A method and device for identifying individual stress state based on physiological and psychological data

Abstract

The present invention provides a method and device for identifying individual stress states based on physiological and psychological data, belonging to the field of stress identification technology. The method includes collecting skin electricity data, pulse wave data and subjective stress psychological data under stress state; processing the above data to obtain physiological characteristic data and subjective stress state characteristics respectively; using Pearson correlation analysis to evaluate the correlation between physiological characteristic data and subjective stress state characteristics, screening key physiological characteristic data, forming a data set for modeling, using Hyperopt tool to screen model structure and optimize hyperparameter combination, automatically screening multiple machine learning algorithms, and determining training models; using training data set to train training model, and using the trained model to predict test data set. The method of the present invention can establish a real-time recognition model of stress state with high robustness, and realize effective detection of individual stress state in different scenarios.

Description

Technical Field

[0001] The present invention belongs to the technical field of stress recognition, and in particular relates to a method and device for identifying an individual stress state based on physiological and psychological data. Background Art

[0002] Stress refers to a state in which an individual's balance is threatened or perceived to be threatened. When an individual perceives stressful stimuli (i.e. encounters a stressor) physiologically or psychologically, their sympathetic nervous system and sympathetic adrenal medullary system will be activated, resulting in a series of fast / slow reactions (such as accelerated heart rate, increased cortisol secretion, increased arousal, and other physical and mental reactions), thereby improving the individual's ability to adapt to the environment. Stress is not only closely related to people's mental health, but also directly affects the individual's work performance in a variety of scenarios. Therefore, accurately identifying an individual's stress state not only helps to avoid dangers at work, but is also one of the important prerequisites for targeted protection of individual mental health.

[0003] Existing stress recognition technologies, for example: CN 115956911 A discloses a detection method based on physiological and psychological state stress, which is mainly limited to specific scenarios and has limited adaptability to complex, natural, and daily use scenarios. For another example, CN109145890A discloses a model for intelligent identification of driver psychological stress intensity based on single-modal physiological signals, and CN117056786A discloses a non-contact stress state identification method and system, which uses subjective classification to distinguish the stress state of the target, mainly for the identification of stress state in a specific task. For another example, CN114398959A discloses a post-traumatic stress disorder identification device, method, and storage medium, and selects post-traumatic stress disorder as a classification label. The algorithm is only applicable to group distinction rather than the real-time state of the target individual.

[0004] At present, the individual stress state recognition technology has the following shortcomings:

[0005] 1) The stress state is mainly based on the subjective classification of the collector, which not only ignores the real psychological feelings of the collector, but also has limited validity in reflecting the individual's stress state;

[0006] 2) Lack of real-time identification of the subject's stress level;

[0007] 3) Existing algorithms are mainly limited to specific scenarios and have limited adaptability to complex, natural, and daily usage scenarios;

[0008] 4) Most algorithms rely on single-modal data. A single data source is easily affected by external noise, which in turn affects the overall accuracy of the algorithm. Summary of the invention

[0009] In view of this, the technical problem to be solved by the present invention is to provide a method for identifying individual stress states based on physiological and psychological data, establish a real-time recognition model of stress states with high robustness, and realize effective detection of individual stress states in different scenarios.

[0010] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0011] A method for identifying individual stress states based on physiological and psychological data, comprising:

[0012] S1: Collect skin electricity data, pulse wave data and subjective stress psychological data under stress state, and mark the data collection stage by combining active marking and preset marking;

[0013] S2: processing the skin electricity data and pulse wave data respectively to obtain a plurality of physiological characteristic data, and processing the subjective stress psychology data to obtain subjective stress state characteristics;

[0014] S3: using Pearson correlation analysis to evaluate the correlation between the physiological characteristic data and the subjective stress state characteristics, and screening out key physiological characteristic data;

[0015] S4: The key physiological characteristic data are used as training characteristic data, the subjective stress state characteristics are used as label data, the training characteristic data and the label data constitute a data set for modeling, the data set for modeling is randomly divided into a training data set and a validation data set, the Hyperopt tool is used to screen the model structure and optimize the hyperparameter combination, Gaussian process regression is selected as the internal optimizer, the model effect is evaluated by the cross-validation method, a variety of machine learning algorithms are automatically screened, and the training model is determined;

[0016] S5: randomly split the modeling data set into a training data set and a test data set again, train the training model using the training data set, predict the test data set using the trained model, and output the prediction result.

[0017] Optionally, collecting skin electrical data, pulse wave data and subjective stress psychological data under stress in S1 includes: obtaining skin electrical data, pulse wave data and subjective stress psychological data of the subject by completing a stress test;

[0018] The stress test includes: social pressure induced stress test and cognitive pressure induced stress test.

[0019] The subjective stress psychological data were collected using a VAS scale with a score of 0-10.

[0020] Optionally, the processing of the pulse wave data in S2 includes:

[0021] S20: performing band-pass filtering on the pulse wave data to remove data higher than 3.5 Hz and lower than 0.7 Hz;

[0022] S21: resampling the data obtained in S20;

[0023] S22: Obtain a crossover point threshold, record the data above the threshold as a peak point, calculate the number of peaks, and obtain HR;

[0024] Calculate the time difference between two adjacent peak points to get the NN interval;

[0025] Based on the NN intervals, HRV features are calculated.

[0026] Optionally, the HRV features in S2 include: SDNN, RMSSD and NN50.

[0027] Optionally, the skin electrical data is processed in S2, including: extracting a tonic component r, a phase component t, and an independent and identically distributed zero-mean Gaussian noise term ε from the skin conductivity y, where: y=r+t+ε.

[0028] Optionally, the subjective stress psychological data is processed in S2, including: re-dividing the collected subjective VAS scale values ​​to obtain different stress level labels, wherein: 0~2 is divided into label 0, corresponding to low stress; 3~4 is divided into label 1, corresponding to medium stress; 5~10 is divided into label 2, corresponding to high stress.

[0029] Optionally, the machine learning algorithm in S4 includes a support vector machine, a K-nearest neighbor classification algorithm, a multilayer perceptron or a random forest.

[0030] Optionally, the training model is a multi-layer perceptron model, and the hyperparameter combination screened by the Hyperopt tool is: epochs=2000, batchsize=32, adam algorithm optimizes the learning rate, the initial learning rate is 0.001, and the regularization parameter value is 1e-3.

[0031] An individual stress state recognition device based on physiological and psychological data, comprising:

[0032] A data acquisition module is used to collect skin electricity data, pulse wave data and subjective stress psychology data under stress state;

[0033] A data processing module, used to process the skin electricity data and the pulse wave data respectively to obtain a plurality of physiological characteristic data, and to process the subjective stress psychology data to obtain subjective stress state characteristics;

[0034] A data analysis module, used for evaluating the correlation between the physiological characteristic data and the subjective stress state characteristics by using Pearson correlation analysis, and retaining the physiological characteristic data if there is a significant correlation, and removing the physiological characteristic data if there is no significant correlation;

[0035] A model screening and optimization module, which is used to form the retained physiological characteristic data into training characteristic data, use the subjective stress state characteristics as label data, the training characteristic data and the label data constitute a data set for modeling, randomly divide the data set for modeling into a training data set and a validation data set, use the Hyperopt tool to perform model structure screening and hyperparameter combination optimization, select Gaussian process regression as the internal optimizer, use the cross-validation method to evaluate the model effect, automatically screen a variety of machine learning algorithms, and determine the training model;

[0036] The model training and prediction module is used to randomly re-split the modeling data set into a training data set and a test data set, use the training data set to train the training model, use the trained model to predict the test data set, and output the prediction result.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] Based on the above technical solution, the present invention provides an individual stress state identification method based on physiological and psychological data, establishes a real-time identification model of stress state with high robustness, and realizes effective detection of individual stress state in different scenarios. The present invention induces individual stress state with ecological validity through stress induction technology combining cognitive induction and social induction; extracts multi-index feature data of heart rate, heart rate variability, and skin electricity and psychological state characteristics with subjective feelings based on the subjective psychological state and objective physiological data during the induction process; explores the mapping relationship between multi-index feature data and individual stress state through Pearson correlation analysis method, and screens key features; uses Hyperopt tool to test and screen model architecture and hyperparameters; further trains based on the best model architecture and hyperparameters to obtain the optimal model, thereby improving the accuracy of real data recognition.

[0039] To avoid the problem of overfitting / underfitting, it is particularly important to select a model architecture and hyperparameters suitable for training data. However, model architecture and hyperparameter combination have nearly infinite combinations. In order to efficiently find their best combination and improve the learning and prediction capabilities of the model to data, the present invention uses Hyperopt tools to screen model structure and hyperparameter combination. Hyperopt function mainly uses the method of Gaussian process regression (GP) when searching for hyperparameter combination. This method regards the optimal hyperparameter combination as an optimal function, and gradually approaches the optimal function through the distribution results of the model test based on random hyperparameters each time. This method has the advantages of high efficiency, strong anti-overfitting / underfitting ability and strong anti-noise ability.

[0040] The real-time stress state recognition model constructed by the present invention is highly robust, with an accuracy of more than 70%, a precision of more than 85%, a recall rate of more than 60%, and an f1 score of more than 70%. It is expected to be used in the fields of individual state detection, talent assessment, etc.

[0041] The technical solution of the present invention is highly transferable and applicable. The model trained under this method is not only robust across scenarios, but also can maintain a stable recognition effect between physiological signals from different sources, thereby achieving effective detection of individual stress states in the scene and reducing the possibility of danger caused by human factors in production and life. DETAILED DESCRIPTION

[0042] In order to better understand the present invention, the content of the present invention is further clearly set forth in conjunction with the embodiments below, but the protection content of the present invention is not limited to the following embodiments. In the following description, a large number of specific details are provided in order to provide a more thorough understanding of the present invention. However, it is obvious to those skilled in the art that the present invention can be implemented without one or more of these details.

[0043] A method for identifying individual stress states based on physiological and psychological data, comprising:

[0044] Collect skin electricity data, pulse wave data and subjective stress psychological data under stress state, and mark the data collection stage by combining active marking and preset marking;

[0045] The skin electricity data and pulse wave data are processed respectively to obtain a plurality of physiological characteristic data, and the subjective stress psychology data are processed to obtain the subjective stress state characteristics;

[0046] Pearson correlation analysis of SPSS software was used to evaluate the correlation between physiological characteristic data and subjective stress state characteristics. If there was a significant correlation, the physiological characteristic data was retained; if there was no significant correlation, the physiological characteristic data was removed, thereby screening out key physiological characteristic data;

[0047] The key physiological characteristic data are used as training characteristic data, and the subjective stress state characteristics are used as label data. The training characteristic data and label data constitute the modeling data set. The modeling data set is randomly divided into a training data set and a validation data set in a ratio of 8:2. The Hyperopt tool is used to screen the model structure and optimize the hyperparameter combination. Gaussian process regression is selected as the internal optimizer. The cross-validation method is used to evaluate the model effect. A variety of machine learning algorithms are automatically screened, and the model corresponding to the machine learning algorithm with the highest accuracy is used as the training model.

[0048] The modeling data set is randomly re-split into a training data set and a test data set in a ratio of 8:2. The training data set is used to train the training model, and the trained model is used to predict the test data set, and the prediction results are output.

[0049] Among them, the cross-validation method selects the 5-fold cross-validation method.

[0050] In some specific embodiments of the present invention, collecting skin electrical data, pulse wave data and subjective stress psychological data under stress state includes: obtaining skin electrical data, pulse wave data and subjective stress psychological data of the subject by completing a stress test;

[0051] Stress tests include: social stress induced stress test and cognitive stress induced stress test.

[0052] Subjective stress psychological data were collected using a VAS scale with a score of 0-10.

[0053] In order to achieve accurate collection of the psychological state and physiological fluctuation data of the target individual and improve the real-time recognition effect of the final model on the individual stress state in real life, the present invention adopts the localized adaptation of the Trier social stress test paradigm. The entire Trier social stress test paradigm consists of three stages (preparation stage, social induction stage, and emotion induction stage), totaling 15 minutes.

[0054] Preparation phase (5 minutes): In the preparation room, the subjects will be told that the next task is to prepare a speech to defend themselves. During the preparation, the subjects can use paper and pen to prepare, but these notes will be taken away during the speech. They are told that the entire speech process will be recorded by a camera for later analysis.

[0055] Social induction phase (5 minutes): The data collectors took the individuals to a lecture room where two store managers and one security guard (all pretending to be data collectors) were sitting and asked them to give a 5-minute defense speech.

[0056] Cognitive induction phase (5 minutes): The subjects are required to perform mental arithmetic for 5 minutes, reporting loudly the result of 102 minus 13 continuously. If the subjects make a mistake, they will be asked to start over.

[0057] During social induction and cognitive induction, three data collectors who played the roles of store managers and security guards wore white coats, maintained neutral facial expressions, and did not give any feedback to the subjects. The subjects needed to stand in front of the microphone and complete the speech and mental arithmetic tasks in front of these three people. During the process, a camera was used to record the audio and video information of the subjects in the stress task. While the subjects were performing the tasks, the skin electricity and pulse wave data were collected through the scientific research bracelet, and the VAS scale with a score of 0-10 was used to collect the subjective psychological stress state.

[0058] In order to ensure the accuracy of cross-modal feature data extraction, marking points are preset at different stages of the task process. At the same time, data collectors will also mark special situations according to the progress of the experiment, that is, ensure control of the collection process through active recording and preset marking points.

[0059] In some specific embodiments of the present invention, the pulse wave data is processed, including:

[0060] Band-pass filtering was performed on the pulse wave data to remove data above 3.5 Hz and below 0.7 Hz;

[0061] Resample the filtered data to 100Hz;

[0062] The moving average value was calculated with a window of 0.75s as the crossing point threshold. The data above the threshold was recorded as the peak point, and the number of peaks was calculated to obtain the HR. The time difference between two adjacent peak points was calculated to obtain the NN interval. Based on the NN interval, the HRV characteristics were calculated.

[0063] More specifically, the pulse wave data is processed, including:

[0064] The filter_signal function in the heartpy package based on Python was used to perform bandpass filtering on the skin electrical data to remove data above 3.5 Hz and below 0.7 Hz;

[0065] The signal function of scipy based on python is used to resample the filtered data and increase it to 100Hz;

[0066] The process function of the heartpy package was used to calculate the moving average with a window of 0.75s as the crossing point threshold. The data above the threshold was recorded as the peak point. The number of peaks was calculated using the built-in count function in Python to obtain the HR. The time difference between two adjacent peak points was calculated to obtain the NN interval, and the HRV characteristics were calculated based on the NN interval.

[0067] In some specific embodiments of the present invention, the HRV features include: SDNN, RMSSD and NN50.

[0068] a. SDNN (Standard Deviation of NN Intervals), in ms. SDNN is defined as the standard deviation of all NN intervals within the detection time. Its calculation formula is as follows:

[0069] .

[0070] b. RMSSD (square root of the mean square of the NN interval difference). RMSSD reflects the shape of the NN interval difference histogram. When the vagus nerve tension is high, the NN interval difference histogram is low and wide, and the RMSSD is large. Therefore, it can be used as an indicator to measure the regulation of the vagus nerve on the heart rate. The calculation formula is as follows:

[0071] .

[0072] c. NN50 (the number of heart beats with a difference of >50 ms between adjacent NN intervals), measured in heart beats.

[0073] In some specific embodiments of the present invention, the skin electrical data is processed, including: extracting the tonic component r, the phase component t, and the independent and identically distributed zero mean Gaussian noise term ε from the skin conductivity y, where: y=r+t+ε. The skin electrical data is extracted using the Matlab-based CvxEDA function, which is based on a convex optimization model and has the advantage of accurately decomposing the skin electrical signals that may overlap. Based on this function, the tonic component (t), the phase component (r), and the independent and identically distributed zero mean Gaussian noise term (ε) can be extracted from the skin conductivity (y) in a certain period of time.

[0074] In some specific embodiments of the present invention, the key physiological characteristic data include HR, SDNN, NN50, and the tonic component (t), phase component (r) and independent and identically distributed zero-mean Gaussian noise term (ε) of skin electricity.

[0075] In some specific embodiments of the present invention, the subjective stress psychological data is processed, including: re-dividing the collected subjective VAS scale values ​​to obtain different stress level labels, wherein: 0~2 is divided into label 0, corresponding to low stress; 3~4 is divided into label 1, corresponding to medium stress; 5~10 is divided into label 2, corresponding to high stress.

[0076] In some specific embodiments of the present invention, the machine learning algorithm includes a support vector machine, a K-nearest neighbor classification algorithm, a multi-layer perceptron or a random forest.

[0077] The present invention also provides an individual stress state identification device based on physiological and psychological data, comprising:

[0078] A data acquisition module is used to collect skin electricity data, pulse wave data and subjective stress psychology data under stress state;

[0079] A data processing module, used to process the skin electricity data and the pulse wave data respectively to obtain a plurality of physiological characteristic data, and to process the subjective stress psychology data to obtain subjective stress state characteristics;

[0080] A data analysis module, for evaluating the correlation between the physiological characteristic data and the subjective stress state characteristics by using Pearson correlation analysis of SPSS software, and retaining the physiological characteristic data if there is a significant correlation, and removing the physiological characteristic data if there is no significant correlation;

[0081] A model screening and optimization module is used to form the retained physiological characteristic data into training characteristic data, use the subjective stress state characteristics as label data, the training characteristic data and label data constitute a modeling data set, randomly divide the modeling data set into a training data set and a validation data set in a ratio of 8:2, use the Hyperopt tool to perform model structure screening and hyperparameter combination optimization, select Gaussian process regression as the internal optimizer, evaluate the model effect with 5-fold cross-validation results, automatically screen a variety of machine learning algorithms, and use the model corresponding to the machine learning algorithm with the highest accuracy as the training model;

[0082] The model training and prediction module is used to randomly divide the modeling data set into a training data set and a test data set in a ratio of 8:2, use the training data set to train the training model, use the trained model to predict the test data set, and output the prediction result.

[0083] The present invention has accumulated representative stressful events in life through preliminary research, literature review, interviews and other techniques, and refined classic stressful scenes and induction paradigms. Further comparison with similar induction paradigms such as MIST and MAST shows that the adapted Trier social stress test paradigm used in the present invention has advantages in inducing stressful conditions, such as obvious induction and long duration, and the scheme is similar to real stressful events in daily life, and thus has higher ecological validity.

[0084] To ensure the extraction of data features, the present invention adopts a setting method combining active marking and preset marking to accurately mark the data collection stage, thereby ensuring the accuracy of data extraction. In terms of data synchronization, due to the different sampling rates of different modal data, the present invention also performs downsampling processing on the signal after signal denoising, making it convenient for subsequent data processing.

[0085] To avoid the problem of overfitting / underfitting, it is particularly important to be suitable for the model architecture and hyperparameters of training data. However, model architecture and hyperparameter combination have nearly infinite combinations. In order to efficiently find their best combination and improve the learning and prediction capabilities of the model to data, the present invention uses the Hyperopt function package based on python code to screen the model structure and hyperparameter combination. Hyperopt function mainly uses the method of Gaussian process regression (GP) when searching for hyperparameter combination. This method regards the optimal hyperparameter combination as an optimal function, and gradually approaches the optimal function through the distribution results of the model test based on random hyperparameters each time. This method has the advantages of high efficiency, strong anti-overfitting / underfitting ability and strong anti-noise ability.

[0086] Case 1: Model structure and hyperparameter optimization based on Hyperopt tool

[0087] In this case, a total of 131 healthy subjects were recruited to participate in the experiment, including 65 males and 66 females, with an age range of 17 to 28 years old and an average age of 22.46 ± 2.47 years old. The subjects participated in the modified Trier social stress test paradigm, collected skin electricity and pulse wave data through scientific research bracelets, and used a VAS scale with a score of 0-10 to collect the subjective psychological stress state of the subjects during the task. In order to ensure the accuracy of cross-modal feature data extraction, while setting preset marking points during the task, data collectors also need to record marking points for special situations according to the progress of the experiment.

[0088] For the pulse wave data, the filter_signal function in the heartpy package based on pyhton was used to perform bandpass filtering on the data to remove data above 3.5Hz (above 210 BPM) and below 0.7Hz (below 42BPM); then the signal function based on scipy in python was used to resample it and increase it to 100Hz; then the process function of the heartpy package was used to calculate the moving average as the crossing point threshold (adaptive threshold) with a window of 0.75s, and the data above the threshold was recorded as the peak point; the number of peaks was calculated by the built-in count function in python to obtain the heart rate (HR); the time difference between two adjacent peak points was calculated to obtain the NN interval; based on the NN interval, the heart rate variability (HRV) features were calculated, including: SDNN, RMSSD, and NN50.

[0089] For the skin electrical data, the CvxEDA function based on Matlab is used to extract the skin electrical data. This function is based on a convex optimization model. Its advantage is that it can accurately decompose the skin electrical signals that may overlap. Based on this function, the tonic component (t), phase component (r) and independent and identically distributed zero-mean Gaussian noise term (ε) can be extracted from the skin conductivity (y) over a certain period of time.

[0090] The subjective stress VAS scale collected in the experiment was reclassified to obtain different stress level labels, that is, subjective stress state characteristics. 0-2 is divided into low stress (label 0), 3-4 is divided into medium stress (label 1), and 5-10 is divided into high stress (label 2).

[0091] Next, a Pearson correlation analysis was performed on the processed pulse wave data and the extracted skin electricity data with the subjective stress state characteristics. It was found that HR, SDNN, NN50, as well as the tonic component (t), phase component (r) and independent and identically distributed zero-mean Gaussian noise term (ε) of skin electricity were significantly correlated (p<0.05). Therefore, these features were selected to form the data set for modeling.

[0092] Before modeling, Hyperopt tool was used to screen the model structure and optimize the hyperparameter combination. Gaussian process regression (GP) was selected as the internal optimizer to improve the model recognition accuracy. The model effect was evaluated by 5-fold cross-validation results. Support Vector Machine (SVM), K-NearestNeighbor (KNN) algorithm, Multilayer Perceptron (MLP) and RandomForest (RF) were automatically screened 3000 times. The ratio of training data set to validation data set in the test process was 8:2.

[0093] The model screening results are shown in the following table:

[0094]

[0095] *Note: The accuracy in the table is the best cross-validation effect when screening each machine learning model, and does not represent the accuracy of the final model test.

[0096] It can be found from the table that the MLP algorithm has the highest accuracy, so MLP is selected as the final training model. The hyperparameter combination screened by Hyperopt is: epochs=2000, batchsize=32, adam algorithm optimizes the learning rate, the initial learning rate is 0.001, and the regularization parameter is 1e-3.

[0097] The data set for modeling was split into training data set and validation data set in a ratio of 8:2. The training data set was then input into the MLP model and hyperparameter combination for training. The validation data set was used to test the model. The final model achieved an accuracy of 78%, a precision of 89%, a recall rate of 62%, and an f1 score of 73%.

[0098] The above case 1 verifies the effectiveness of the method of the present invention itself, proving that the method of the present invention can stably train a model for identifying individual stress states.

[0099] Case 2: Ecological validity verification based on cross-scenario data

[0100] In this case, data from 16 test subjects were collected, including 5 female test subjects, aged 22.06±0.77 years old. All test subjects were participating in an interview assessment, and data collection was interspersed in their interview process. Therefore, the speech theme in the adapted Trier social stress test paradigm was changed to "You will have 5 minutes to explain your personal characteristics in more detail, and the importance and necessity of these characteristics for the assessment." The rest of the process is the same as Case 1.

[0101] The ecological validity verification process is as follows: input the feature data set processed in this case into the model trained in Case 1, and output the prediction results; compare the output results with the true labels to obtain the ecological validity verification results.

[0102] Based on the above verification method, the ecological validity verification accuracy of Case 2 is 68.75%. Although the verification accuracy value is slightly lower than that of Case 1, this may be due to the difference between the two collection subject groups. The verification result of Case 2 based on the real scene is still much higher than the guessing probability (30%), which proves that the model trained based on the method of the present invention has certain validity in different scenes, confirming that it has high ecological validity and application value.

[0103] The above case 2 starts from a real-life scenario and tests the ecological validity of the model trained based on the method shown in case 1, proving that the method of the present invention has high ecological validity and application value.

[0104] Case 3: Robustness verification based on non-homologous physiological signals

[0105] In order to verify the robustness of the method of the present invention, Case 3 uses a different modeling data source from Case 1, that is, the original extraction of heart rate and heart rate variability data from pulse wave data is changed to extraction of heart rate and heart rate variability based on ECG data. This is to test whether the accuracy of the method of the present invention will drop significantly when comparing physiological signals from different sources, thereby verifying the robustness / versatility of the scheme across physiological signal sources.

[0106] In this case, a total of 90 healthy subjects were recruited to participate in data collection, including 47 males and 43 females, with an age range of 17 to 28 years old. The experiment collected the subjects' skin electrical data, ECG data, and subjective VAS data. The entire implementation process is the same as in Case 1, the only difference is that this case uses the heart rate and heart rate variability features extracted based on ECG data.

[0107] The results of the non-homologous physiological signal modeling test are shown in the following table:

[0108]

[0109] From the results, although non-homologous physiological data were used, the features extracted and the model trained based on the method of the present invention still maintained a high accuracy, thus proving the robustness of the scheme itself.

[0110] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0111] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for identifying individual stress states based on physiological and psychological data, characterized in that: include: S1: Collect skin electricity data, pulse wave data and subjective stress psychological data under stress state, and mark the data collection stage by combining active marking and preset marking; The collecting of skin electrical data, pulse wave data and subjective stress psychological data under stress state includes: obtaining the skin electrical data, pulse wave data and subjective stress psychological data of the subject by completing a stress test; The stress test includes: social pressure induced stress test and cognitive pressure induced stress test. The subjective stress psychological data are collected using a VAS scale with a score of 0-10; S2: processing the skin electricity data and pulse wave data respectively to obtain a plurality of physiological characteristic data, and processing the subjective stress psychology data to obtain subjective stress state characteristics; The processing of the pulse wave data comprises: S20: performing band-pass filtering on the pulse wave data to remove data higher than 3.5 Hz and lower than 0.7 Hz; S21: resampling the data obtained in S20; S22: Obtain the intersection threshold, record the data above the threshold as the peak point, calculate the number of peaks, and obtain HR; calculate the time difference between two adjacent peak points to obtain the NN interval; calculate the HRV feature based on the NN interval; The processing of the skin electrical data includes: extracting a tonic component r, a phase component t, and an independent and identically distributed zero-mean Gaussian noise term ε from the skin conductivity y, wherein: y=r+t+ε; S3: using Pearson correlation analysis to evaluate the correlation between the physiological characteristic data and the subjective stress state characteristics, and screening out key physiological characteristic data; S4: The key physiological characteristic data are used as training characteristic data, the subjective stress state characteristics are used as label data, the training characteristic data and the label data constitute a data set for modeling, the data set for modeling is randomly divided into a training data set and a validation data set, the Hyperopt tool is used to screen the model structure and optimize the hyperparameter combination, Gaussian process regression is selected as the internal optimizer, the model effect is evaluated by the cross-validation method, a variety of machine learning algorithms are automatically screened, and the training model is determined; The machine learning algorithm includes a support vector machine, a K-nearest neighbor classification algorithm, a multilayer perceptron or a random forest; S5: randomly split the modeling data set into a training data set and a test data set again, train the training model using the training data set, predict the test data set using the trained model, and output the prediction result.

2. The method for identifying individual stress status based on physiological and psychological data according to claim 1, characterized in that: The HRV features in S2 include: SDNN, RMSSD and NN50.

3. The method for identifying individual stress status based on physiological and psychological data as claimed in claim 2, characterized in that: The subjective stress psychological data is processed in S2, including: re-dividing the collected values ​​of the subjective VAS scale to obtain different stress level labels, wherein: 0~2 is divided into label 0, corresponding to low stress; 3~4 is divided into label 1, corresponding to medium stress; 5~10 is divided into label 2, corresponding to high stress.

4. A device for identifying individual stress status based on physiological and psychological data, characterized in that: A method for identifying an individual stress state based on physiological and psychological data according to any one of claims 1 to 3, comprising: A data acquisition module is used to collect skin electricity data, pulse wave data and subjective stress psychology data under stress state; A data processing module, used to process the skin electricity data and the pulse wave data respectively to obtain a plurality of physiological characteristic data, and to process the subjective stress psychology data to obtain subjective stress state characteristics; A data analysis module, used for evaluating the correlation between the physiological characteristic data and the subjective stress state characteristics by using Pearson correlation analysis, and retaining the physiological characteristic data if there is a significant correlation, and removing the physiological characteristic data if there is no significant correlation; A model screening and optimization module, which is used to form the retained physiological characteristic data into training characteristic data, use the subjective stress state characteristics as label data, the training characteristic data and the label data constitute a data set for modeling, randomly divide the data set for modeling into a training data set and a validation data set, use the Hyperopt tool to perform model structure screening and hyperparameter combination optimization, select Gaussian process regression as the internal optimizer, use the cross-validation method to evaluate the model effect, automatically screen a variety of machine learning algorithms, and determine the training model; The model training and prediction module is used to randomly divide the data set used for modeling into a training data set and a test data set, use the training data set to train the training model, use the trained model to predict the test data set, and output the prediction result.

Images