal data for multi-modal mental health evaluation. In this way, the mental health status of the patient can be accurately evaluated, but single or multiple mental status evaluations can only grasp the current mental status of the patient, and it is difficult to realize the long-term trend and short-term fluctuation of the patient's mental status in advance, so that it is difficult to realize the long-term trend and short-term fluctuation of the patient's mental status in advance. pre-intervention, there is a certain lag in attention and intervention. SUMMARY

[0003] The present application aims to provide a mental health state trend analysis method based on an AI model to solve the technical problem that it is difficult to realize the long-term trend and short-term fluctuation of the patient's mental status in advance in the prior art.

[0004] To solve the above technical problems, the present application specifically provides the following technical solutions: A mental health state trend analysis method based on an AI model, comprising the following steps: Record the mental health level evaluation results of the same test object at each evaluation time point one by one to form a mental health state sequence; Use a Bi-LSTM network combined with an attention mechanism to form a Bi-LSTM-Attention network architecture, train based on the mental health state sequence, and obtain a mental health trend prediction model for predicting the long-term and short-term results of the mental health level evaluation of the test object; According to the long-term and short-term results of the mental health of the test object output by the mental health trend prediction model, the long-term trend and short-term fluctuation of the mental health of the test object are analyzed by least square method and standard deviation calculation, respectively.

[0005] As a preferred embodiment of the present application, the mental health level evaluation results include healthy, mild, moderate and severe, and the healthy, mild, moderate and severe are converted into numerical expressions 0, 1, 2 and 3, respectively, to form a mental health state sequence , , n is the total number of evaluation time points.

[0006] As a preferred scheme of the present application, the method for constructing the mental health trend prediction model comprises: obtaining a mental health state sequence As an input item of the Bi-LSTM network layer in the Bi-LSTM-Attention network architecture, output a hidden state vector , is the hidden state at each evaluation time point; obtaining a hidden state vector As an input of the attention layer in the Bi-LSTM-Attention network architecture, obtain an attention weight vector corresponding to the hidden state vector , and utilize the attention weight vector and the hidden state vector to form a context vector , is the attention weight of ; concatenate the context vector and the hidden state vector as an input of the Bi-LSTM-Attention network architecture, output a health grade evaluation short-term prediction sequence by the first output layer, and output a health grade evaluation long-term prediction sequence by the second output layer; ; As a preferred scheme of the present application, the loss function of the Bi-LSTM-Attention network architecture is ; ; ; ; In the formula, is a short-term prediction loss, is a long-term prediction loss, is a continuity loss of the short-term and long-term predictions, is a mean square error operation formula; is the health grade evaluation short-term prediction sequence, is the true value of the health grade evaluation short-term prediction sequence, is the health grade evaluation long-term prediction sequence, is the true value of the health grade evaluation long-term prediction sequence, is in the formula The sequence portion corresponding to the time sequence, To be Input the predicted sequence output by the first output layer in the Bi-LSTM-Attention network architecture; for Zhongyu The sequence portion corresponding to the time sequence. As a preferred embodiment of the present invention, the method for analyzing the long-term trend of the mental health of the tested subject includes: Long-term prediction series for health grade assessment using the least squares method Perform linear regression analysis, Convert to 2m data points Constructing a linear fitting model ,in, b is the slope, and b is the intercept; Establish a solution The objective function is to minimize the sum of squared residuals of b. Using S respectively Taking the partial derivatives of b and setting them equal to 0, we obtain the normal system of equations: ; ; Obtained from the normal equations Find the analytical solution for b, and use it To determine the long-term trend of the mental health of the test subjects; when When the test subjects' mental health deteriorates, the long-term trend of their mental health worsens. when When the test results show that the long-term trend of the test subject's mental health is improving; when When the test results are within a certain range, the long-term trend of the mental health of the tested subjects is stable.

[0007] As a preferred embodiment of the present invention, the method for analyzing short-term fluctuations in the mental health of the tested subject includes: Calculate the short-term predictive sequence of health level assessment Standard deviation , for The mean; when If the value exceeds the preset threshold, the short-term fluctuations in the mental health of the tested subject will be drastic. when If the value is less than or equal to the preset threshold, the short-term fluctuations in the mental health of the tested subject will be mild. Short-term predictive sequence of health level assessment The extreme value points found in the middle are marked as short-term wave crests and short-term wave troughs.

[0008] As a preferred scheme of the present application, the method for obtaining the mental health level evaluation result comprises: In the process of filling in the PHQ-9 / GAD-7 scale by the testee for self-evaluation, the expression data and the physiological data of the testee are collected; The self-evaluation result of the testee on the PHQ-9 / GAD-7 scale, the emotion recognition result and the physiological data are input into a pre-established classifier to output the mental health level evaluation result.

[0009] As a preferred scheme of the present application, the Bi-LSTM network is composed of a forward LSTM network and a reverse LSTM network, the forward LSTM network is used to obtain the forward mental health level prediction result output at each time point, the reverse LSTM network is used to obtain the reverse mental health level prediction result output at each time point, and the output result of the mental health level prediction result output in the two directions after splicing is taken as the final mental health level prediction result output at each time point. The structure expression of the Bi-LSTM network is: ; In the formula, 、 、 、 respectively, the forward propagation hidden layer state, the reverse propagation hidden layer state, the input mental health state sequence and the hidden state at the tth evaluation time point, is an activation function of the hidden layer, 、 、 、 respectively, the weight matrix of in 、 in 、 、 , 、 respectively, the forward propagation hidden layer state and the reverse propagation hidden layer state at the t-1th evaluation time point, 、 respectively, the bias of the forward propagation hidden layer and the reverse propagation hidden layer, is a vector splicing symbol.

[0010] As a preferred scheme of the present application, the physiological data comprises: heart rate HR, blood oxygen SpO2, respiratory rate RR, heart rate variability HRV.

[0011] As a preferred scheme of the present application, the emotion recognition result comprises: anger, disgust, fear, happy, normal, sad, and surprised.

[0012] Compared with the prior art, the present application has the following beneficial effects: The present application utilizes the Bi-LSTM network combined with the attention mechanism to form the Bi-LSTM-Attention network architecture, realizes the prediction of the long-term result and the short-term result of the mental health grade evaluation of the measured object, and thus analyzes the long-term trend and the short-term volatility of the mental health of the measured object, so as to achieve the long-term trend and the short-term volatility of the mental state of the patient. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are only exemplary, and for those skilled in the art, other drawings can be obtained from the provided drawings without creative labor.

[0014] Figure 1 The AI model-based mental health state trend analysis method flowchart is provided for the embodiments of the present application; Figure 2 The Bi-LSTM-Attention network architecture diagram is provided for the embodiments of the present application; Figure 3 The intelligent terminal schematic diagram is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0016] As shown in the drawings, Figure 1 The present application provides an AI model-based mental health state trend analysis method, comprising the following steps: The mental health grade evaluation results of the same measured object at each evaluation time point are recorded one by one to form a mental health state sequence; The Bi-LSTM network is combined with an attention mechanism to form a Bi-LSTM-Attention network architecture, and the psychological health trend prediction model for predicting long-term and short-term results of the psychological health grade evaluation of the measured object is obtained based on the training of the psychological health state sequence. According to the long-term and short-term results of the measured object psychological health output by the psychological health trend prediction model, the long-term trend and short-term volatility of the measured object psychological health are analyzed by least square method and standard deviation calculation respectively.

[0017] In order to master the psychological health status of the measured object in advance, the present data (psychological health grade evaluation results) recorded in the multiple evaluation process are subjected to time series prediction to obtain the prediction results of the psychological health status in the future time.

[0018] In order to master the long-term trend and short-term volatility of the measured object psychological health in the future time, the prediction results of the psychological health status in the future time are divided into two results, one is the long-term (such as one week, half a month, one month, etc. in the future) prediction result, corresponding to the analysis of long-term trend, and the other is the short-term (such as two weeks, one month, two months, etc. in the future, corresponding to the short-term twice) prediction result, corresponding to the analysis of short-term volatility.

[0019] Therefore, before analyzing the long-term trend and short-term volatility, the long-term prediction result and the short-term prediction result of the psychological health grade in the future time are obtained respectively according to the existing psychological health grade evaluation results.

[0020] In order to realize the time series prediction of the psychological health grade in the future time, the time series prediction model Bi-LSTM network is adopted, and the bidirectional LSTM network structure is used, the model can consider the past and future information in the time series data at the same time, so as to more comprehensively capture the trend of time series, allow the model not only to learn the input of the current time, but also to use the information of the future time at the same time, improve the modeling ability of the time series relationship. The advantage of bidirectional LSTM is that it can more effectively handle long-term dependencies, which helps to capture complex information in time series data. In addition, the bidirectional LSTM structure allows the model to learn the features of different network layers, which helps to better represent the structure of time series data.

[0021] The attention mechanism Attention is combined with the Bi-LSTM network, so that the Bi-LSTM network can better deal with noise and outliers in time series data, thereby enhancing its robustness and generalization, and being suitable for various time series prediction tasks. When dealing with complex time series data, the combination of LSTM and attention mechanism provides the model with stronger modeling ability, making it better adapt to different data forms and trends.

[0022] The mental health level evaluation result includes health, mild, moderate and severe, and the health, mild, moderate and severe are converted into numerical expressions 0, 1, 2 and 3 to form a mental health state sequence , wherein, , is the mental health level evaluation result at the tth evaluation time point, and n is the total number of evaluation time points.

[0023] The construction method of the mental health trend prediction model comprises: taking the mental health state sequence as an input item of a Bi-LSTM network layer in a Bi-LSTM-Attention network architecture (as shown in Figure 2 ), outputting a hidden state vector , from the Bi-LSTM network layer, and taking the hidden state vector as an input of an attention layer in the Bi-LSTM-Attention network architecture, obtaining an attention weight vector corresponding to the hidden state vector, and forming a context vector , by weighted summation of the attention weight vector and the hidden state vector ; taking the concatenation result of the context vector and the hidden state vector as an input of the Bi-LSTM-Attention network architecture, outputting, by a first output layer, a health level evaluation short-term prediction sequence , and outputting, by a second output layer, a health level evaluation long-term prediction sequence ; The present application accesses two output layers on the Bi-LSTM-Attention network, the first output layer is used for mapping the features learned by the model into the final short-term prediction result, and the second output layer is used for mapping the features learned by the model into the final long-term prediction result, so as to realize the synchronous output of the short-term prediction result and the long-term prediction result, without outputting one by one, and improve the prediction efficiency.

[0024] The loss function of the Bi-LSTM-Attention network architecture ; ; ; ; In the formula, is a short-term prediction loss, is a long-term prediction loss, is a continuity loss of short-term and long-term prediction, is a mean square error operation formula; is a health grade evaluation short-term prediction sequence, is a true value of the health grade evaluation short-term prediction sequence, is a health grade evaluation long-term prediction sequence, is a true value of the health grade evaluation long-term prediction sequence, is a sequence part corresponding to a time sequence in ; is a prediction sequence output by a first output layer in an input Bi-LSTM-Attention network architecture; is a sequence part corresponding to a time sequence in ;

[0025] ; ; is a sequence part at in , that is, a first half part of the long-term prediction sequence; is a sequence part at in , that is, a second half part of the long-term prediction sequence; is a sequence at ; the short-term prediction sequence output by the first output layer in the Bi-LSTM-Attention network architecture matches the second half part.

[0026] In order to ensure the prediction accuracy of the two output layers, the short-term prediction loss and the long-term prediction loss are respectively used, so that the short-term prediction result of the first output layer is close to the true value, the accuracy of the short-term prediction of the first output layer is ensured, and similarly, the long-term prediction result of the second output layer is close to the true value, the accuracy of the long-term prediction of the second output layer is ensured. Further, there is continuity between the short-term prediction result and the long-term prediction result, the short-term prediction result should be a component of the first half time of the long-term prediction result, that is, ​​and the short-term prediction result should be the latter part of the long-term prediction result In order to maintain the continuity between the short-term prediction result and the long-term prediction result, the continuity loss of short-term and long-term prediction is adopted, so as to achieve the time sequence continuity, thereby avoiding the time sequence discontinuity or contradiction between the short-term prediction result and the long-term prediction result, and affecting the final long-term trend analysis and short-term fluctuation analysis result.

[0027] The analysis method of the long-term trend of the mental health of the measured object includes: The long-term prediction sequence of the health level assessment is obtained by using the least square method The linear regression analysis is performed is converted into 2m data points The linear fitting model is constructed , wherein, is the slope, and b is the intercept; The residual sum of squares minimization objective function of a and b is established The partial derivatives of a and b are obtained by using S respectively, and are equal to 0, and the normal equation group is obtained by arranging: ; ; ; The analytical solution of a and b is obtained from the normal equation group, and the mental health of the measured object is determined by using ; When , the long-term trend of the mental health of the measured object is in a deteriorating trend; When , the long-term trend of the mental health of the measured object is in a good trend; When , the long-term trend of the mental health of the measured object is in a stable trend.

[0028] Other trend analysis methods (such as symbol-based trend judgment or more complex time sequence model) can also be adopted, but the current method can effectively reflect the overall change direction.

[0029] The analysis method of the short-term fluctuation of the mental health of the measured object includes: The standard deviation of the short-term prediction sequence of the health level assessment is calculated , is the mean value of ; When is greater than a preset threshold value, the short-term fluctuation of the mental health of the measured object is severe; When​​ less than or equal to a preset threshold, the short-term fluctuation of the mental health of the measured object is gentle; In the health level evaluation short-term prediction sequence extreme points are marked as short-term peaks and short-term troughs, such as positions of adjacent points in the sequence that have a level jump (such as directly changing from 1 to 3, or 4 to 2), which are used to locate points of sharp changes in mental health status.

[0030] After obtaining the long-term prediction result, the present application performs linear fitting on it using the least square method, and the slope of the linear fitting straight line is used to master the trend. If the slope is greater than 0, it means that the value corresponding to the mental health level has an increasing trend in the long term, and corresponding to the health, mild, moderate and severe levels, it is gradually developing towards the severe direction, that is, the mental condition is deteriorating. Similarly, if the slope is less than 0, it means that the value corresponding to the mental health level has a decreasing trend in the long term, and corresponding to the health, mild, moderate and severe levels, it is gradually developing towards the health direction, that is, the mental condition is improving.

[0031] After obtaining the short-term prediction result, the present application calculates the short-term prediction sequence using the standard deviation to judge the oscillation of the short-term prediction sequence. The greater the oscillation, the more severe the mental health of the measured object in the short term, or the more severe the emotional fluctuation. For example, emotional oscillation, instability and sharp changes within a few days can be captured, and the short-term fluctuation peaks and troughs can be identified, so as to locate whether the concern is associated with specific events, such as acute stress events, medication or certain mental states (such as cyclic episodes of borderline personality disorder and bipolar affective disorder).

[0032] The Bi-LSTM network is composed of a forward LSTM network and a reverse LSTM network. The forward LSTM network is used to obtain the forward mental health level prediction result output at each time point, and the reverse LSTM network is used to obtain the reverse mental health level prediction result output at each time point. The output result formed by splicing the mental health level prediction result outputs of the two directions is used as the final mental health level prediction result output at each time point; The structure expression of the Bi-LSTM network is: ; In the formula, 、 、 、 respectively, the forward propagation hidden layer state, the reverse propagation hidden layer state, the input mental health state sequence and the hidden state of the tth evaluation time point, is an activation function of the hidden layer, 、 , , , , , , , , weight matrix of the t-th evaluation time point, , forward propagation hidden layer state and backward propagation hidden layer state of the t-1-th evaluation time point, , bias of the forward propagation hidden layer and the backward propagation hidden layer, is a vector concatenation symbol.

[0033] In the process of filling in the PHQ-9 / GAD-7 scale for self-evaluation by the measured object (filling in through an intelligent terminal, as shown in Figure 3 , the camera on the intelligent terminal collects the expression data of the measured object, and the physiological data is collected through a wearable device; The method for obtaining the mental health level evaluation result comprises: In the process of filling in the PHQ-9 / GAD-7 scale for self-evaluation by the measured object, the expression data and the physiological data of the measured object are collected; The self-evaluation result of the PHQ-9 / GAD-7 scale of the measured object, the emotion recognition result and the physiological data are input into a pre-established classifier, and the mental health level evaluation result is output.

[0034] The physiological data comprises heart rate HR, blood oxygen SpO2, respiratory rate RR and heart rate variability HRV.

[0035] The method for constructing the emotion classification model comprises: The expression image is taken as input, and the emotion category is taken as output, a classification network is trained to obtain an emotion classification model, and a structure expression of the emotion classification model is: ; In the formula, is an emotion category output by the emotion classification model, is an enhanced expression image, is a classification network; The ResNet network solves the training difficulty problem of a traditional deep network through a residual learning framework and a jump connection (shortcut channel). Unlike the traditional network which directly learns the mapping from input to output, the ResNet divides the complex function into multiple residual mappings, and directly transmits information by bypassing part of the network layer through the jump connection, thereby reducing the training difficulty, and is widely used in computer vision tasks such as image classification, detection, etc.

[0036] The commonly used classification network in the application can also adopt a VGG network (Visual Geometry Group Network, the entire network is composed of alternating convolutional layers and pooling layers, finally connected with a full connection layer and a Softmax classifier, and the feature extraction capability of the model can be improved by stacking a deeper network, and due to the powerful feature extraction capability, the VGG network is commonly used for image classification), and a DenseNet (Dense Convolutional Network, by establishing dense connection between all previous layers and subsequent layers, each layer receives the feature map of all previous layers as input, and feature reuse is realized through channel dimension splicing. This design reduces the number of parameters and computing cost, enhances gradient propagation efficiency, and is widely used in computer vision tasks such as image classification).

[0037] The emotion category is taken as the emotion recognition result.

[0038] The emotion category includes anger, disgust, fear, happy, normal, sad, and surprised. The method for constructing the mental health assessment model comprises the following steps: The mental health self-evaluation result, the emotion recognition result, and the physiological data are taken as inputs, and the mental health grade is taken as output, a classifier is trained, and a mental health assessment model is obtained. The structure expression of the mental health assessment model is as follows: ; In the formula, The mental health grade output by the mental health assessment model, The mental health self-evaluation result, The emotion category, All physiological data collected by the measured object in the mental health self-evaluation process, The support vector machine can also be replaced by a classifier model such as a decision tree, a random forest, etc.

[0039] The Bi-LSTM network combined with the attention mechanism is used to form a Bi-LSTM-Attention network architecture in the application, long-term and short-term results of the mental health grade evaluation of the measured object are realized, long-term trends and short-term volatility of the mental health of the measured object are analyzed, and long-term trends and short-term fluctuations of the mental state of the patient are pre-mastered and continuously concerned.

[0040] The above examples are only exemplary embodiments of the present application, and are not intended to limit the present application, and the protection scope of the present application is defined by the claims. Those skilled in the art can make various modifications or equivalent replacements to the present application within the spirit and protection scope of the present application, and such modifications or equivalent replacements are also considered to fall within the protection scope of the present application.

Claims

1. A method for trend analysis of mental health status based on an AI model, characterized in that, The method comprises the following steps: sequentially recording the mental health level evaluation results of the same test object at each evaluation time point to form a mental health state sequence; a Bi-LSTM-Attention network architecture is constructed by using a Bi-LSTM network combined with an attention mechanism, and the mental health trend prediction model for predicting long-term and short-term results of the mental health level evaluation of the test object is obtained based on the training of the mental health state sequence; the long-term trend and short-term volatility of the mental health of the test object are analyzed according to the long-term and short-term results of the mental health of the test object output by the mental health trend prediction model through least square method and standard deviation calculation, respectively.

2. The AI model-based mental health status trend analysis method of claim 1, wherein: The mental health level evaluation result includes health, mild, moderate and severe, and the health, mild, moderate and severe are converted into numerical expressions 0, 1, 2 and 3 respectively to form a mental health state sequence wherein, , is the mental health level evaluation result at the tth evaluation time point, and n is the total number of evaluation time points.

3. The AI model-based mental health status trend analysis method of claim 2, wherein: The construction method of the mental health trend prediction model comprises: a sequence of mental health states as input items for the Bi-LSTM network layer in the Bi-LSTM-Attention network architecture, the hidden state vectors output by the Bi-LSTM network layer , for each assessment time point; The hidden state vector As the input of the attention layer in the Bi-LSTM-Attention network architecture, the hidden state vector corresponds to the attention weight vector , and the context vector is formed by weighted summation of the attention weight vector and the hidden state vector , is the attention weight of . concatenation result of the context vector and the hidden state vector As input to the Bi-LSTM-Attention network architecture, the short-term prediction sequence of health class assessment is output by the first output layer The long-term prediction sequence of health class assessment is output by the second output layer .​ 4. The AI model-based mental health status trend analysis method of claim 3, wherein: The loss function of the Bi-LSTM-Attention network architecture ; ； ； ； In the formula, is a short-term prediction loss, is a long-term prediction loss, is a continuity loss of the short-term and long-term predictions, is a mean square error operation formula; short-term prediction sequences for health level assessment, ground truth for short-term prediction sequences for health level assessment, long-term prediction sequences for health level assessment, ground truth for long-term prediction sequences for health level assessment, for a sequence portion in the middle corresponding to a time sequence, a sequence portion in the middle corresponding to a time sequence, for inputting the prediction sequence output by the first output layer into the Bi-LSTM-Attention network architecture; for a sequence portion in the middle corresponding to a time sequence, a sequence portion in the middle corresponding to a time sequence.

5. The AI model-based mental health status trend analysis method of claim 4, wherein: The analysis method of the long-term trend of the mental health of the test object comprises: Least squares method is used to evaluate long-term prediction sequence of health rating Linear regression analysis is performed to convert into 2m data points Linear fitting model is constructed wherein, is the slope and b is the intercept; Establishing solving and the residual sum of squares of b minimization objective function , respectively, using S on and b partial derivative, and let it equal to 0, and get the normal equations: ； ； The analytical solution of a and b is obtained from the normal equations and the analytical solution of a and b is obtained from the normal equations determining long-term trends in mental health of the subject; When the long-term trend of the mental health of the measured object is a deteriorating trend; When the long-term trend of the mental health of the measured object is a good trend; When the long-term trend of the mental health of the measured subject is stable.

6. The AI model-based mental health status trend analysis method of claim 4, wherein: The analysis method of the short-term volatility of the mental health of the test object comprises: Computing health grade assessment short-term prediction sequences of the standard deviation , is the mean value of ​ When greater than a preset threshold, the short-term fluctuation of the mental health of the measured object is intense; When If the value is less than or equal to a preset threshold, the short-term fluctuation of the mental health of the measured object is gentle. In the short-term prediction sequence of the health level assessment The extreme points are marked as short-term peaks and short-term troughs.

7. The AI model-based mental health status trend analysis method of claim 1, wherein: The method for obtaining the mental health level evaluation results comprises: In the process of filling out the PHQ-9 / GAD-7 scale by the test object for self-evaluation, the expression data and physiological data of the test object are collected; the self-evaluation results of the PHQ-9 / GAD-7 scale, the emotion recognition results and the physiological data of the test object are input into the pre-established classifier to output the mental health level evaluation results.

8. The AI model-based mental health status trend analysis method of claim 3, wherein: The Bi-LSTM network is composed of a forward LSTM network and a reverse LSTM network, the forward LSTM network is used to obtain the forward sequence mental health level prediction result output at each time point, the reverse LSTM network is used to obtain the reverse sequence mental health level prediction result output at each time point, and the output results of the mental health level prediction results of the two directions are spliced to form the final mental health level prediction result output at each time point; The structure expression of the Bi-LSTM network is: ； wherein , , , are the forward-propagated hidden layer state, the backward-propagated hidden layer state, the input sequence of mental health states, the hidden state at the tthevaluation time point, respectively, is an activation function of the hidden layer, , , , are weight matrices of in , in , , , , are the forward-propagated hidden layer state and the backward-propagated hidden layer state at the t-1thevaluation time point, respectively, , are the bias of the forward-propagated hidden layer and the backward-propagated hidden layer, respectively, is a vector concatenation symbol.

9. The AI model-based mental health status trend analysis method of claim 7, wherein: The physiological data comprises heart rate HR, blood oxygen SpO2, respiratory rate RR and heart rate variability HRV.

10. The AI model-based mental health status trend analysis method of claim 7, wherein: The emotion recognition results comprise anger, disgust, fear, happy, normal, sad and surprised.

Images