Physiological state assessment method, system and equipment and storage medium

Abstract

The invention discloses a physiological state assessment method, system and device and a storage medium, and relates to the technical field of physiological state assessment, and the method comprises the steps: collecting a multi-source physiological signal of a user; calculating the complexity of each physiological signal; setting a threshold standard of signal complexity, decomposing signals with the complexity higher than a threshold by adopting a signal decomposition method, and combining sub-signals obtained after decomposition with undecomposed models; calculating correlation degrees among the signals by adopting a clustering analysis method, screening out the signal with the highest average correlation degree with each signal, and placing the signal with the highest correlation degree with the signal in the residual signals adjacent to the signal to obtain a reordered feature matrix; using a convolutional neural network to extract high-order information features in the reordered feature matrix; evaluating the physiological state of the user according to the high-order features; according to the method, the correlation between the signals is fully utilized, and the physiological state evaluation accuracy of the multi-source physiological signals with complexity is effectively improved.

Description

A method, system, device, and storage medium for assessing physiological state Technical Field

[0001] This invention relates to the field of physiological state assessment technology, specifically to a physiological state assessment method, system, device, and storage medium. Background Technology

[0002] Physiological data contains a wealth of potentially valuable information that plays a crucial role in areas such as telemedicine and interpreting pathological mechanisms. However, due to the inherent uncertainty of each individual's physiological state, even when exhibiting the same complex symptoms, the acquired physiological data can vary significantly. This makes physiological data significantly affected by various individual differences. Therefore, when assessing the physiological state of each individual, it is particularly important to incorporate data from multiple sources.

[0003] In recent years, academia and industry have explored numerous methods for identifying and analyzing physiological data. However, current technologies primarily focus on subject-specific modeling of individual physiological data, while subject-independent modeling of non-individual physiological data in real-world contexts is relatively lacking. Among the limited methods for modeling non-individual physiological data, existing technologies rely too heavily on purely physiological data-driven approaches, neglecting a wealth of valuable individual subject information. This situation limits the widespread application and in-depth development of physiological data modeling, necessitating new methods and approaches to overcome this bottleneck.

[0004] Currently, the identification and analysis of multi-source physiological information data still faces several challenges: First, physiological data exhibits complex characteristics such as nonlinearity, non-stationarity, and non-Gaussianity, making it particularly difficult to accurately capture data dynamics. Second, the inherent complexity of physiological data leads to low feature discrimination. Taking epilepsy as an example, the physiological data features of the preictal and interictal phases are intertwined and complex, making them difficult to distinguish clearly, which poses a significant challenge to the identification of physiological data at different stages. Third, the significant differences in physiological data between individuals limit the widespread application of modeling results based on existing data, making it difficult to ensure the effectiveness and universality of the model. Finally, the diversity and categorization of individual information present numerous obstacles when fusing this information with multi-source physiological data, making it difficult to achieve effective integration and utilization of information.

[0005] In summary, due to the complex characteristics of multi-source physiological data, such as nonlinearity, non-stationarity, and non-Gaussianity, traditional physiological state assessment methods struggle to accurately capture dynamic changes in the data, resulting in low feature discrimination and reduced accuracy in assessing physiological state. Summary of the Invention

[0006] To address the shortcomings of existing technologies in accurately capturing dynamic changes in data, resulting in low feature discrimination and reduced accuracy in assessing physiological states, this invention proposes a physiological state assessment method, system, device, and storage medium. By decomposing highly complex signals into less complex sub-signals, it enables a more comprehensive analysis of the diverse characteristics of individual information and effectively extracts complex features from multi-source physiological data, thereby solving the problems existing in the prior art.

[0007] A physiological state assessment method includes the following steps: collecting multi-source physiological signals from the user under test; calculating the complexity of each signal in the multi-source physiological signals; decomposing signals with complexity higher than a set threshold to obtain sub-signals with complexity lower than the set threshold; merging the decomposed sub-signals with the undecomposed signals to obtain an expanded feature matrix; using cluster analysis to calculate the correlation between signals in the expanded feature matrix, selecting the signal with the highest average correlation with each signal, and placing the signal with the highest correlation with the remaining signal and the signal with the highest average correlation with each signal next to each other to obtain a reordered feature matrix; using a convolutional neural network to extract high-order information features from the reordered feature matrix and concatenating them into a high-order information feature group; calculating the high-order information features that contribute the most to the physiological state assessment using an attention mechanism as key high-order features; and assessing the physiological state of the user under test based on the key high-order features.

[0008] Furthermore, the complexity of each signal in the multi-source physiological signal is calculated using analysis of variance, specifically expressed as follows: ,definition Indicates physiological signals The complexity of each physiological signal is determined; the signals are sorted from highest to lowest complexity to obtain the sorted feature matrix. ,in Representing the Secondary sorting.

[0009] Furthermore, a signal decomposition method is used to decompose signals with complexity exceeding a set threshold, obtaining sub-signals with complexity below the set threshold. The decomposed sub-signals are then merged with the undecomposed signals to obtain an expanded feature matrix. This process includes the following steps: determining the complexity threshold of each physiological signal using an exponential decreasing strategy based on the complexity of each signal. Physiological signals with complexity higher than a complexity threshold are selected; empirical mode decomposition (EMD) is used to decompose these signals into physiological signals with complexity lower than the threshold, represented as follows: ,in This indicates the number of signals after decomposition; for each signal to be decomposed... Find the maximum and minimum points in the signal, and connect the extreme points with spline curves to obtain... and If the complexity of the two signals is less than a set threshold, then the decomposition is stopped. Otherwise, the signals that do not meet the threshold are decomposed again to obtain the decomposed sub-signal groups. The obtained sub-signals are merged to obtain the decomposed signal set. The decomposed signal set is merged with the undecomposed signals to obtain the expanded feature matrix.

[0010] Furthermore, the step of calculating the correlation degree between signals using cluster analysis, selecting the signal with the highest average correlation degree with all other signals, and placing the signal with the highest correlation degree with the selected signal and the signal with the highest average correlation degree with all other signals in the remaining signals adjacent to each other to obtain a reordered feature matrix includes the following steps: calculating the correlation degree matrix between signals in the expanded feature matrix using cluster analysis; selecting the signal with the highest average correlation degree with other signals and placing it first; selecting the signal with the highest correlation degree with the selected signal from the remaining signals and placing it adjacent to the selected signal with the highest average correlation degree with other signals, and so on, to obtain the reordered feature matrix. .

[0011] Furthermore, the step of extracting high-order information features from the reordered feature matrix using a convolutional neural network and concatenating them into a high-order information feature group specifically includes the following steps: selecting convolutional blocks with different kernel sizes. and Perform convolution operations to extract the feature matrix. Higher-order related information is obtained by concatenating the outputs of each convolutional block along the channel dimension. .

[0012] Furthermore, the step of calculating the high-order information features that contribute the most to the physiological state assessment through the attention mechanism as key high-order features specifically includes: A channel attention mechanism is used to enhance the proportion of information that contributes most to physiological state assessment in the high-order information feature set, and a three-dimensional matrix is ​​output. ;Will A spatial attention mechanism is used to enhance the proportion of spatially correlated information in high-order information feature groups, and a three-dimensional matrix is ​​output. , as a key high-order feature.

[0013] Furthermore, after collecting the multi-source physiological signals of the user to be tested, the multi-source physiological signals are preprocessed using an outlier detection method, specifically including the following steps: using the Z-score method to detect and remove outliers and missing values ​​in each physiological signal, and using KNN or nearest neighbor interpolation to complete the data; using the mode frequency as the unified sampling frequency, interpolating and resampling the non-uniform frequency signals; and normalizing the resampled signals to obtain the preprocessed signals.

[0014] The present invention also includes a physiological state assessment system, comprising: a data acquisition module for acquiring multi-source physiological signals from the user under test; a decomposition module for calculating the complexity of each physiological signal; decomposing signals with complexity higher than a set threshold to obtain sub-signals with complexity lower than the set threshold; merging the decomposed sub-signals with the undecomposed signals to obtain an expanded feature matrix; a filtering module for calculating the correlation between signals in the expanded feature matrix using cluster analysis, filtering out the signal with the highest average correlation with each signal, and placing the signal with the highest correlation with the remaining signal and the signal with the highest average correlation with each signal adjacent to each other to obtain a reordered feature matrix; a calculation module for extracting high-order information features from the reordered feature matrix using a convolutional neural network and concatenating them into a high-order information feature group; calculating the high-order information features that contribute the most to the physiological state assessment using an attention mechanism as key high-order features; and an assessment module for assessing the physiological state of the user under test based on the key high-order features.

[0015] The present invention also includes a physiological state assessment computer device, characterized in that it comprises: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the physiological state assessment method.

[0016] The present invention also includes a readable storage medium, characterized in that the readable storage medium stores a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform the steps of the physiological state assessment method.

[0017] This invention provides a method for assessing physiological state, which has the following beneficial effects: The signal complexity decomposition and correlation rearrangement strategy adopted in this invention effectively eliminates the interference of differences in physiological signals between individuals, unifying signals from different sources into a comparable low-complexity subspace; by extracting multiple high-order features through convolutional neural networks, and then using an attention mechanism to intelligently evaluate the contribution of each feature to the current physiological state assessment task, and weighting and filtering accordingly, it can automatically suppress redundancy and noise, amplify key information, and ultimately achieve more discriminative information with fewer features, significantly improving the accuracy of physiological state assessment; This invention not only improves the accuracy of utilizing multi-source physiological information with inherent complexity, but also expands the application of multi-source physiological information on different individuals, making the method scalable and with broad application prospects in the field of physiological state analysis. Attached Figure Description

[0018] Figure 1 is a schematic diagram of the working process of the channel attention mechanism in an embodiment of the present invention; Figure 2 is a schematic diagram of the working process of the spatial attention mechanism in an embodiment of the present invention; Figure 3 is a flowchart of the data fusion algorithm for high-dimensional physiological information in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0020] Physiological data is typically individualized, meaning that its use in medical diagnosis and status assessment is highly susceptible to individual differences. For example, an electrocardiogram (ECG) is a crucial tool for assessing cardiac function; however, ECG waveforms can vary significantly between individuals. Some athletes, due to long-term training, may have slower heart rates and different ECG waveforms than the average person. If doctors are unaware of these individual differences, they might misdiagnose heart disease. Blood pressure is an important indicator of cardiovascular health, but the normal blood pressure range can differ from person to person. Some elderly individuals may have elevated blood pressure due to arteriosclerosis or other reasons, but this does not necessarily mean they have hypertension. Therefore, when assessing blood pressure data, factors such as individual age and health status must be considered.

[0021] This invention proposes a physiological state assessment method, which includes the following steps: S1, acquiring the high-dimensional physiological signals of the user to be tested, and using outlier detection technology to preprocess the raw data to obtain the original feature matrix.

[0022] S1.1, Set the raw physiological signal to be input as follows: ,in Indicates the input number of the first... Physiological signal data sequences. An outlier detection model was used to preprocess the original physiological signal sequences.

[0023] The outlier detection model is the Z-score detection method, and the specific steps are as follows: Calculate the Z-score sequence of each physiological signal: ,in To calculate the mean of a data series, To calculate the standard deviation of a data series, To obtain the absolute value sequence of each element in a data sequence, the outlier criterion is set as follows: ,when If an element is missing (missing value) or exceeds the outlier threshold (outlier), it is removed.

[0024] Missing values ​​are imputed using KNN or nearest neighbor imputation methods to obtain the original feature matrix. .

[0025] S1.2 Analyze the sampling time of each raw signal and set the sampling frequency of each raw physiological signal as follows: Then the sampling frequency of the feature matrix is ,in To calculate the mode of the data, after determining the sampling frequency, for samples with a sampling frequency that is not zero... The signal is updated to its sampling frequency using interpolation and resampling techniques. Then, feature alignment is performed. The original feature matrix can then be represented as... ,in To diagnose cycle time, Indicates in A data point obtained from sampling at any given time.

[0026] S1.3. Normalize each signal data to obtain the normalized feature matrix. ,in , To find the minimum value of a data sequence, To find the maximum value of a data sequence, After data normalization at time The corresponding sampled data. Normalization facilitates the calculation and analysis of signal complexity in S2.

[0027] S2. Using analysis of variance, calculate the complexity of each signal and sort the signals according to their complexity.

[0028] S2.1 Calculate the variance of each signal after normalization. ,use Representative signal Complexity; S2.2, Sort the signals according to their complexity from largest to smallest to obtain the sorted feature matrix. Combining the process in S4, superscript... Representing the Secondary sorting.

[0029] S3. Set a threshold standard and use signal decomposition technology to decompose a signal with high complexity into a sub-signal with lower complexity.

[0030] S3.1, Considering the initial complexity of each signal Design an exponentially decreasing strategy Determine the initial signal complexity threshold Signals with complexity exceeding a threshold are selected. Signals with complexity exceeding a threshold are defined as... ,in This represents the number of signals whose complexity exceeds a threshold. A value of 0 indicates that there is no signal above the threshold;

[0031] S3.2. Utilize Empirical Mode Decomposition (EMD) technology to transform highly complex signals. Decompose into signals with low complexity ,in This indicates the number of signals after decomposition.

[0032] S3.2.1 For each signal to be decomposed Find the maximum and minimum points in the signal, and connect the extreme points with spline curves to obtain... and At this point, we can obtain two decomposed signals, which are respectively and S3.2.2 Calculate whether the two signals satisfy the following conditions: , If the condition is met, the decomposition process stops; otherwise, the unmet condition is re-executed in step S3-2-1 to obtain the final signal. Decomposed sub-signal groups S3.2.3, will After each signal completes the decomposition operations S3.2.1 and S3.2.2, the sub-signals are merged to obtain the final result. .

[0033] S3.3, Decompose the signal With undecomposed signals The features are merged to form the expanded feature matrix. .

[0034] S4. Reset the new threshold and repeat steps S2 and S3. Continue this process until all signals reach a low complexity level.

[0035] S4.1, the exponential decreasing strategy is expressed as: Where THR represents the complexity threshold, THR0 represents the initial complexity threshold, α is a parameter, and k represents the number of executions. Using an exponentially decreasing strategy, when the number of executions is k, it first checks if there is a signal indicating that the complexity exceeds the threshold. If not, the loop ends; if it does, the complexity is decreased according to the exponentially decreasing strategy. Continue executing operations S2 and S3.

[0036] S4.2. Determine the parameter α in S4.1. This involves considering both limiting the number of signals to be decomposed and minimizing the complexity of each signal, aiming for a low complexity level for all signals. Optimize this using the Particle Swarm Optimization (PSO) algorithm to obtain the decomposed feature matrix. ,in This represents the number of characteristic matrix sequences after decomposition.

[0037] S5. Using correlation analysis technology, the relative positions of each signal are rearranged so that strongly correlated features have closer spatial positions.

[0038] S5.1. Using cluster analysis (CA), calculate Correlation matrix between signals ,in, express middle and The correlation degree is calculated as follows: when The closer it gets to 1, and The higher the correlation, the higher the average correlation with other signals should be, and the signal with the highest correlation with other signals should be selected and placed first.

[0039] S5.2 Select the signal with the highest correlation to the selected signal from the remaining signals, and place it next to the previously selected signal. Repeat this process to form the rearranged feature matrix. .

[0040] S6. Using a convolutional network module, extract high-order information features from the data, and concatenate the various feature quantities to obtain a high-order information feature group.

[0041] S6.1, Define the rearranged feature matrix for Two-dimensional matrix, selecting convolutional blocks with different kernel sizes ,and Perform convolution operations to extract higher-order relevant information from the feature matrix, using padding different from the kernel size. Set the stride to 1 so that different convolutional blocks and The size remains unchanged after the convolution operation. The final convolution output is... ,in ,for Two-dimensional matrix.

[0042] S6.2, will By combining them, we obtain higher-order information feature groups. ,for Three-dimensional matrix.

[0043] S7. Utilizing the attention mechanism module, redundant information and noise interference are eliminated to identify key high-order features. Based on specific scenario requirements, data dimensionality reduction and fusion are adaptively performed. The key high-order matrix refers to an abstract set of features highly correlated with the target physiological state, extracted from multi-source physiological signals through a convolutional neural network (CNN) and the attention mechanism. CNN captures the nonlinear relationships between signals through multi-layer convolution operations, while the attention mechanism further calculates the features that contribute most to the physiological state assessment (such as changes in EEG theta wave power and nonlinear components of heart rate variability) through weight allocation. This entire process is performed using a deep learning neural network. Its computational training process requires a certain amount of raw data and corresponding evaluation index data to fit the network. During the fitting process, the network automatically learns and mines information from the data. At this point, the attention mechanism extracts the features most correlated with the evaluation index data; these features are called "key high-order features."

[0044] S7.1, will By employing channel attention, the proportion of useful information in high-order feature groups is enhanced, ultimately outputting a three-dimensional matrix. Its size is .

[0045] S7.1.1, Input matrix Make a copy and set it as .for any ,beg The mean of all elements in the matrix is ​​obtained. Finally obtained mean group .

[0046] S7.1.2, Set the weight parameters as follows Multiply the weight parameters by each element of the mean set to obtain the weighted mean set. .

[0047] S7.1.3 Calculate the sigmoid value of each element in the weighted mean group. The calculation method is as follows: S7.1.4, Combine the calculated output from S7.1.3 with... Multiply to get the final output. .

[0048] S7.2, will By employing spatial attention, the proportion of spatial correlation information in high-order feature sets is enhanced, ultimately outputting a three-dimensional matrix. Its size is .

[0049] S7.2.1, Input matrix Make a copy and set it as .because for A three-dimensional matrix, on the one hand, is to find The mean over the m-dimensional dimension is obtained. On the other hand, seeking The maximum value in the m-dimensional dimension is obtained. ,in , All Two-dimensional matrix.

[0050] S7.2.2, will , Combine to get one A 3D matrix is ​​passed through a convolutional network module with a kernel size of 3, a stride of 1, and padding of 1. Dimensionality reduction of the data is performed, and finally a result is obtained. Two-dimensional matrix .

[0051] S7.2.3, will and Multiplying along the m-dimensional axis yields the final output. .

[0052] S7.3, using the convolution-fully connected module to... Dimensionality reduction is performed to output the final one-dimensional output (fatigue assessment value).

[0053] Using actual collected physiological data from an individual as an example, this study assesses whether the individual is in a state of fatigue. The individual is 25 years old and male. Data collected includes electrocardiogram (ECG), electroencephalogram (EEG), electromyography (EMG), blood pressure, and respiratory data. ECG, EEG, and EMG data were sampled using a physiological electrical signal sensor with a sampling rate of 512 Hz; blood pressure was sampled using a sphygmomanometer with a sampling rate of 1 Hz; and respiration was sampled using a spirometer with a sampling rate of 2 Hz. The total sampling time was 30 minutes, with each diagnostic cycle lasting 60 seconds. The main steps included:

[0054] S1. Acquire the high-dimensional physiological signals to be used, and preprocess the raw data using outlier detection technology to obtain the original feature matrix. High-dimensional physiological signals can be collected through sensors such as electrocardiogram (ECG), electroencephalogram (EEG), electromyography (EMG), blood pressure monitors, respiratory sensors, blood analyzers, and thermometers. Outlier detection technology can be implemented using the NumPy numerical computing library, Pandas data structures and data analysis tools, and the scikit-learn machine learning library in Python.

[0055] S1.1: The raw physiological signal to be input is ,in Indicates the input number of the first... The input physiological signal data sequence is {ECG, EEG, EMG, blood pressure, respiration}. An outlier detection model is used to preprocess the original physiological signal sequences.

[0056] The outlier detection model is the Z-score detection method, and the specific steps are as follows: Calculate the Z-score sequence of each physiological signal: ,in To calculate the mean of a data series, To calculate the standard deviation of a data series, To obtain the absolute value sequence of each element in a data sequence, the outlier criterion is set as follows: ,when If an element is missing (missing value) or exceeds the outlier threshold (outlier), it is removed. The Z-score detection method is an outlier detection method based on statistical principles. It assumes that the data follows a normal distribution and identifies outliers by calculating the distance (in standard deviation) of each data point from the mean. The outlier thresholds for each feature are {0.45, 0.229, 0.334, 1.3, 1.79}.

[0057] Missing values ​​are imputed using KNN or nearest neighbor imputation methods to obtain the original feature matrix. The KNN method, for each missing value, finds the two most similar data points in the dataset (usually based on a distance metric), and then fills the missing value with the mean or median of these two data points.

[0058] S1.2: Analyze the sampling time of each raw signal and set the sampling frequency of each raw physiological signal as follows: If the frequency of the feature matrix is ​​{512Hz, 512Hz, 512Hz, 1Hz, 2Hz}, then the sampling frequency of the feature matrix is... =512Hz, where To calculate the mode of the data, after determining the sampling frequency, for signals with a sampling frequency other than 512Hz (i.e., blood pressure and respiration signals), interpolation and resampling techniques are used to update their sampling frequency to 512Hz for feature alignment. The original feature matrix can then be represented as follows: ,in =60s is the diagnostic cycle time. Indicates in A data point obtained from sampling at any given time.

[0059] S1.3: Normalize each signal data to obtain the normalized feature matrix. ,in , To find the minimum value of a data sequence, To find the maximum value of a data sequence, After data normalization at time The corresponding sampled data. Normalization facilitates the calculation and analysis of signal complexity in S2.

[0060] S2. Using analysis of variance, calculate the complexity of each signal and sort the signals according to their complexity. This can be achieved using NumPy numerical computation.

[0061] S2.1: Calculate the variance of each signal after normalization. ={0.114,0.147,0.246,0.024,0.037}, using Representative signal Complexity.

[0062] S2.2: Sort the signals in descending order of complexity to obtain the sorted feature matrix. The corresponding data are {electromyography, electroencephalography, electrocardiography, respiration, blood pressure}, combined with the process in S4, superscript... Representing the Secondary sorting.

[0063] S3. Set a threshold standard and use signal decomposition technology to decompose a signal with high complexity into a sub-signal with lower complexity.

[0064] S3.1: Considering the initial complexity of each signal Design an exponentially decreasing strategy Determine the initial signal complexity threshold =0.1, select signals with complexity higher than the threshold. Set the signal complexity higher than the threshold as... The corresponding signals are {electromyography, electroencephalography, electrocardiography}, among which... This represents the number of signals whose complexity exceeds a threshold. =3, when A value of 0 indicates that there is no signal above the threshold.

[0065] S3.2: Utilize Empirical Mode Decomposition (EMD) technology to transform highly complex signals. Decompose into signals with low complexity ,in This indicates the number of signals after decomposition. For EMD decomposition, PyEMD is a Python library specifically designed for EMD and its variants, such as ensemble EMD and bimodal EMD. It provides an easy-to-use interface and a variety of options to customize the EMD process.

[0066] S3.2.1: For each signal to be decomposed Find the maximum and minimum points in the signal, and connect the extreme points with spline curves to obtain... and At this point, we can obtain two decomposed signals, which are respectively and .

[0067] S3.2.2: Calculate whether the two signals satisfy... , If the condition is met, the decomposition process stops; otherwise, the unmet condition is re-executed in step S3-2-1 to obtain the final signal. Decomposed sub-signal groups .

[0068] S3.2.3: Will After each signal in the process completes the decomposition operations S3-2-1 and S3-2-2, the sub-signals are merged to obtain the final result. .

[0069] S3.3: Decompose the signal With undecomposed signals The features are merged to form the expanded feature matrix. Electromyography, electroencephalography, and electrocardiography all generated two sub-signals.

[0070] S4. Reset the new threshold and repeat steps S2 and S3. Continue this process until all signals reach a low complexity level.

[0071] S4.1: The exponentially decreasing strategy is expressed as: Where THR represents the complexity threshold, THR0 represents the initial complexity threshold, α is a parameter, and k represents the number of executions. Using an exponentially decreasing strategy, when the number of executions is k, it first checks if there is a signal indicating that the complexity exceeds the threshold. If not, the loop ends; if it does, the complexity is decreased according to the exponentially decreasing strategy. Continue executing operations S2 and S3. During the decomposition of the data in the case study, the loop was executed twice, with THR values ​​of 0.1 and 0.05 respectively, and α = 0.693.

[0072] S4.2: Determine the parameter α from S4.1. This involves considering both limiting the number of signals to be decomposed and minimizing the complexity of each signal, aiming for a low complexity level for all signals. Optimize this using the Particle Swarm Optimization (PSO) algorithm to obtain the decomposed feature matrix. ,in This represents the number of feature matrix sequences after decomposition. Ultimately, the original ECG signal generated 2 sub-signals, the original EEG signal generated 2 sub-signals, and the original EMG signal generated 3 sub-signals. It can be represented as .

[0073] S5. Using correlation analysis technology, the relative positions of each signal are rearranged so that strongly correlated features have closer spatial positions.

[0074] S5.1: Use cluster analysis (CA) to calculate Correlation matrix between signals ,in, express middle and The correlation degree is calculated as follows: when The closer it gets to 1, and The higher the correlation, the higher the average correlation with other signals should be, and the higher that signal should be placed first. Cluster analysis can be performed using the KMeans correlation function in the scikit-learn library.

[0075] S5.2: Select the signal with the highest correlation to the selected signal from the remaining signals, place it next to the previously selected signal, and so on, to form the rearranged feature matrix. .

[0076] S6. Using a convolutional network module, extract high-order information features from the data, and concatenate the various feature values ​​to obtain a high-order information feature set. In Python, PyTorch is a widely used deep learning library that provides all the tools and functions needed to build and train convolutional modules and attention mechanism modules.

[0077] S6.1: Define the rearranged feature matrix for Two-dimensional matrix, selecting convolutional blocks with different kernel sizes In this case study, a total of m=4 convolutional blocks were selected, with kernel sizes of 1, 3, 3, and 5 respectively. The convolutional blocks and... Perform convolution operations to extract higher-order relevant information from the feature matrix, using padding different from the kernel size. ={0,1,1,2}, with a stride of 1, such that different convolutional blocks and The size remains unchanged after the convolution operation. The final convolution output is... ,in ,for Two-dimensional matrix.

[0078] S6.2: Will By combining them, we obtain higher-order information feature groups. ,for Three-dimensional matrix.

[0079] S7. Utilize the attention mechanism module to eliminate redundant information and noise interference, identify key high-order features, and adaptively perform data dimensionality reduction and fusion according to specific scenario requirements.

[0080] S7.1: Will By employing channel attention, the proportion of useful information in high-order feature groups is enhanced, ultimately outputting a three-dimensional matrix. Its size is The structure of the channel attention mechanism is shown in Figure 1.

[0081] S7.1.1: Input matrix Make a copy and set it as .for any ,beg The mean of all elements in the matrix is ​​obtained. Finally obtained mean group .

[0082] S7.1.2: Set the weight parameters as follows Multiply the weight parameters by each element of the mean set to obtain the weighted mean set. .

[0083] S7.1.3: Calculate the sigmoid value of each element in the weighted mean group. The calculation method is as follows: S7.1.4: Combine the calculated output from S7-1-3 with... Multiply to get the final output. .

[0084] S7.2: Will By employing spatial attention, the proportion of spatial correlation information in high-order feature sets is enhanced, ultimately outputting a three-dimensional matrix. Its size is The spatial attention mechanism structure is shown in Figure 2.

[0085] S7.2.1: Input matrix Make a copy and set it as .because for A three-dimensional matrix, on the one hand, is to find The mean over the m-dimensional dimension is obtained. On the other hand, seeking The maximum value in the m-dimensional dimension is obtained. ,in , All Two-dimensional matrix.

[0086] S7.2.2: Will , Combine to get one A 3D matrix is ​​passed through a convolutional network module with a kernel size of 3, a stride of 1, and padding of 1. Dimensionality reduction of the data is performed, and finally a result is obtained. Two-dimensional matrix .

[0087] S7.2.3: Will and Multiplying along the m-dimensional axis yields the final output. .

[0088] S7.3: Using a convolution-fully connected module to... Dimensionality reduction is performed to obtain the output.

[0089] The overall network model structure is shown in Figure 3. By feeding physiological signal data into the network, training it, and optimizing the model parameters, a multi-source physiological signal fusion analysis model with good predictive performance can be obtained. Feeding physiological signal data into the model yields the final physiological condition and medical diagnosis results.

[0090] In this case study, individual fatigue levels are mapped to a range of 0 to 1. The closer the model output is to 0, the better the individual's mental state; conversely, the closer the model output is to 1, the more fatigued the individual. This transforms the model input / output into a regression prediction problem. The root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are used to predict the case input results, and the resulting errors are shown in Table 1. As can be seen from Table 1, the method proposed in this invention has high accuracy and is practical, applicable to various physiological condition analysis fields such as health monitoring and disease diagnosis.

[0091] Table 1 Prediction results of the method of the present invention Based on the same inventive concept, this invention also proposes a physiological state assessment system, including: a data acquisition module for acquiring multi-source physiological signals of the user to be tested.

[0092] The decomposition module is used to calculate the complexity of each physiological signal; decompose signals with a complexity higher than a set threshold to obtain sub-signals with a complexity lower than the set threshold; and merge the decomposed sub-signals with the undecomposed signals to obtain the expanded feature matrix.

[0093] The filtering module is used to calculate the correlation between signals in the expanded feature matrix using cluster analysis, filter out the signal with the highest average correlation with each signal, and place the signal with the highest correlation with the remaining signal and the signal with the highest average correlation with each signal next to each other to obtain the reordered feature matrix.

[0094] The computation module is used to extract high-order information features from the reordered feature matrix using a convolutional neural network and concatenate them into a high-order information feature group; the high-order information features that contribute the most to the physiological state assessment are calculated through an attention mechanism as the key high-order features.

[0095] The assessment module is used to evaluate the physiological state of the user under test based on key high-order features.

[0096] The present invention also proposes a computer device for assessing physiological state, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the physiological state assessment method.

[0097] The present invention also proposes a readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform steps of a physiological state assessment method.

[0098] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for assessing physiological state, characterized in that, Includes the following steps: Collect multi-source physiological signals from the user under test; Calculate the complexity of each signal in a multi-source physiological signal; Signals with complexity exceeding a set threshold are decomposed to obtain sub-signals with complexity below the set threshold. The decomposed sub-signals are then merged with the undecomposed signals to obtain an expanded feature matrix. Cluster analysis is used to calculate the correlation between signals in the expanded feature matrix, and the signal with the highest average correlation with each signal is selected. The signal with the highest correlation with the selected signal and the signal with the highest average correlation with each signal are then placed next to each other to obtain a reordered feature matrix. A convolutional neural network is used to extract high-order information features from the reordered feature matrix and concatenate them into a high-order information feature group. The high-order information features that contribute the most to the assessment of physiological state are calculated using an attention mechanism and selected as key high-order features; the physiological state of the user to be tested is then assessed based on these key high-order features.

2. The physiological state assessment method according to claim 1, characterized in that, The complexity of each signal in the multi-source physiological signal was calculated using analysis of variance, and is specifically expressed as follows: ,definition Indicates physiological signals The complexity of each physiological signal is determined; the signals are sorted from highest to lowest complexity to obtain the sorted feature matrix. ,in Representing the Secondary sorting.

3. The physiological state assessment method according to claim 1, characterized in that, Signal decomposition is used to decompose signals with complexity exceeding a set threshold, resulting in sub-signals with complexity below the set threshold. The decomposed sub-signals are then merged with the undecomposed signals to obtain an expanded feature matrix. This process includes the following steps: determining the complexity threshold of each physiological signal using an exponential decreasing strategy based on the complexity of each signal. Physiological signals with complexity higher than a complexity threshold are selected; empirical mode decomposition (EMD) is used to decompose these signals into physiological signals with complexity lower than the threshold, represented as follows: ,in This indicates the number of signals after decomposition; for each signal to be decomposed... Find the maximum and minimum points in the signal, and connect the extreme points with spline curves to obtain... and If the complexity of the two signals is less than a set threshold, then the decomposition is stopped. Otherwise, the signals that do not meet the threshold are decomposed again to obtain the decomposed sub-signal groups. The obtained sub-signals are merged to obtain the decomposed signal set. The decomposed signal set is merged with the undecomposed signals to obtain the expanded feature matrix.

4. The physiological state assessment method according to claim 1, characterized in that, The process involves using cluster analysis to calculate the correlation between signals, selecting the signal with the highest average correlation to all other signals, and placing the remaining signals with the highest correlation to the selected signal and the signal with the highest average correlation to all other signals adjacent to each other to obtain a reordered feature matrix. This includes the following steps: calculating the correlation matrix between signals in the expanded feature matrix using cluster analysis; selecting the signal with the highest average correlation to other signals and placing it first; selecting the signal with the highest correlation to the selected signal from the remaining signals and placing it adjacent to the selected signal with the highest average correlation to other signals, and so on, to obtain the reordered feature matrix. 。 5. The physiological state assessment method according to claim 4, characterized in that, The process of extracting high-order information features from the reordered feature matrix using a convolutional neural network and concatenating them into a high-order information feature group specifically includes the following steps: selecting convolutional blocks with different kernel sizes. and Perform convolution operations to extract the feature matrix. Higher-order related information is obtained by concatenating the outputs of each convolutional block along the channel dimension. 。 6. The physiological state assessment method according to claim 5, characterized in that, The process of calculating the high-order information features that contribute the most to the assessment of physiological state through an attention mechanism as key high-order features specifically includes: A channel attention mechanism is used to enhance the proportion of information that contributes most to physiological state assessment in the high-order information feature set, and a three-dimensional matrix is ​​output. ;Will A spatial attention mechanism is used to enhance the proportion of spatially correlated information in high-order information feature groups, and a three-dimensional matrix is ​​output. , as a key high-order feature.

7. The physiological state assessment method according to claim 1, characterized in that, After collecting multi-source physiological signals from the user to be tested, the multi-source physiological signals are preprocessed using an outlier detection method, specifically including the following steps: using the Z-score method to detect and remove outliers and missing values ​​in each physiological signal, and using KNN or nearest neighbor interpolation to complete the data; using the mode frequency as the unified sampling frequency, interpolating and resampling non-uniform frequency signals; and normalizing the resampled signals to obtain the preprocessed signals.

8. A physiological state assessment system, characterized in that, include: The acquisition module is used to acquire multi-source physiological signals from the user under test; The decomposition module is used to calculate the complexity of each signal in a multi-source physiological signal. The signal with a complexity higher than a set threshold is decomposed to obtain sub-signals with a complexity lower than the set threshold; the sub-signals obtained after decomposition are merged with the undecomposed signal to obtain the expanded feature matrix. filter The module is used to calculate the correlation between signals in the expanded feature matrix using cluster analysis, filter out the signal with the highest average correlation with each signal, and place the signal with the highest correlation with the remaining signal and the signal with the highest average correlation with each signal next to each other to obtain the reordered feature matrix. The computation module is used to extract high-order information features from the reordered feature matrix using a convolutional neural network and concatenate them into a high-order information feature group; the high-order information features that contribute the most to the physiological state assessment are calculated through an attention mechanism as the key high-order features. The assessment module is used to evaluate the physiological state of the user under test based on key high-order features.

9. A computer device for assessing physiological state, characterized in that, include: A memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the physiological state assessment method according to any one of claims 1-7.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program, which includes program instructions that, when executed by a processor, perform the steps of the physiological state assessment method according to any one of claims 1-7.

Images