Physiological stress state monitoring method in limited diet or specific diet mode

Abstract

The invention relates to the technical field of artificial intelligence and physiological status monitoring, in particular to a physiological stress status monitoring method in a limited diet or specific diet mode, and the method specifically comprises the following steps: collecting physiological index data of a subject in a diet intervention period, marking the physiological index data, and dividing the collected data; aligning the multi-source asynchronous physiological index time sequence data by using a dynamic time warping algorithm based on mutual information, and outputting a regular physiological index matrix; constructing a physiological stress state monitoring model, inputting the normalized physiological index matrix into the model for classification probability calculation, then calculating the total loss of the model, and training the model; and inputting the physiological index data of a new subject into the trained model, outputting the probability corresponding to each physiological stress state category, and selecting the category with the highest probability value as a final result. The physiological stress state monitoring method provided by the invention is more accurate and comprehensive and meets clinical requirements in a limited diet or specific diet mode.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and physiological state monitoring, and in particular to a physiological stress state monitoring method under a restricted diet or specific dietary pattern. BACKGROUND

[0002] With the development of modern medicine, the monitoring means of physiological indicators is increasingly diversified, including electrocardiogram, electroencephalogram, blood oxygen saturation, body temperature and the collection of various heterogeneous physiological data. These data show high asynchrony and diversity. Traditional medical diagnosis methods mostly rely on single, relatively fixed physiological indicators, which limits their processing capacity for complex pathological conditions. However, in clinical applications, the physiological state of patients often presents diverse and complex changes, and the traditional single data source analysis method cannot fully reflect the actual condition of the patient. The fusion analysis of multi-source asynchronous physiological data provides more information sources, but still faces many challenges in data alignment, feature extraction, model training, etc. In particular, how to maintain the intrinsic physiological relationship in asynchronous data streams and make effective prediction and decision-making combined with clinical pathological mechanisms has become a difficulty in the current medical field.

[0003] The prior art has the following disadvantages in practical application: the existing dynamic time warping algorithm does not consider the biological correlation between indicators, resulting in the neglect of the nonlinear dependence between physiological indicators when processing multi-source asynchronous data, which may cause information loss or mismatch in the data alignment process; the traditional feature extraction method lacks biological and clinical prior knowledge, which may easily lead to the inability of the model to effectively capture clinical key information, poor interpretability, and inability to provide valuable diagnostic basis for clinicians; the conventional convolutional neural network is difficult to effectively process multi-scale physiological stress signals, and cannot fully extract fast and slow varying features in the same network architecture, resulting in inaccurate or insufficient signal processing; although the traditional attention mechanism can improve the performance of the model, it usually lacks clinical guidance, resulting in unreasonable allocation of model attention, which may focus on features irrelevant to the disease mechanism, reducing the clinical value and interpretability of the model.

[0004] Therefore, the present application proposes a physiological stress state monitoring method under a restricted diet or specific dietary pattern to solve the above problems. SUMMARY

[0005] The present application is aimed at the deficiencies of the prior art and develops a physiological stress state monitoring method under a restricted diet or specific dietary pattern. The present application provides a more accurate, comprehensive and clinically relevant physiological stress state monitoring method under a restricted diet or specific dietary pattern.

[0006] The technical solution of the present application is a physiological stress state monitoring method under a restricted diet or specific dietary pattern, comprising the following steps: S1, collect physiological indicator data of a subject during a dietary intervention, form a time series, and manually label physiological stress state categories, divide the collected data into a training set, a validation set, and a test set; S2, align the multi-source asynchronous physiological indicator time series data using a dynamic time warping algorithm based on mutual information, and output the aligned physiological indicator matrix; S3, construct a physiological stress state monitoring model, which includes a physiological constraint feature encoding module, a multi-scale feature decoupling module, an interpretable feature distillation module, and a dynamic sparse classifier. After inputting the aligned physiological indicator matrix into the model, the physiological constraint feature encoding module is used to strengthen the consistency of pathological mechanisms, the multi-scale feature decoupling module is used to separate fast-changing features and slow-changing features, the interpretable feature distillation module is used to focus on key feature patterns of clinical decision rules, and the dynamic sparse classifier is used to calculate the classification probability; Then calculate the total loss of the model, and iteratively train and update the parameters of the model using the data in the training set and the validation set, obtain the trained model, and evaluate the trained model using the test set; S4, input the physiological indicator data of a new subject into the trained model after alignment, output the probability corresponding to each physiological stress state category, and select the category with the highest probability value as the final result.

[0007] S1 is as follows: Physiological indicator data of a subject during a dietary intervention is collected through clinical experiments and wearable devices; The physiological indicator data is used as sample data, and its collection process covers consecutive time points, forming physiological indicator time series data corresponding to the entire intervention period; The collected data is labeled by multiple clinical experts in combination with clinical standards, and the final label is determined through consistency testing. The label is three categories of physiological stress states, namely normal state, mild stress, and severe stress.

[0008] S2 is as follows: Based on the input physiological indicator time series data, the mutual information function is used to quantify the nonlinear dependence between the original physiological indicator data and the physiological indicator data target value, an optimal alignment path solving function is constructed to minimize the Euclidean distance between the two data, the mutual information weight matrix is used to strengthen the correlation priority of metabolic related indicators, and the multi-source asynchronous physiological indicator time series data is aligned in the premise of preserving biological correlation, and the aligned physiological indicator matrix is output; The physiological indicator data target value represents the value of the original physiological indicator data after mapping through the optimal alignment path.

[0009] The operations in the physiological stress state monitoring model are as follows: The normalized physiological index matrix is input into the constructed physiological stress state monitoring model, and is first subjected to a physiological constraint feature coding module. The physiological constraint feature coding module adopts a physiological constraint coding layer to convert the clinical guidelines into differentiable biological constraints, forces the network to learn the feature representation conforming to the pathological mechanism of hyperuricemia, and outputs the feature matrix subjected to physiological constraint coding; The feature matrix subjected to physiological constraint coding is then input into a multi-scale feature decoupling module. The multi-scale feature decoupling module adopts a double-path decoupling structure, extracts fast-changing and slow-changing features through differentiated convolution kernel parameters, and introduces a channel attention gate mechanism for adaptive feature fusion, and outputs the feature matrix subjected to multi-scale fusion. The feature matrix subjected to physiological constraint coding and the feature matrix subjected to multi-scale fusion are then spliced and subjected to interpretable feature distillation. A decision tree-based distillation attention mechanism is introduced, the feature split path is extracted from the CART decision tree as an attention prior, the network is forced to focus on clinically interpretable key feature patterns, and an interpretable attention weight matrix is output. A dynamic sparse classifier is used to input the interpretable attention weight matrix and the feature matrix subjected to multi-scale fusion for classification prediction, and the classifier weight sparsity is controlled through a Bernoulli mask that monotonically increases over time.

[0010] The operations in the physiological constraint feature coding module are as follows: The normalized physiological index matrix is subjected to feature mapping through a linear transformation weight matrix and a bias vector, the mapping result is converted to the (0, 1) interval through a Sigmoid activation function, and the input data is subjected to biological correlation enhancement through a mutual information weight matrix, and a feature matrix subjected to physiological constraint coding is output. The mutual information weight matrix is obtained by calculating the mutual information between all pairs of indicators for each sample, forming a symmetric matrix, and expanding it to the full time sequence according to the time point. The feature matrix subjected to physiological constraint coding is then optimized by defining constraint conditions. Specifically, two groups of Frobenius norm constraint conditions are defined according to the clinical guidelines corresponding to the disease, the relevant indicator pairs contained in the coding layer weight matrix are indexed, and lower and upper norm constraints are applied respectively, the coding layer is forced to strengthen key pathological correlations and weaken redundant correlations, and a constraint loss term is constructed based on the correlation strength of the indicator pairs to optimize the feature matrix subjected to physiological constraint coding.

[0011] The operations in the multi-scale feature decoupling module are as follows: The physiological constraint coded feature matrix is based on one-dimensional convolution operation of a small convolution kernel to capture local rapid change mode, combined with the focusing effect of the gating weight matrix on the transient sensitive area, and outputs a fast-changing feature matrix representing high-frequency transient physiological phenomena. The physiological constraint coded feature matrix and its average pooling result are based on one-dimensional convolution operation of a large convolution kernel to extract global slow change mode, and outputs a slow-changing feature matrix representing low-frequency trend physiological phenomena. The fast-changing feature matrix and the slow-changing feature matrix are spliced in the channel dimension, the channel attention score is calculated through the gating weight matrix, the key biological indicators are strengthened by the down-sampling mutual information weight matrix, and the channel weight distribution is generated by using the Sigmoid function to generate a multi-scale fused feature matrix.

[0012] The interpretable feature distillation operation is as follows: Based on the splicing result of the multi-scale fused feature matrix and the physiological constraint coded feature matrix, the feature split path and the clinical split threshold are extracted through the decision tree classifier training, the attention prior matrix coding the clinical decision rule is constructed, and the query matrix with interpretability is output; The dot product similarity is calculated based on the transpose matrix of the query matrix and the feature dimension scaling factor, the fusion strength of the attention prior matrix and the mutual information matrix is balanced by adjusting the coefficient, and the attention weight matrix consistent with the clinical decision logic is generated by using the Softmax function; The indicator function of the clinical key feature set is introduced in the decision tree split criterion, the split node is guided to preferentially select the feature threshold with diagnostic value by applying an additional reward weight to the clinical key feature, and the focus of the attention mechanism on the clinical pathological feature mode is realized.

[0013] The operation in the dynamic sparse classifier is as follows: The mask matrix with monotonically increasing activation probability with training rounds is generated by Bernoulli distribution sampling, and the dynamic sparse weight matrix is obtained by element-wise multiplication of the mask matrix and the basic weight matrix, realizing the dynamic sparsification of the classifier weight connection strength adaptive adjustment with the training process; Based on the weighted result of the attention weight matrix on the multi-scale feature, the linear transformation is performed through the dynamic sparse weight matrix, combined with the nonlinear filtering effect of the ReLU activation function, and the classification probability vector of the physiological stress state is output, the dimension of the classification probability vector corresponds to the number of physiological stress state categories.

[0014] The loss function calculation is as follows: The soft label supervision of the integrated decision tree distillation is based on the KL divergence constraint to constrain the consistency of the probability distribution of the deep model output and the decision tree pathological rule, and the mutual information alignment factor is introduced to strengthen the loss weight of the key physiological indicators, and the classification consistency loss is obtained. The constraint loss term and the orthogonal regular term of the encoding layer weight matrix are introduced into the physiological constraint feature encoding module to form a total loss function.

[0015] The training process of the model is as follows: The model training adopts a small batch gradient descent algorithm, the gradient of the loss function to the model parameters is calculated through back propagation, and the adaptive matrix estimation optimizer is used to update the trainable parameters such as the weight matrix and the bias vector; In each training round, a batch of samples are randomly extracted from the training data set, and the classification prediction output is obtained through the model; Then the total loss function of the model is calculated, and the model parameters are optimized by minimizing the total loss; After the parameter update, the performance indicators are evaluated on the validation set, and the stopping condition is based on the change monitoring of the validation set loss, when the validation set loss no longer significantly decreases or starts to rise in continuous multiple training rounds, the early stopping mechanism is triggered, and the maximum training round is set as a hard stop condition; After the training is completed, the model parameters with the optimal performance on the validation set are saved to obtain the trained model.

[0016] The effects provided in the summary of the invention are only the effects of the embodiments, not all the effects of the invention, and the above technical solutions have the following advantages or beneficial effects: The dynamic time warping algorithm with mutual information weighting can retain biological relevance and achieve a balance between "time alignment" and "semantic alignment" when processing multi-source asynchronous physiological indicator data, and can overcome the defects of ignoring biological dependence of conventional methods; The physiological constraint encoding layer is introduced in the feature encoding stage, the model is supervised by combining clinical pathological mechanism, the clinical interpretability of feature representation is guaranteed, and the learning of incorrect patterns that do not conform to biological laws is effectively avoided; The multi-scale feature decoupling module is introduced, different feature extraction channels of fast and slow signals are distinguished, and adaptive feature fusion is performed combined with the channel attention mechanism, the understanding and processing ability of multi-scale physiological phenomena is improved; The feature distillation is performed through the attention mechanism based on the decision tree, the clinical decision rules are embedded into the deep learning model, the interpretability of the model is enhanced, the model is more consistent with the clinical practice, and has stronger clinical applicability.

[0017] In summary, the present application realizes accurate processing and biological relevance retention of multi-source asynchronous physiological data through an optimization algorithm, guarantees the clinical interpretability of features and avoids incorrect patterns relying on physiological constraints, improves the processing ability of different rate physiological phenomena through multi-scale decoupling and attention fusion, and further enhances the interpretability and practical applicability of the model by embedding clinical decision rules, which provides more accurate, comprehensive and clinically demand-oriented technical support for physiological stress state monitoring. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0019] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0020] Figure 2 A comparison chart of error distributions for multi-source asynchronous time series alignment methods.

[0021] Figure 3 A comparative chart showing the interpretable distribution of attention weights. Detailed Implementation

[0022] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0023] Example 1 like Figure 1 As shown, a method for monitoring physiological stress under restricted or specific dietary patterns includes the following steps: S1. Collect physiological index data of subjects during dietary intervention, form a time series, and manually label the physiological stress state categories. Divide the collected data into training set, validation set and test set. S2. Apply the dynamic time warping algorithm based on mutual information to align the time series data of multi-source asynchronous physiological indicators and output the warped physiological indicator matrix. S3. Construct a physiological stress state monitoring model. The model includes a physiological constraint feature encoding module, a multi-scale feature decoupling module, an interpretable feature distillation module, and a dynamic sparse classifier. After the normalized physiological index matrix is ​​input into the model, the physiological constraint feature encoding module strengthens the consistency features of pathological mechanisms, the multi-scale feature decoupling module separates fast-changing features and slow-changing features, the interpretable feature distillation module focuses on the key feature patterns of clinical decision-making rules, and the dynamic sparse classifier calculates the classification probability. Then, the total loss of the model is calculated, and the model is iteratively trained and its parameters are updated using data from the training set and the validation set to obtain a trained model. Finally, the trained model is evaluated using the test set. S4. After the new subjects' physiological index data are normalized and aligned, they are input into the trained model, and the probability corresponding to each physiological stress state category is output. The category with the highest probability value is selected as the final result.

[0024] In a specific implementation, S1 is as follows: The present application collects physiological index data of subjects during dietary intervention through clinical experiments and wearable devices, including 10 key indicators of blood uric acid, systolic pressure, diastolic pressure, fasting blood glucose, triglyceride, body mass index, purine intake, total cholesterol, high-density lipoprotein and low-density lipoprotein; The data collection process follows a standardized protocol, blood uric acid and fasting blood glucose are measured daily by venous blood samples, systolic and diastolic blood pressure are recorded multiple times daily using a dynamic blood pressure monitor, body mass index is calculated by weight and height and tracked daily, purine intake is quantified by dietary diary and nutrition analysis software, and total cholesterol, high-density lipoprotein and low-density lipoprotein are obtained by regular blood tests.

[0025] The data collection of each sample covers consecutive time points, forming an original time series, with a total of 1024 time points corresponding to intensive monitoring throughout the intervention period.

[0026] Further, the collected data is labeled, and the data labeling is based on clinical expert evaluation, combined with clinical standards such as blood uric acid level, blood pressure fluctuation, and abnormal glucose metabolism, to divide the physiological stress state into three categories: normal state, mild stress and severe stress. The labeling process is completed independently by multiple medical experts and verified for consistency to ensure the accuracy of the labeling, and finally each sample is attached with a 3-dimensional one-hot encoding label vector for supervised model training.

[0027] The data set is divided into training set, validation set and test set to evaluate the generalization ability of the model.

[0028] In the specific implementation, S2 is specifically as follows: The physiological index data has multi-source asynchronous characteristics, such as blood uric acid and systolic pressure, which respond at different speeds to dietary intervention, resulting in time series that cannot be aligned, and conventional data standardization methods will destroy the correlation between time series, while conventional dynamic time warping algorithms can align the time axis, but ignore the inherent biological correlation between indicators, and cannot preserve their nonlinear dependence.

[0029] The present application is based on the original physiological index time series data, and the nonlinear biological dependence relationship between the indicators is quantified by the mutual information function, an optimal alignment path solving function is constructed with the goal of minimizing the weighted Euclidean distance, the correlation priority of metabolic related indicators is strengthened by using the mutual information weight matrix, the time series alignment of multi-source asynchronous time series is realized under the premise of preserving the biological correlation, and the normalized physiological index matrix is output, which is represented as: In the formula, represents the output matrix of the i-th sample after dynamic time warping, with a dimension of ​, for storing the aligned time series data; represents the optimal alignment path, which can be represented as a mapping function for mapping the original time index to the target time index; represents the optimal alignment path is solved to minimize the objective function; represents the mutual information function, which is used to quantify the nonlinear dependence between two random variables; represents the original value of the th sample at the th time point of the th index, which is an element of the input time series data; represents the value of the th sample at the aligned th time point of the th index, which is an element of the time series data aligned by dynamic time warping; represents the time point index, and the total number of time points of the original sample is 1024; represents the index of the index, and the total number of indexes of the original sample is 10, including blood uric acid, systolic pressure, diastolic pressure, fasting blood glucose, triglyceride, body mass index, purine intake, total cholesterol, high-density lipoprotein, and low-density lipoprotein; represents the sample index; represents the time index mapped by the optimal alignment path , which is used to indicate the time position after alignment; represents the L2 norm, which is equivalent to the calculation method of the Euclidean distance, and is used to calculate the difference between two vectors.

[0030] In specific implementation, the alignment path is the path mapping in dynamic programming, and the optimal curved path is determined by iterative calculation of the minimum cumulative distance to ensure the alignment of physiological indexes after the time axis is twisted.

[0031] In specific implementation, the value of the th sample at the aligned th time point of the th index is derived from the optimal alignment path calculation in the dynamic time warping process, and is obtained by solving the objective function. Specifically, under the optimal alignment path , the value of the It is the first The first sample after alignment The first time point The value of the indicator The direct mapping value ensures the biological consistency of the aligned time series.

[0032] It should be noted that there are non-linear biological dependencies between physiological indicators, such as the metabolic relationship between serum uric acid and fasting blood glucose. Directly using Euclidean distance will weaken this relationship. Weighted summation of mutual information terms representing terms. The Euclidean distance term, represented by the term, achieves dual optimization, namely, through... The term minimizes the time offset of the distorted path. Item-prioritized alignment of indicators with strong nonlinear dependencies, such as blood uric acid and purine intake, inhibits irrelevant temporal alignment, uses mutual information as a weighting factor, and enables the alignment process to simultaneously preserve the temporal local morphology and cross-indicator biological coupling. It embeds biological priors into the alignment process, resolving the contradiction between "temporal alignment" and "semantic alignment" in multi-source asynchronous data.

[0033] S3 is as follows: In a specific implementation, the physiological constraint feature encoding module operates as follows: Conventional feature extraction methods ignore prior medical knowledge, resulting in poor model interpretability and an inability to effectively utilize biological constraints in clinical guidelines. Conventional coding layers focus only on data fitting and may learn feature associations that contradict pathological mechanisms, such as overemphasizing redundant correlations between blood glucose and triglycerides while weakening the key association between blood uric acid and purine intake, thus affecting the clinical applicability of the model.

[0034] This invention employs a physiological constraint coding layer to transform clinical guidelines into differentiable biological constraints, forcing the network to learn feature representations that conform to the pathological mechanism of hyperuricemia. The specific steps are as follows: 1) Constructing the basic coding layer Based on the alignment matrix output by dynamic time warping, feature mapping is performed by linearly transforming the weight matrix and bias vector. The mapping result is then transformed to the (0,1) interval using the Sigmoid activation function. Finally, the mutual information weight matrix is ​​used to enhance the biological associations of the input data, resulting in a feature matrix encoded with physiological constraints, as shown below: In the formula, Indicates the first The feature matrix of each sample after physiological constraint encoding has a dimension of . , used to store encoded physiological characteristics; denotes the Sigmoid activation function, mapping the linear transformation result to interval; denotes the mutual information weight matrix, with dimension , to strengthen the biological relevance of the data after alignment, by calculating the mutual information between all pairs of indicators for each sample, forming a symmetric matrix, and then expanding to the full time series at each time point; denotes the weight matrix of the encoding layer, with dimension is a trainable parameter, used to learn the linear mapping relationship from input to feature; denotes the bias vector of the encoding layer, with dimension is a trainable parameter, used to adjust the threshold of the activation function.

[0035] 2) Define physiological constraints According to the clinical guidelines for hyperuricemia, two groups of Frobenius norm constraints are defined, by performing column index operations on the encoding layer weight matrix with blood uric acid-purine intake indicators and fasting blood glucose-triglyceride indicators, respectively, to impose lower and upper norm constraints, to force the encoding layer to strengthen key pathological correlations and weaken redundant correlations, denoted as: In the formula, denotes the constraint condition, used to regularize the weight matrix of the encoding layer , limiting the learning direction, to ensure that the feature representation conforms to the pathological mechanism of hyperuricemia; denotes the Frobenius norm, which calculates the square sum of all elements of the matrix and takes the square root, used to measure the overall energy of the matrix; denotes the first column index operation, selecting the columns corresponding to blood uric acid and purine intake indicators from the weight matrix of the encoding layer to form a submatrix; denotes blood uric acid, which is one of the physiological indicators; denotes purine intake, which is one of the physiological indicators; denotes the second column index operation, selecting the columns corresponding to fasting blood glucose and triglyceride indicators from the weight matrix of the encoding layer to form a submatrix; denotes fasting blood glucose, which is one of the physiological indicators; denotes triglyceride, which is one of the physiological indicators; represents the lower bound constraint strength, preferably taking the value of , to ensure the sufficient size of the blood uric acid and purine intake associated weight; represents the upper bound constraint strength, preferably taking the value of , to limit the excessive growth of the blood glucose and triglyceride associated weight.

[0036] 3) Realize constraint optimization In the model training process, the constraint loss term is constructed by calculating the maximum value of the blood uric acid-purine intake associated strength deficiency and the blood glucose-triglyceride associated strength excess, and the constraint loss is added to the main loss function for synchronous optimization, so that the coding layer weight matrix is modified to conform to the clinical pathological mechanism, which is represented as: In the formula, represents the constraint loss term, which is used to measure the deviation of the current weight matrix from the physiological constraint; represents the maximum value operation, which uses the realization method of hingeloss function, only produces punishment when the constraint is violated, and the punishment is zero when the constraint is met.

[0037] It should be noted that, the item represents the deficiency of the blood uric acid-purine associated strength, and when the norm is less than , a positive punishment is generated, and forces the coding layer to strengthen the association between blood uric acid and purine intake, and reflects the key pathology of hyperuricemia, the item represents the excess of the blood glucose-triglyceride associated strength, and when the norm is greater than , a positive punishment is generated, and weakens the redundant correlation between blood glucose and triglyceride, avoids learning false patterns, and ensures that the feature coding conforms to the clinical guidelines.

[0038] In the specific implementation, the operation of the multi-scale feature decoupling module is as follows: The physiological stress signal has the coupling of fast-changing characteristics such as systolic pressure fluctuation and slow-changing characteristics such as body mass index change, and the conventional convolutional neural network uses a single scale convolution kernel, which is difficult to effectively separate such multi-scale features, resulting in that the fast-changing transient signal is smoothed or the slow-changing trend feature is not fully extracted.

[0039] The present application adopts a double-path decoupling structure, extracts fast-changing and slow-changing features through differentiated convolution kernel parameters, and introduces a channel attention gate mechanism for adaptive feature fusion, effectively decouples the time-domain multi-scale features, and the specific steps are as follows: 1) Fast-changing feature extraction path Based on the physiological constraint encoding feature matrix, a one-dimensional convolution operation with a small convolution kernel is used to capture local rapid change patterns. Combined with the focusing effect of the gating weight matrix on transient sensitive regions, a fast-changing feature matrix characterizing high-frequency transient physiological phenomena is output, represented as: In the formula, Indicates the first The rapidly changing feature matrix extracted from each sample has a dimension of . Used to store high-frequency transient physiological characteristics; This represents a one-dimensional convolution operation with a kernel size of 3 and a stride of 2, where the subscript... This indicates that the kernel size is 3, used to capture locally rapidly changing patterns. (Subscript) This indicates that the convolution stride is 2, which is used to reduce the resolution of the time dimension. For gating weights, dimensions Focusing on rapidly changing characteristics, the calculation method is expressed as follows: Output via Sigmoid Interval weighting, with high-weight regions focusing on transient features, such as features related to a sudden rise in blood pressure; These are gating training weights, which are trainable parameters used to learn transiently sensitive regions of physiological indicators. The kernel size; is the convolution stride.

[0040] 2) Slow-varying feature extraction pathway Based on the concatenation operation of the physiological constraint encoded feature matrix and its average pooling result, a global slow-change pattern is extracted through a one-dimensional convolution operation with a large convolution kernel. The output is a slowly varying feature matrix representing low-frequency trend physiological phenomena, represented as follows: In the formula, Indicates the first The slowly varying feature matrix extracted from each sample has dimensions of . Used to store low-frequency trend physiological characteristics; Indicates feature splicing; This represents a one-dimensional average pooling operation; This represents a one-dimensional convolution operation with a kernel size of 31 and a stride of 16, where the subscript... This indicates that the kernel size is 31, used to capture slow-moving global patterns. (Subscript) indicates that the convolution step is 16, which is used to significantly reduce the resolution of the time dimension.

[0041] It should be noted that in the fast-changing feature extraction channel, indicates that small convolution kernels are used to capture short-term fluctuations such as blood pressure, indicates that moderate dimension reduction is used to retain details, and the slow-changing feature extraction channel, indicates that large convolution kernels are used to cover long-period trends such as body mass index, indicates that strong dimension reduction is used to extract macro patterns.

[0042] 3) Channel attention gate fusion The fast-changing feature matrix and the slow-changing feature matrix are spliced in the channel dimension, the channel attention score is calculated through the gate weight matrix, the strengthening effect of the down-sampling mutual information weight matrix on the biological key indicators is combined, the channel weight distribution is generated using the Sigmoid function, and the multi-scale fused feature matrix is generated., realize the adaptive fusion output of cross-scale features, expressed as: In the formula, indicates the feature matrix of the th sample after multi-scale fusion, which avoids relying only on statistical correlation of channel attention, but strengthens the cross-scale contribution of clinical key indicators; indicates the down-sampling mutual information weight matrix, which is the mutual information weight matrix after down-sampling matching the feature matrix dimension; indicates the channel attention gate function, which is used to adaptively adjust the importance of the feature channel, The calculation method of the term is expressed as ; indicates that the fast-changing feature matrix extracted from the th sample and the slow-changing feature matrix extracted from the th sample are spliced in the channel dimension; indicates the gate weight matrix, which is a trainable parameter used to calculate the channel attention score; indicates the Hadamard product.

[0043] It should be noted that, The term combines the down-sampling mutual information weight matrix , so that the gate training weight not only depends on statistical correlation, but also is modulated by biological correlation, for example, the transient numerical fluctuation of blood uric acid is small but the clinical value is high,​​​ The channel weight will be increased, and the feature decoupling will be guided by the clinical explainability. Even if the fast-varying signal is weaker than the slow-varying signal in numerical energy, the channel weight can still be strengthened through high mutual information, so as to avoid important pathological features from being submerged.

[0044] In the specific implementation, the operation of the explainable feature distillation module is as follows: Deep neural networks are usually regarded as black box models, and the decision-making process lacks transparency and explainability, which is difficult to meet the requirements of clinical medicine on traceability of model decision-making. Although the conventional attention mechanism can provide feature importance weight, it lacks the guidance of medical prior knowledge and may focus on feature patterns irrelevant to clinical pathology.

[0045] The application introduces a decision tree distillation-based attention mechanism, which extracts feature split paths from a CART decision tree as attention priors to force the network to focus on clinically explainable key feature patterns. The specific steps are as follows: 1) Constructing a decision tree attention prior Based on the splicing result of the multi-scale fusion feature matrix and the physiological constraint coding feature matrix, the feature split path and the clinical split threshold are extracted through decision tree classifier training to construct an attention prior matrix coding clinical decision rules, and an explainable query matrix is output, which is represented as: In the formula, represents the query matrix of the i-th sample after decision tree attention processing, which is used to store the feature representation based on the decision tree prior; represents the decision tree attention operation, including the decision tree training and feature extraction process, The specific operation of the term is as follows: for the input feature Train the CART decision tree classifier, extract the feature index and split threshold of all split nodes in the decision tree, and then construct the attention prior matrix according to the split path.

[0046] 2) Calculate the explainable attention weight Calculate the dot product similarity based on the transpose matrix of the query matrix and the feature dimension scaling factor, balance the fusion strength of the attention prior matrix and the mutual information matrix through the adjustment coefficient, generate an attention weight matrix consistent with the clinical decision logic using the Softmax function, and is represented as: In the formula, represents the explainable attention weight matrix, representing the importance correlation between features and quantifying the contribution of different features to decision-making; ​To adjust the coefficients, the weights of the prior knowledge and data-driven association of the decision tree are preferably set to [value missing]. ; It is the first The query matrix after processing samples using decision tree attention. The mutual information matrix; This represents the softmax activation function, which transforms the input into a probability distribution. Indicates the first The attention prior matrix for each sample, whose elements encode clinical prior knowledge of the decision tree splitting path, such as feature split points and thresholds, is used to quantify the importance associations between features based on clinical rules. express The transpose of the matrix; Indicates the size of the feature dimension, specifically The feature dimension size is used to scale the dot product to prevent gradient explosion.

[0047] In specific implementation, the first Attention prior matrix for each sample Derived from the attention operation of decision trees, it first processes the input features Train a CART decision tree classifier, extract the feature indices and split thresholds of all split nodes, and then construct a prior matrix to encode the decision rules.

[0048] 3) Clinical threshold-guided feature focusing Introducing an indicator function for the set of key clinical features into the decision tree splitting criterion, and by applying additional reward weights to key clinical features, guides splitting nodes to preferentially select feature thresholds with diagnostic value, thereby achieving the focus of the attention mechanism on clinicopathological feature patterns, expressed as: In the formula, The evaluation criteria for selecting splitting features in a decision tree node are used to determine the optimal feature split. The Gini impurity metric, representing the standard, measures the degree of disorder in a dataset; This represents the clinical prior weight coefficient, which controls the importance of clinical prior in the splitting criterion; the preferred value is 0.1. This indicates an indicator function that has a value of 1 when the condition is met and 0 otherwise. Indicates candidate splitting characteristics, such as physiological indicators like blood uric acid and fasting blood glucose; It represents a set of key clinical features, including physiological indicators known to have diagnostic value, derived from a set of key indicators defined in clinical guidelines, such as serum uric acid and fasting blood glucose. This indicates a membership relationship and determines whether a feature is in the set of clinical key features.

[0049] In practical implementation, during the decision tree splitting process, the set of key clinical features is... Additional rewards are applied to features that guide preferential splitting, and the splitting criteria directly affect... The generation process, thereby constraining attention weights Focus on clinical threshold characteristics.

[0050] It should be noted that, This approach combines decision tree rules with mutual information. For example, if transient blood uric acid features simultaneously satisfy high mutual information and high clinical weight, they receive double reinforcement. The splitting paths extracted by the decision tree are encoded as follows: The matrix allows the attention mechanism of the deep model to naturally align with clinical guidelines, ensuring that the weight allocation conforms to the pathological mechanism.

[0051] In a specific implementation, the operation of the dynamic sparse classifier is as follows: Conventional fully connected layers are prone to overfitting when processing small sample physiological data. The fixed network connection structure cannot adapt to the needs of different stages in the training process. Although the conventional Dropout method can prevent overfitting, its randomness will destroy the biological association between features. Moreover, it maintains a fixed sparsity rate throughout the training process, which cannot balance the contradictory needs of preventing overfitting in the early stage of training and maintaining performance in the later stage of training.

[0052] This invention employs a dynamic sparsity mechanism, controlling the sparsity of classifier weights through a Bernoulli mask that monotonically increases over time. This achieves adaptive adjustment, with high sparsity in the early stages of training to prevent overfitting and low sparsity in the later stages to retain more connections and improve performance. The specific steps are as follows: 1) Generate a dynamic sparse weight matrix A mask matrix with monotonically increasing activation probabilities is generated by sampling using a Bernoulli distribution. A dynamically sparse weight matrix is ​​obtained by element-wise multiplication of this mask matrix with the base weight matrix. This achieves dynamic sparsity, where the classifier weight connection strength adaptively adjusts with the training process. This is represented as: In the formula, Indicates the first The dynamic sparse weight matrix during round training has dimensions of . Adaptive sparsity of weight connections is implemented in the forward computation of the classifier to prevent overfitting during training with small samples. denotes the base weight matrix, with dimension , containing complete connection weight parameters, which are updated by training after random initialization; denotes the Bernoulli mask matrix at the th training round, with probability randomly breaking the weight connection, realizing structured sparsity, with dimension , and the element value is 0 or 1, and the calculation method is represented as ; denotes Bernoulli distribution sampling, with probability generating 1, and with probability generating 0; denotes the mask activation probability at the th training round, with a value range of , controlling the sparsity of the network, and the calculation method is represented as ; denotes the growth rate coefficient, controlling the speed of probability increasing with the training round, and the typical value is 0.01; denotes the natural constant; denotes the training round.

[0053] 2) Constructing the forward calculation of the dynamic sparse classifier Based on the weighted results of multi-scale features based on attention weight matrix, linear transformation is performed through dynamic sparse weight matrix, combined with the nonlinear filtering effect of ReLU activation function, and the classification probability vector of physiological stress state is output, represented as: In the formula, denotes the classification probability vector of the th sample, with dimension , corresponding to the classification probability of three physiological stress states; denotes the rectified linear unit activation function, which is used to introduce nonlinearity and filter negative values.

[0054] In the specific implementation, the calculation process of the loss function is as follows: Conventional loss function designs typically only consider prediction accuracy, ignoring clinical interpretability requirements, and are difficult to balance multi-objective optimization. Conventional regularization methods such as Dropout or weight decay can prevent overfitting, but they cannot incorporate medical prior knowledge and maintain a fixed strength throughout the training process, failing to adapt to training dynamics. This leads to the model being prone to overfitting in the early stages of training and having limited performance in the later stages.

[0055] This invention integrates classification loss and decision consistency loss, introduces dynamically decaying constraint loss and orthogonal regularization terms, and constructs a total loss function to achieve synergistic optimization of prediction accuracy, clinical interpretability, and pathological consistency. The specific steps are as follows: 1) Integration of classification loss and decision consistency loss The soft-label supervision of integrated decision tree distillation is used to constrain the consistency between the output of the deep model and the probability distribution of the pathological rules of the decision tree through KL divergence, and a mutual information alignment factor is introduced to strengthen the loss weight of key physiological indicators, thereby achieving the optimization objective guided by clinical interpretability, as expressed as: In the formula, Represents the classification consistency loss, integrating cross-entropy and KL divergence constraints; This indicates the training batch size and controls the sample size used in loss calculations. This represents the cross-entropy loss function, which calculates the difference between the predicted probability and the true label. Indicates the first The one-hot encoded true label vector of each sample, with dimension [missing information]. ; This represents the decision consistency weighting coefficient, which balances the contributions of cross-entropy and KL divergence, with a preferred value of 0.6. This represents the KL divergence calculation function, which measures the difference between the predicted distribution and the decision tree distribution. Indicates the first The decision tree soft label probability vector for each sample is The CART decision tree leaf node category distribution during the implementation process; Represents the mutual information alignment factor vector, with dimension To enhance the loss weighting of key clinical indicators, when a sample is misclassified, a higher penalty is applied if the key indicator is abnormal. The calculation method is expressed as follows: ; Represents the downsampled mutual information weight matrix The Middle a column vector of indicators.

[0056] In a specific implementation, the first a decision tree soft label probability vector of the sample generated by a decision tree attention operation, specifically by After training the CART decision tree, the probability vector is calculated according to the class distribution of the sample falling into the leaf node.

[0057] 2) Dynamic regularization total loss integration The constraint loss intensity is regulated by the attenuation coefficient dependent on the training round, and the weight matrix orthogonal regularization term is introduced to prevent feature redundancy, solving the overfitting and pathological violation coupling problem in small sample training, expressed as: In the formula, the total loss function is represented, integrating the classification loss, the constraint loss, and the regularization term; The constraint loss attenuation coefficient is represented, which is dynamically decreased with the training round, and the calculation method is represented as ; The maximum training round is represented, and the preferred value is 200; The attenuation rate factor is represented, which controls the the descending speed of the term, and the preferred value is 3.0; The orthogonal regularization strength coefficient is represented, and the preferred value is ; The unit matrix is represented, and the dimension .

[0058] It should be noted that The constraint loss is dynamically attenuated, and in the early stage of training, is small, , the network is forced to learn the pathological constraint, and in the later stage of training, is close to , , the model capacity is released to improve the fitting ability.

[0059] It should also be noted that The term forces the encoding layer weight column vector to be orthogonal, eliminating feature redundancy such as false association between blood glucose and triglycerides, and improving feature representation compactness. Through multi-objective optimization of the total loss, the model prediction performance, clinical interpretability, and pathological consistency are synergistically improved.

[0060] In the specific implementation, the model training process is as follows: The model training adopts a small-batch gradient descent algorithm, calculates the gradient of the loss function on the model parameters through back propagation, and updates the weight matrix and bias vector and other trainable parameters using the adaptive matrix estimation optimizer.

[0061] In each training round, a batch of samples is randomly selected from the training dataset and input into the physiological stress state monitoring model, which sequentially passes through the dynamic time warping based on mutual information, physiological constraint feature encoding, multi-scale feature decoupling, interpretable feature distillation, and dynamic sparse classifier modules, and the prediction output is calculated forwardly; The total loss function includes classification consistency loss, dynamically decaying constraint loss, and orthogonal regularization term, and the model parameters are optimized by minimizing the total loss.

[0062] During training, the mask activation probability of the dynamic sparse classifier monotonically increases with the training round, achieving adaptive adjustment from high sparsity to low sparsity. High sparsity in the initial stage prevents overfitting, and low sparsity in the later stage retains more connections to improve performance.

[0063] After parameter update, the performance indicators are evaluated on the validation set, and the stopping condition is based on the change monitoring of the validation set loss. When the validation set loss no longer significantly decreases or starts to rise for consecutive multiple training rounds, the early stopping mechanism is triggered to avoid overfitting. At the same time, the maximum training round is set as a hard stop condition to ensure that the training process terminates within a reasonable time.

[0064] After training is completed, the model parameters with the best performance on the validation set are saved for subsequent physiological stress state monitoring.

[0065] In the specific implementation, S4 is specifically as follows: After the model training is completed, it is deployed in actual application scenarios to realize physiological stress state monitoring under restricted diet or specific dietary patterns.

[0066] In the monitoring stage, the new subject's physiological indicator time series data is input, and then the dynamic time warping algorithm based on mutual information is applied to align the multi-source asynchronous time series, and the biological relevance of key indicators such as blood uric acid and purine intake is retained. The regularized data is input into the trained physiological stress state monitoring model, which sequentially passes through the physiological constraint feature encoding module to strengthen the consistency of pathological mechanisms, the multi-scale feature decoupling module to separate fast-changing features such as systolic pressure fluctuations and slow-changing features such as body mass index changes, the interpretable feature distillation module to focus on key feature patterns of clinical decision rules, and the dynamic sparse classifier to calculate the classification probability. The final output is a three-dimensional probability vector representing the probability distribution of the subject being in a normal state, mild stress, or severe stress, and the physiological stress state category is determined according to the maximum probability.

[0067] The monitoring results can be combined with interpretable attention weights and decision tree splitting paths to provide a basis for clinical decision-making, such as the strength of the association between high uric acid and purine intake, to guide dietary adjustments or medical interventions, and are applicable to personalized health management and clinical practice.

[0068] Example 2 like Figure 2 As shown, a comparative analysis of the alignment error distribution of multi-source asynchronous time series data was conducted to verify the superiority of the proposed dynamic time warping algorithm based on mutual information in processing multi-source asynchronous physiological index data. The experiment compared five different time series alignment methods, including the method of this invention, the conventional dynamic time warping method, the linear interpolation method, the nearest neighbor interpolation method, and the spline interpolation method. In the experimental configuration, each method processed the same multi-source physiological index time series data, including ten indicators such as serum uric acid, systolic blood pressure, and diastolic blood pressure. These indicators exhibit time-series asynchrony due to varying response speeds to dietary interventions. The experimental results show that the method of this invention demonstrates a significant advantage in terms of alignment error distribution. Figure 2 The error distribution is displayed using a combination of violin plots and box plots. The width of the violin represents the density of the data distribution, while the box plot shows the median and interquartile range. The violin shape of the method in this invention is the most compact and lowest, indicating that its alignment error not only has the smallest numerical value but also the smallest fluctuation range, demonstrating the best stability. Although conventional dynamic time warping methods are superior to various interpolation methods, their error distribution is significantly more dispersed than that of the method in this invention. The error distributions of various interpolation methods are relatively dispersed and have high numerical values, especially the nearest neighbor interpolation method, which performs the worst. Experimental results show that this invention, by introducing a mutual information weighting mechanism, effectively preserves the biological correlation between indicators and is effective in resolving the contradiction between "temporal alignment" and "semantic alignment."

[0069] Example 3 like Figure 3As shown, the interpretable attention weight distribution comparison is performed to verify the effect of the attention mechanism based on decision tree distillation in improving the model interpretability. The experiment compares three kinds of attention mechanisms: the decision tree distillation attention proposed in the present application, the self-attention mechanism and the feature importance of the conventional convolutional neural network. The experimental configuration selects five clinical key features, including blood uric acid mutation, purine intake peak, blood pressure surge, blood sugar abnormality and blood lipid fluctuation, which have important value in clinical diagnosis. The experimental results clearly show the attention weight distribution of different methods on key features through stacked column charts. The method of the present application allocates the highest attention weight to the two most important features of blood uric acid mutation and purine intake peak, which is consistent with the focus of clinical experts in actual diagnosis. Although the self-attention mechanism can also identify some important features, its weight distribution is relatively scattered, and it allocates too much weight to features such as blood pressure surge and blood sugar abnormality. The feature importance distribution of the conventional convolutional neural network is the most unreasonable, and it pays insufficient attention to blood uric acid mutation and purine intake peak. The experimental results prove that the present application embeds clinical diagnosis rules into the deep model by introducing the decision tree splitting path as the attention prior, so that the attention distribution of the model is consistent with clinical practice, and the interpretability and clinical applicability of the model are significantly improved.

[0070] Example 4 The overall performance of the complete system in the physiological stress state monitoring task is comprehensively investigated. The experimental purpose is to comprehensively evaluate the comprehensive performance of the present application in terms of accuracy, interpretability and clinical applicability. The experiment sets up ablation comparison and conventional method comparison, including a simplified version removing the dynamic sparse mechanism, a simplified version removing the interpretable feature distillation, and conventional deep learning, support vector machine, random forest and logistic regression, etc. Conventional machine learning methods. The experimental data comes from real clinical acquisition of ten physiological index time series, covering normal state, mild stress and severe stress three physiological states. As shown in Table 1, the experimental results show that the complete system of the present application achieves the optimal level in all evaluation indicators, especially in maintaining high accuracy while achieving excellent interpretability, solving the black box problem of conventional deep learning models. Ablation experiments further prove the important contribution of the dynamic sparse mechanism and the interpretable feature distillation, and the comparison with conventional methods highlights the significant advantages of the present application in clinical applicability and comprehensive performance, providing a more reliable technical solution for physiological stress monitoring under dietary restriction or specific dietary patterns.

[0071] Table 1 Ablation experiment comparison table The above describes the specific embodiments of the application in combination with the drawings, but is not a limitation on the protection scope of the application. Various modifications or variations made by those skilled in the art on the basis of the technical solutions of the application without creative labor are still within the protection scope of the application.

Claims

1. A method for monitoring physiological stress under restricted or specific dietary patterns, characterized in that, Includes the following steps: S1. Collect physiological index data of subjects during dietary intervention, form a time series, and manually label the physiological stress state categories. Divide the collected data into training set, validation set and test set. S2. Apply the dynamic time warping algorithm based on mutual information to align the time series data of multi-source asynchronous physiological indicators and output the warped physiological indicator matrix. S3. Construct a physiological stress state monitoring model. The model includes a physiological constraint feature encoding module, a multi-scale feature decoupling module, an interpretable feature distillation module, and a dynamic sparse classifier. After the normalized physiological index matrix is ​​input into the model, the physiological constraint feature encoding module strengthens the consistency features of pathological mechanisms, the multi-scale feature decoupling module separates fast-changing features and slow-changing features, the interpretable feature distillation module focuses on the key feature patterns of clinical decision-making rules, and the dynamic sparse classifier calculates the classification probability. Then, the total loss of the model is calculated, and the model is iteratively trained and its parameters are updated using data from the training set and the validation set to obtain a trained model. Finally, the trained model is evaluated using the test set. S4. After the new subjects' physiological index data are normalized and aligned, they are input into the trained model, and the probability corresponding to each physiological stress state category is output. The category with the highest probability value is selected as the final result.

2. The method for monitoring physiological stress under restricted or specific dietary patterns according to claim 1, characterized in that, S2 is as follows: Based on the input physiological index time series data, the mutual information function is used to quantify the nonlinear dependency between the original physiological index data and the target value of the physiological index data. An optimal alignment path solution function is constructed with the goal of minimizing the Euclidean distance between the two data. The mutual information weight matrix is ​​used to strengthen the association priority of metabolic related indicators. The time series alignment of multi-source asynchronous physiological index time series data is performed while preserving biological associations, and the normalized physiological index matrix is ​​output. The target value of the physiological indicator data represents the value of the original physiological indicator data after being mapped through the optimal alignment path.

3. The method for monitoring physiological stress under restricted or specific dietary patterns according to claim 1, characterized in that, physiological The specific operations in the stress state monitoring model are as follows: The normalized physiological index matrix is ​​input into the constructed physiological stress state monitoring model. It first passes through the physiological constraint feature encoding module. The physiological constraint feature encoding module adopts the physiological constraint encoding layer to transform clinical guidelines into differentiable biological constraints, forcing the network to learn feature representations that conform to the pathological mechanism of hyperuricemia, and outputs the feature matrix after physiological constraint encoding. The feature matrix after physiological constraint encoding is then input into the multi-scale feature decoupling module. The multi-scale feature decoupling module adopts a dual-path decoupling structure, extracts fast-changing and slow-changing features through differentiated convolution kernel parameters, and introduces a channel attention gating mechanism for adaptive feature fusion, outputting the feature matrix after multi-scale fusion. Then, the feature matrix after physiological constraint encoding and the feature matrix after multi-scale fusion are concatenated and interpreted feature distillation is performed. An attention mechanism based on decision tree distillation is introduced. By extracting feature splitting paths from CART decision trees as attention priors, the network is forced to focus on key clinically interpretable feature patterns and outputs an interpretable attention weight matrix. The dynamic sparse classifier employs a dynamic sparsity mechanism, inputting an interpretable attention weight matrix and a feature matrix fused by multiple scales for classification prediction. The sparsity of the classifier weights is controlled by a Bernoulli mask that increases monotonically over time.

4. The method for monitoring physiological stress under restricted or specific dietary patterns according to claim 3, characterized in that the operation in the physiological constraint feature encoding module is as follows: The normalized physiological index matrix is ​​mapped to the bias vector through a linear transformation weight matrix. The mapping result is then transformed to the (0,1) interval using the Sigmoid activation function. The mutual information weight matrix is ​​used to enhance the biological association of the input data, and the output feature matrix is ​​encoded with physiological constraints. in, The mutual information weight matrix is ​​obtained by calculating the pairwise mutual information between all indicators of each sample, forming a symmetric matrix, and then expanding it to the full time series according to the time points. Then, constraints are defined to optimize the feature matrix after physiological constraint encoding. Specifically, two sets of Frobenius norm constraints are defined according to the clinical guidelines corresponding to the disease. By indexing the relevant index pairs contained in the weight matrix of the encoding layer, lower bound norm constraints and upper bound norm constraints are applied respectively to force the encoding layer to strengthen the key pathological associations and weaken redundant correlations. Then, a constraint loss term is constructed based on the association strength of the index pairs to optimize the feature matrix after physiological constraint encoding.

5. The method for monitoring physiological stress under restricted or specific dietary patterns according to claim 3, characterized in that, The specific operations in the multi-scale feature decoupling module are as follows: Based on the feature matrix encoded by physiological constraints, a one-dimensional convolution operation with a small convolution kernel is used to capture local rapid change patterns. Combined with the focusing effect of the gate weight matrix on transient sensitive regions, a fast-changing feature matrix representing high-frequency transient physiological phenomena is output. Based on the concatenation operation of the feature matrix encoded by physiological constraints and its average pooling result, the global slow change pattern is extracted by one-dimensional convolution operation with large convolution kernel, and the slow-changing feature matrix representing low-frequency trend physiological phenomena is output. The fast-changing feature matrix and the slow-changing feature matrix are concatenated along the channel dimension. The channel attention score is calculated through the gated weight matrix. The downsampled mutual information weight matrix is ​​combined with the enhancement effect of the biological key indicators. The channel weight distribution is generated using the Sigmoid function to generate the multi-scale fused feature matrix.

6. The method for monitoring physiological stress under restricted or specific dietary patterns according to claim 3, characterized in that, The specific characteristics of the distillation operation can be explained as follows: Based on the concatenation result of the feature matrix after multi-scale fusion and the feature matrix after physiological constraint encoding, feature splitting paths and clinical splitting thresholds are extracted by training a decision tree classifier, an attention prior matrix encoding clinical decision rules is constructed, and an interpretable query matrix is ​​output. The dot product similarity is calculated based on the transpose of the query matrix and the feature dimension scaling factor. The fusion strength of the attention prior matrix and the mutual information matrix is ​​balanced by adjusting the coefficients. The attention weight matrix that conforms to the clinical decision-making logic is generated using the Softmax function. Introducing an indicator function of the set of key clinical features into the decision tree splitting criterion, by applying additional reward weights to key clinical features, guides splitting nodes to prioritize feature thresholds with diagnostic value, thereby enabling the attention mechanism to focus on clinical pathological feature patterns.

7. The method for monitoring physiological stress under restricted or specific dietary patterns according to claim 3, characterized in that, The specific operations in the dynamic sparse classifier are as follows: A mask matrix with monotonically increasing activation probability is generated by sampling using Bernoulli distribution. The dynamic sparse weight matrix is ​​obtained by multiplying the mask matrix with the basic weight matrix element by element, thereby realizing the dynamic sparsity of the classifier weight connection strength adaptively adjusted with the training process. Based on the weighted result of multi-scale features by the attention weight matrix, a linear transformation is performed through the dynamic sparse weight matrix, combined with the nonlinear filtering effect of the ReLU activation function, to output the classification probability vector of physiological stress state. The dimension of the classification probability vector corresponds to the number of categories of physiological stress state.

8. The method for monitoring physiological stress under restricted or specific dietary patterns according to claim 1, characterized in that, The loss function is calculated as follows: The soft label supervision of integrated decision tree distillation is used to constrain the consistency between the output of the deep model and the probability distribution of the pathological rules of the decision tree through KL divergence, and the loss weight of key physiological indicators is strengthened by introducing mutual information alignment factor to obtain classification consistency loss. Then, the constraint loss term from the physiological constraint feature encoding module and the orthogonal regularization term from the encoding layer weight matrix are introduced to form the total loss function.

9. The method for monitoring physiological stress under restricted or specific dietary patterns according to claim 1, characterized in that, The training process of the model is as follows: The model training uses the mini-batch gradient descent algorithm, which calculates the gradient of the loss function with respect to the model parameters through backpropagation, and uses the adaptive moment estimator optimizer to update trainable parameters such as the weight matrix and bias vector. In each round of training, batches of samples are randomly drawn from the training dataset and processed by the model to obtain classification prediction output; Then, the total loss function of the model is calculated, and the model parameters are optimized by minimizing the total loss. After the parameters are updated, the performance metrics are evaluated on the validation set. The decision to stop the iteration is based on the monitoring of changes in the validation set loss. When the validation set loss no longer decreases significantly or begins to increase in multiple consecutive training rounds, the early stopping mechanism is triggered. At the same time, the maximum number of training rounds is set as a hard stopping condition. After training is complete, save the parameters of the model with the best performance on the validation set to obtain the trained model.

10. The method for monitoring physiological stress under restricted or specific dietary patterns according to claim 1, characterized in that, S1 is as follows: Physiological data of subjects were collected during the dietary intervention through clinical trials and wearable devices; Physiological index data, as sample data, are collected over a continuous time period to form time series data of physiological indicators, corresponding to the entire intervention cycle. Multiple clinical experts independently labeled the collected data in accordance with clinical standards, and then the final labeling was determined through a consistency test. The labels were divided into three categories of physiological stress state: normal state, mild stress, and severe stress.

Images