Brain oxygen frequency domain feature screening method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of biomedical signal processing, in particular to a brain oxygen frequency domain feature screening method and device, electronic equipment and a storage medium, and the method comprises the steps: collecting first near-infrared signal data and second near-infrared signal data, and obtaining brain oxygen time sequence data; at least one frequency domain feature is extracted from brain oxygen time sequence data, and a Markov decision process based on a finite frequency diffusion feature search strategy is established on the basis of the at least one frequency domain feature in combination with the characteristics that frequency concentration of brain oxygen signal features is similar to blood oxygen homologous feature distribution; and performing approximate optimization training of a state-action strategy by using a deep Q network to obtain an optimal feature subset for disease recognition. According to the method, the feature selection logic of the Markov decision process can be optimized by using the deep Q network and the finite frequency diffusion feature search mode, and meanwhile, the time for repeatedly selecting the highly dependent features and training the model is shortened.

Description

Technical Field

[0001] The present application relates to the field of biomedical signal processing technology, and in particular to a method, device, electronic device and storage medium for screening brain oxygen frequency domain characteristics. Background Art

[0002] In recent years, near-infrared spectroscopy (NIRS) has been widely used in the study of neurological diseases (such as Alzheimer's disease and depression) and brain function due to its noninvasive, real-time monitoring of brain oxygen metabolism. By collecting signals of changes in the concentrations of oxyhemoglobin (HbO) and deoxyhemoglobin (HbR) in brain regions, NIRS can reflect neurovascular coupling mechanisms and the dynamic characteristics of brain functional networks. However, problems such as high dimensionality, low signal-to-noise ratio, and multiple sources of physiological interference make extracting disease-related features from massive amounts of near-infrared information a significant challenge.

[0003] In related technologies, near-infrared signals are generally processed by bandpass filtering, independent processing of multi-channel signals, black box models, etc.; traditional screening methods such as filter methods, wrapper methods, and embedded methods are used to screen features, and reinforcement learning is applied to feature screening.

[0004] However, the above methods used in the related art all have significant defects, as follows:

[0005] 1. Limitations of feature screening technology

[0006] Filter method:

[0007] It quickly screens features based on statistical indicators (such as t-test, mutual information, and correlation coefficient), but ignores the interactions between features and cannot be dynamically coupled with classifier performance.

[0008] Wrapper method:

[0009] Feature subsets are optimized iteratively through recursive feature elimination or genetic algorithms combined with classifier performance, but the computational complexity is extremely high, making it difficult to apply to multi-channel brain oxygen time series data (e.g., 32 channels × 10 Hz sampling × 30 minutes).

[0010] Embedded method:

[0011] Automatically select features during model training (such as Lasso regression and random forest importance scoring), but the model dependency is too strong and there is a lack of cross-model generalization capabilities.

[0012] 2. Special Difficulties in Near-Infrared Signal Processing

[0013] Physiological noise pollution: Low-frequency interference such as breathing (0.2-0.3Hz) and heartbeat (0.8-1.2Hz) overlaps with the frequency band of brain function signals (0.01-0.1Hz). Traditional bandpass filtering methods can easily cause effective signal distortion.

[0014] Cross-brain spatiotemporal correlation: The spatial distribution of multi-channel signals (such as prefrontal-parietal functional connectivity) and temporal synchronization (such as event-related potentials) need to be jointly modeled, while existing methods mostly process spatiotemporal features independently.

[0015] Interpretability requirements: In medical scenarios, features need to be associated with physiological mechanisms (such as "reduced low-frequency power of HbO in the left prefrontal lobe is associated with cognitive decline"), but black box models (such as deep learning) are difficult to provide clinically acceptable explanations.

[0016] 3. Bottlenecks in the application of reinforcement learning in feature screening

[0017] Although reinforcement learning has shown great potential in automatic feature engineering, its application in medical data still faces the following challenges:

[0018] Single state representation: Existing solutions only use mutual information or classification error as state input, which cannot fully characterize the frequency-time-spatial multidimensional characteristics of brain oxygen signal.

[0019] Reward design bias: Directly using classification accuracy as a reward can easily lead to local optimality, ignoring key factors such as feature redundancy and computational cost.

[0020] In summary, the current feature screening methods of near-infrared signals in related technologies often ignore key factors such as the interaction between features, the comprehensive characterization of the multidimensional features of brain oxygen signals, feature redundancy, and computational cost, resulting in significant deficiencies in efficiency, robustness, and interpretability. They also lack the ability to dynamically optimize multidimensional brain oxygen data, and it is difficult to determine disease characteristics. These problems urgently need to be improved. Summary of the Invention

[0021] The present application provides a brain oxygen frequency domain feature screening method, device, electronic device and storage medium to address the problems in the related art that current feature screening methods for near-infrared signals often ignore key factors such as the interaction between features, the comprehensive characterization of the multidimensional features of brain oxygen signals, feature redundancy, and computational cost, resulting in significant deficiencies in efficiency, robustness, and interpretability, lack of dynamic optimization capabilities for multidimensional brain oxygen data, and difficulty in determining disease characteristics.

[0022] The first aspect of the present application provides a method for screening brain oxygen frequency domain features, comprising the following steps: respectively collecting first near-infrared signal data of a target population that meets preset normal conditions and second near-infrared signal data of a target disease population; preprocessing the first near-infrared signal data and the second near-infrared signal data to obtain multi-position and multi-type brain oxygen time series data; extracting at least one frequency domain feature from the brain oxygen time series data, and based on the at least one frequency domain feature, combined with the similarity in frequency concentration of brain oxygen signal features and distribution of blood oxygen homologous features, establishing a Markov decision process based on a finite frequency diffusion feature search strategy, and using a deep Q network to perform approximate optimization training of a state-action strategy to obtain an optimal feature subset for disease identification.

[0023] Through the above technical solution, we can construct a Markov decision process and use a deep Q network to perform an approximate optimization strategy for the state-action strategy. Combined with the characteristics of the frequency concentration of brain oxygen signal features and the similar distribution of blood oxygen homologous features, the feature selection strategy is optimized according to the finite frequency diffusion feature search method to obtain the optimal feature subset for disease identification. In this way, the feature selection logic is optimized, the repeated selection of highly dependent features is reduced, the model training time is reduced, and effective feature guidance is provided for disease analysis.

[0024] Optionally, in one embodiment of the present application, the method further includes: adding the optimal feature subset and the interference feature subset to a pre-trained machine learning classification model to verify the rationality of the optimal feature subset.

[0025] Through the above technical solution, the optimal feature subset can be selected by verifying the rationality of the optimal feature subset. Selecting the optimal feature subset can reduce noise and redundant features, thereby improving the model's performance indicators such as accuracy and recall rate, effectively improving the performance of the model. The optimal feature subset can also reduce overfitting and improve the model's generalization ability and interpretability on unseen data.

[0026] Optionally, in one embodiment of the present application, extracting at least one frequency domain feature from the brain oxygen time series data includes: dividing the brain oxygen time series data into multiple sub-frequency bands; calculating the frequency domain features of the blood oxygen signal of each type and / or each position based on the multiple sub-frequency bands; and normalizing the frequency domain features of the blood oxygen signal of each type and / or each position to generate the at least one frequency domain feature.

[0027] Through the above technical solution, at least one frequency domain feature can be extracted from brain oxygen time series data. Based on the extracted frequency domain feature, the frequency characteristics of brain activity can be reflected, abnormal patterns in brain oxygen signals can be captured, and brain diseases can be diagnosed in an auxiliary manner, providing important information and support for brain function research, disease diagnosis, and health monitoring.

[0028] Optionally, in one embodiment of the present application, the establishment of a Markov decision process based on a finite frequency diffusion feature search strategy includes: determining a first state component by the mutual information between the current feature and the label; determining a second state component by the sum of the mutual information between the current feature and other features; training a pre-built basic classification model based on all features, and calculating the static Shapley additive explanation value of each feature to obtain an absolute value mean, and determining a third state component based on the absolute value mean; determining the current state based on the first state component, the second state component and the third state component to determine an immediate reward, and constructing the Markov decision process.

[0029] Through the above technical solution, by using mutual information and Shapley additive explanations, namely SHAP values (SHapleyAdditive exPlanations, Shapley additive explanations) as state components, information can be provided from two perspectives: the global correlation between features and labels and the local contribution of features to model prediction. Mutual information provides the global correlation between features and labels, and SHAP values provide the local contribution of features to model prediction. The combination of the two can construct a more comprehensive state representation, enhance the richness of state representation and model performance, and improve the accuracy of feature screening.

[0030] Optionally, in one embodiment of the present application, the current state is expressed as:

[0031]

[0032]

[0033] Among them, S t is the current state; is the first state component; is the second state component; is the third state component; f t is the current feature; C is the label; I(C; f t ) is the current feature f t Mutual information with the label; f′ is other features in the feature set F′; I(f t ,f′) is the current feature f t Mutual information with other features in feature set F′; ∑ f′∈F′ I(f t ,f′) is the current feature f t The sum of the mutual information between the other features in the feature set F′; φ mean (f t ) is the Shapley additive explanatory value of the current feature in the benchmark classification model.

[0034] Through the above technical solution, the current state can be represented by utilizing the correlation between the current feature and the label, the redundancy of the current feature and other features in the feature set, and the classification importance of the current feature in the baseline classification model. This can build a more comprehensive state representation and support the modeling of complex tasks and dynamic feature weight adjustment.

[0035] Optionally, in one embodiment of the present application, the finite frequency diffusion feature search strategy includes: constructing an initial feature set based on mutual information sorting between features and labels;

[0036] A seed feature set is selected from the initial feature set; a dynamic pruning search is performed in the homologous frequency band neighborhood of the seed feature of the seed feature set, feature addition decisions are made with an ε decay strategy, the initial feature set is improved, and path pruning is performed based on the change in classifier loss after the candidate features are added. After all seed feature searches are completed, the next round of seed feature selection and dynamic pruning search is performed until the target number of rounds is reached, and the feature subset corresponding to the minimum loss is output.

[0037] Through the above technical solution, the feature selection strategy can be optimized through the finite frequency diffusion feature search method, that is, the transfer strategy of the Markov decision process can be optimized, which can limit the search frequency range and focus on the key frequency band. Key features can be selected efficiently and accurately in the frequency domain, thereby improving the efficiency, quality and performance of feature selection.

[0038] A second aspect of the present application provides a brain oxygen frequency domain feature screening device, including: an acquisition module, used to respectively collect first near-infrared signal data of a target population that meets preset normal conditions and second near-infrared signal data of a target disease population; a processing module, used to preprocess the first near-infrared signal data and the second near-infrared signal data to obtain multi-location and multi-type brain oxygen time series data; a screening module, used to extract at least one frequency domain feature from the brain oxygen time series data, and based on the at least one frequency domain feature, use a pre-constructed Markov decision process and an approximate optimization strategy of a state-action strategy based on a deep Q network, combined with the frequency concentration of brain oxygen signal features and the similarity of the distribution of blood oxygen homologous features, optimize the feature selection strategy according to the finite frequency diffusion feature search method to obtain the optimal feature subset for disease identification.

[0039] Through the above technical solution, we can construct a Markov decision process and use a deep Q network to perform an approximate optimization strategy for the state-action strategy. Combined with the characteristics of the frequency concentration of brain oxygen signal features and the similar distribution of blood oxygen homologous features, the feature selection strategy is optimized according to the finite frequency diffusion feature search method to obtain the optimal feature subset for disease identification. In this way, the feature selection logic is optimized, the repeated selection of highly dependent features is reduced, the model training time is reduced, and effective feature guidance is provided for disease analysis.

[0040] Optionally, in one embodiment of the present application, it further includes: a verification module, used to add the optimal feature subset and the interference feature subset to the pre-trained machine learning classification model to verify the rationality of the optimal feature subset.

[0041] Through the above technical solution, the optimal feature subset can be selected by verifying the rationality of the optimal feature subset. Selecting the optimal feature subset can reduce noise and redundant features, thereby improving the model's performance indicators such as accuracy and recall rate, effectively improving the performance of the model. The optimal feature subset can also reduce overfitting and improve the model's generalization ability and interpretability on unseen data.

[0042] Optionally, in one embodiment of the present application, the screening module includes: a division unit for dividing the brain oxygen time series data into multiple sub-frequency bands; a calculation unit for calculating the frequency domain characteristics of the blood oxygen signal of each type and / or each position based on the multiple sub-frequency bands; and a generation unit for normalizing the frequency domain characteristics of the blood oxygen signal of each type and / or each position to generate the at least one frequency domain feature.

[0043] Through the above technical solution, at least one frequency domain feature can be extracted from brain oxygen time series data. Based on the extracted frequency domain feature, the frequency characteristics of brain activity can be reflected, abnormal patterns in brain oxygen signals can be captured, and brain diseases can be diagnosed in an auxiliary manner, providing important information and support for brain function research, disease diagnosis, and health monitoring.

[0044] Optionally, in one embodiment of the present application, it also includes: a first state component determination module, which is used to determine the first state component by the mutual information between the current feature and the label before obtaining the optimal feature subset for disease identification; a second state component determination module, which is used to determine the second state component by the sum of the mutual information between the current feature and other features before obtaining the optimal feature subset for disease identification; a third state component determination module, which is used to train a pre-built basic classification model based on all features before obtaining the optimal feature subset for disease identification, and calculate the static Shapley additive explanation value of each feature to obtain the absolute value mean, and determine the third state component based on the absolute value mean; a construction module, which is used to determine the current state based on the first state component, the second state component and the third state component before obtaining the optimal feature subset for disease identification, so as to determine the immediate reward and construct the Markov decision process.

[0045] Through the above technical solution, by using mutual information and Shapley additive interpretation value, that is, SHAP value as state components, information can be provided from two perspectives: the global correlation between features and labels and the local contribution of features to model prediction. Mutual information provides the global correlation between features and labels, and SHAP value provides the local contribution of features to model prediction. The combination of the two can construct a more comprehensive state representation, greatly enhance the richness of state representation and model performance, and improve the accuracy of feature screening.

[0046] Optionally, in one embodiment of the present application, the screening module is also used to: construct an initial feature set based on the mutual information sorting of features and labels; select a seed feature set from the initial feature set; perform a dynamic pruning search within the homologous frequency band neighborhood of the seed features of the seed feature set, make feature addition decisions with an ε attenuation strategy, and at the same time improve the initial feature set, and perform path pruning based on the change in classifier loss after the candidate features are added. After all seed feature searches are completed, the next round of seed feature selection and dynamic pruning search is performed until the target number of rounds is reached, and the feature subset corresponding to the minimum loss is output.

[0047] Through the above technical solution, the feature selection strategy can be optimized through the finite frequency diffusion feature search method, that is, the transfer strategy of the Markov decision process can be optimized, which can limit the search frequency range and focus on the key frequency band. Key features can be selected efficiently and accurately in the frequency domain, thereby improving the efficiency, quality and performance of feature selection.

[0048] A third aspect of the present application provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the brain oxygen frequency domain feature screening method as described in the above embodiment.

[0049] A fourth aspect of the present application provides a computer-readable storage medium, which stores a computer program. When the program is executed by a processor, it implements the above-mentioned brain oxygen frequency domain feature screening method.

[0050] A fifth aspect of the present application provides a computer program product, which stores a computer program that, when executed by a processor, implements the above-mentioned brain oxygen frequency domain feature screening method.

[0051] The embodiment of the present application can extract at least one frequency domain feature from the acquired multi-location and multi-type brain oxygen time series data, and construct a Markov decision process based on the finite frequency diffusion feature search strategy based on the frequency domain feature, and perform an approximate optimization strategy of the state-action strategy through a deep Q network. In combination with the characteristics of the frequency concentration of brain oxygen signal features and the similarity of the distribution of blood oxygen homologous features, the feature selection strategy is optimized according to the finite frequency diffusion feature search method to obtain the optimal feature subset for disease identification, thereby optimizing the feature selection logic, reducing the repeated selection of highly dependent features, reducing the time of model training, and providing effective feature guidance for disease analysis. Thus, the problem that the current feature screening methods of near-infrared signals in the related art often ignore the key factors such as the interaction between features, the comprehensive characterization of the multidimensional features of brain oxygen signals, feature redundancy, and computational cost, resulting in significant deficiencies in efficiency, robustness, and interpretability, lack of dynamic optimization capabilities for multi-dimensional brain oxygen data, and difficulty in determining disease features is solved.

[0052] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0054] Figure 1 This is a flow chart of a method for screening brain oxygen frequency domain features according to an embodiment of the present application;

[0055] Figure 2 1 is a flow chart of a method for screening brain oxygen frequency domain features according to a specific embodiment of the present application;

[0056] Figure 3 1 is a flow chart of a finite frequency diffusion feature search method according to one embodiment of the present application;

[0057] Figure 4 Schematic diagram of a brain oxygen frequency domain feature screening device according to an embodiment of the present application;

[0058] Figure 5 A schematic diagram of the structure of an electronic device provided according to an embodiment of the present application.

[0059] Reference numerals:

[0060] 10- Brain oxygen frequency domain feature screening device: 100- acquisition module, 200- processing module and 300- screening module; 501- memory, 502- processor and 503- communication interface. DETAILED DESCRIPTION

[0061] The following describes in detail embodiments of the present application, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0062] The following describes the brain oxygen frequency domain feature screening method, device, electronic device and storage medium of the embodiment of the present application with reference to the accompanying drawings. In view of the fact that the current feature screening method for near-infrared signals in the related art mentioned in the background technology center has significant deficiencies in efficiency, robustness and interpretability, and lacks the ability to dynamically optimize multi-dimensional brain oxygen data, the present application provides a brain oxygen frequency domain feature screening method, in which at least one frequency domain feature can be extracted from the acquired multi-position and multi-type brain oxygen time series data, and a Markov decision process based on a finite frequency diffusion feature search strategy is constructed based on the frequency domain feature. An approximate optimization strategy of the state-action strategy is performed through a deep Q network. Combined with the characteristics of the frequency concentration of brain oxygen signal features and the similarity of the distribution of blood oxygen homologous features, the feature selection strategy is optimized according to the finite frequency diffusion feature search method to obtain the optimal feature subset for disease identification, thereby optimizing the feature selection logic, reducing the repeated selection of highly dependent features, reducing the model training time and providing effective feature guidance for disease analysis. This solves the problem that current feature screening methods for near-infrared signals in related technologies often ignore key factors such as the interaction between features, the comprehensive characterization of the multidimensional features of brain oxygen signals, feature redundancy, and computational cost, resulting in significant deficiencies in efficiency, robustness, and interpretability, a lack of dynamic optimization capabilities for multidimensional brain oxygen data, and difficulty in determining disease characteristics.

[0063] Specifically, Figure 1 A flowchart of a method for screening brain oxygen frequency domain features provided in an embodiment of the present application.

[0064] like Figure 1 As shown, the brain oxygen frequency domain feature screening method includes the following steps:

[0065] In step S101, first near-infrared signal data of a target population meeting preset normal conditions and second near-infrared signal data of a target disease population are collected respectively;

[0066] It is understandable that the near-infrared signal can be a signal related to the interaction between light and matter obtained through near-infrared spectroscopy technology, including but not limited to HbO (Oxyhemoglobin, oxygenated hemoglobin) concentration, HbR (Hemoglobin, deoxyhemoglobin) concentration, HbT (Total Hemoglobin, total hemoglobin) concentration, tissue oxygen saturation, hemodynamic information and other signals.

[0067] In actual implementation, embodiments of the present application can use a functional near-infrared multi-channel acquisition device to collect near-infrared signal data from a target population and a normal control group. For example, embodiments of the present application can use a dual-lead near-infrared device for acquisition, with the probes placed in the left and right frontal lobe regions of the subject, respectively. The population to be collected can be stroke patients and healthy people.

[0068] The embodiments of the present application can collect near-infrared signal data of healthy people and diseased people for comparative analysis. By comparing the two types of data, the differences in near-infrared signals in healthy and diseased states can be analyzed, which can be better used for disease diagnosis or research.

[0069] In step S102 , the first near-infrared signal data and the second near-infrared signal data are preprocessed to obtain multi-position and multi-type brain oxygen time series data.

[0070] Among them, brain oxygen time series data can be understood as brain hemodynamic data that changes over time and is continuously collected through near-infrared spectroscopy technology (such as fNIRS (functional near-infrared spectroscopy)). It reflects the changes in the concentrations of HbO, HbR, and HbT in the cerebral cortex over time.

[0071] For example, the preprocessing method can be set by those skilled in the art according to actual conditions. For example, the original optical density signal collected from multiple channels can be band-pass filtered at 0.1-2 Hz to retain as much frequency domain information as possible; TDDR (Temporal Derivative Distribution Repair) is used to correct the signal head motion; and the optical density signal is converted into the variation of HbO, HbR, and HbT using the improved Beer-Lambert Law. After preprocessing, multi-location and multi-type brain oxygen time series data can be obtained.

[0072] The embodiment of the present application can extract brain oxygen time series data of multiple spatial positions (channels) by preprocessing the first and second near-infrared signals, thereby reflecting the activities of different brain regions and supporting the construction and analysis of brain functional networks; it can simultaneously extract the time series data of HbO, HbR and HbT to provide more comprehensive hemodynamic information and support the comprehensive analysis of oxygen supply, consumption and blood volume changes.

[0073] In step S103, at least one frequency domain feature is extracted from the brain oxygen time series data. Based on the at least one frequency domain feature, combined with the similarity in the frequency concentration of brain oxygen signal features and the distribution of blood oxygen homologous features, a Markov decision process based on a finite frequency diffusion feature search strategy is established. The deep Q network is used to perform approximate optimization training of the state-action strategy to obtain the optimal feature subset for disease identification.

[0074] It is understandable that the Markov Decision Process (MDP) can be used as a mathematical model to describe sequential decision-making problems. The core elements include state S, action A, reward R, and transition strategy P; Deep Q-Network (DQN) is a reinforcement learning algorithm that combines Q learning and deep neural networks, and can solve high-dimensional state space problems through experience replay and target networks.

[0075] In practice, the Markov decision process can effectively reduce model overfitting, improve classification model training speed, and discover specific feature subsets by screening large-dimensional frequency-domain features. Feature screening can be addressed by constructing a Markov decision process, which consists of four main elements: state S, action A, reward R, and transition strategy P. Approximate optimization of the state-action strategy is achieved using a deep Q-network. A dual-channel deep Q-network architecture can be designed, in which the global channel implements feature space dimensionality reduction using a mutual information weight matrix, while the local channel adopts a feature selection strategy constrained by frequency band homology. An ε-greedy search is performed within the sub-band neighborhood to rapidly discover the optimal disease-related feature subset. This optimizes the feature selection strategy, i.e., the transition strategy of the Markov decision process, ultimately resulting in the optimal feature subset for disease recognition.

[0076] The embodiment of the present application can optimize the state-action strategy of the Markov decision process through a deep Q network, and optimize the transfer strategy of the Markov decision process through a finite frequency diffusion feature search method. It combines the characteristics of the frequency concentration of brain oxygen signal features and the similar distribution of blood oxygen homologous features, optimizes the feature selection logic, and reduces the repeated selection of highly dependent features. It reduces the time of model training and provides effective feature guidance for disease analysis, improves the efficiency, robustness and interpretability of near-infrared signal feature screening, and realizes dynamic optimization of multi-dimensional brain oxygen data.

[0077] Optionally, in one embodiment of the present application, it also includes: adding an optimal feature subset and an interference feature subset to a pre-trained machine learning classification model to verify the rationality of the optimal feature subset.

[0078] It is understandable that the machine learning classification model includes but is not limited to support vector machines, k-nearest nearest neighbor algorithms, decision trees, random forest hidden Markov models, and the like.

[0079] In the actual implementation process, the embodiment of the present application can adopt a tree classifier model. With the advantage of the tree classifier model being able to sort the importance of input features, the optimal feature subset F with m features is obtained. * Afterwards, an interference feature set outside the optimal subset is added to the subset to obtain a mixed feature subset F″. The mixed feature subset F″ is used to train the tree classifier. After the model training is completed, the rationality of the filtered subset is verified by feature importance ranking. If more than 80% of the top m features belong to the optimal subset, the screening is judged to be effective.

[0080] The embodiments of the present application can select the optimal feature subset by verifying the rationality of the optimal feature subset. Selecting the optimal feature subset can reduce noise and redundant features, thereby improving performance indicators such as model accuracy and recall rate, effectively improving the performance of the model. The optimal feature subset can also reduce overfitting and improve the generalization ability and interpretability of the model on unseen data.

[0081] Optionally, in one embodiment of the present application, extracting at least one frequency domain feature from the brain oxygen time series data includes: dividing the brain oxygen time series data into multiple sub-frequency bands; calculating the frequency domain features of the blood oxygen signal of each type and / or each position based on the multiple sub-frequency bands; and normalizing the frequency domain features of the blood oxygen signal of each type and / or each position to generate at least one frequency domain feature.

[0082] Among them, frequency domain features can be understood as features extracted after converting the signal from the time domain to the frequency domain. They reflect the distribution and energy information of the signal in different frequency components and are widely used in signal processing, image processing, speech recognition, biomedical signal analysis and other fields.

[0083] Specifically, extracting at least one frequency domain feature may include: dividing the preprocessed blood oxygen signal into k sub-bands, calculating the frequency domain features of the blood oxygen signal for each type (including HbO, HbR, HbT, etc.) and each location (including the left frontal lobe, the right frontal lobe, etc.), including power spectral density, wavelet coefficients, sample entropy, etc., and normalizing these features to make their ranges consistent.

[0084] The embodiments of the present application can extract at least one frequency domain feature from brain oxygen time series data, reflect the frequency characteristics of brain activity based on the extracted frequency domain feature, capture abnormal patterns in brain oxygen signals, assist in the diagnosis of brain diseases, and provide important information and support for fields such as brain function research, disease diagnosis, and health monitoring.

[0085] Optionally, in one embodiment of the present application, a Markov decision process based on a finite frequency diffusion feature search strategy is established, including: determining a first state component by the mutual information between the current feature and the label; determining a second state component by the sum of the mutual information between the current feature and other features; training a pre-built basic classification model based on all features, and calculating the static Shapley additive explanation value of each feature to obtain the absolute value mean, and determining the third state component based on the absolute value mean; determining the current state based on the first state component, the second state component, and the third state component to determine the immediate reward, and constructing a Markov decision process.

[0086] It is understandable that mutual information can be used to measure the statistical dependence between two random variables, which represents the amount of information about another variable that can be obtained by observing one variable. In machine learning and data analysis, mutual information is commonly used in feature selection to measure the correlation between features and labels. SHAP value (SHapley Additive exPlanations, Shapley additive explanation) can be used to explain the prediction results of the machine learning model. The Shapley additive explanation value of the embodiment of the present application comes from the pre-training model and is static. It provides a fair and consistent feature importance measure by calculating the contribution of each feature to the model output. The SHAP value combines the Shapley value in game theory and the local explanation method in machine learning, and is widely used in model interpretability analysis.

[0087] During the actual implementation process, the embodiment of the present application can use all feature data to train the benchmark classification model after completing the frequency domain feature extraction processing, and calculate the absolute mean of the SHAP value corresponding to each feature in the model; for example, XGBoost is selected as the benchmark classification model, and the SHAP optimization algorithm Tree SHAP is used to calculate the SHAP value; for frequency domain features with huge dimensions, features need to be screened to effectively reduce model overfitting, improve the classification model training speed, and discover specific feature subsets.

[0088] For the problem of feature screening, the embodiment of the present application can construct a Markov decision process and implement approximate optimization of the state-action strategy based on the deep Q network. The Markov decision process contains four main elements, namely state S, action A, reward R, and transfer strategy P.

[0089] In the feature screening problem, the embodiment of the present application can define the state S as a vector, which may specifically include the correlation between the current feature and the label, the redundancy of the current feature and other features in the feature set, and the classification importance of the current feature in the benchmark classification model.

[0090] The embodiments of the present application can use mutual information and SHAP values as state components to provide information from two perspectives: the global correlation between features and labels and the local contribution of features to model prediction. Mutual information provides the global correlation between features and labels, and SHAP values provide the local contribution of features to model prediction. The combination of the two can construct a more comprehensive state representation, greatly enhance the richness of state representation and model performance, and improve the accuracy of feature screening.

[0091] Optionally, in one embodiment of the present application, the current state is expressed as:

[0092]

[0093] Among them, S t is the current state; is the first state component; is the second state component; is the third state component; f t is the current feature; C is the label; I(C; f t ) is the current feature f t Mutual information with the label; f′ is other features in the feature set F′; I(f t ,f′) is the current feature f t Mutual information with other features in feature set F′; ∑ f′∈F′ I(f t ,f′) is the current feature f t The sum of the mutual information between the other features in the feature set F′; φmean (f t ) is the Shapley additive explanatory value of the current feature in the benchmark classification model.

[0094] During actual execution, the embodiment of the present application may combine the first state component, the second state component, and the third state component to represent the current state after determining the first state component, the second state component, and the third state component.

[0095] The embodiments of the present application can represent the current state by utilizing the correlation between the current feature and the label, the redundancy of the current feature and other features in the feature set, and the classification importance of the current feature in the baseline classification model. It can build a more comprehensive state representation and support the modeling of complex tasks and dynamic feature weight adjustment.

[0096] Optionally, in one embodiment of the present application, the finite frequency diffusion feature search strategy includes: constructing an initial feature set based on the mutual information sorting of features and labels; selecting a seed feature set from the initial feature set; performing a dynamic pruning search within the homologous frequency band neighborhood of the seed features of the seed feature set, and after all seed feature searches are completed, performing the next round of seed feature selection and dynamic pruning search until the target number of rounds is reached, and outputting the feature subset corresponding to the minimum loss.

[0097] It can be understood that the limited frequency diffusion feature search method can be a frequency-based feature selection method, which reduces the search space and improves the search efficiency by limiting the frequency range or diffusion range of the search; dynamic pruning search can be an optimization strategy that dynamically prunes unimportant or redundant branches during the search process, aiming to reduce the search space and improve the search efficiency.

[0098] In the specific implementation process, the embodiment of the present application can optimize the feature selection strategy according to the finite frequency diffusion feature search method, make feature addition decisions with the ε decay strategy, and at the same time improve the initial feature set. Path pruning is performed based on the change in classifier loss after the candidate features are added. After all seed feature searches are completed, the next round of seed feature selection and dynamic pruning search is performed until the target number of rounds is reached and the feature subset corresponding to the minimum loss is output. The finite frequency diffusion feature search method can be implemented by the following steps:

[0099] 1) Based on the mutual information ranking of features and labels, construct the initial feature set FB;

[0100] 2) Randomly select a seed feature set F′ from FB;

[0101] 3) Perform dynamic pruning search in the neighborhood of the same frequency band of the seed feature:

[0102] a) Control the exploration-exploitation balance with a decaying random rate ∈ to choose whether to add candidate features;

[0103] The decay of the random rate ∈ follows ∈ = ∈ 0 × exp(-t / τ);

[0104] Among them, ∈0 is the initial random rate, t is the number of times the feature is selected, and τ is the time constant.

[0105] b) If the classification loss decreases after adding features, add F′ and update FB and continue searching, otherwise terminate the current branch;

[0106] 4) After all seed feature searches are completed, the next round of random seed feature selection and dynamic pruning search is performed until the target number of rounds n is reached, and the optimal feature subset F corresponding to the minimum loss is output * .

[0107] The embodiment of the present application can optimize the feature selection strategy through a limited frequency diffusion feature search method, that is, optimize the transfer strategy of the Markov decision process, limit the frequency range of the search, focus on the key frequency band, and efficiently and accurately select key features in the frequency domain, thereby improving the efficiency, quality and performance of feature selection.

[0108] In order to facilitate those skilled in the art to better understand the Markov decision process for feature selection established in this application, a specific embodiment is described in detail below. Figure 2 As shown, in step S205, the Markov decision process includes four main elements, namely state S, action A, reward R, and transfer strategy P. The deep Q network can be used to achieve approximate optimization of the state-action strategy, and realize efficient exploration of the feature space and iteration of the value function.

[0109] In the feature screening problem, the state S is defined as a vector, which specifically includes the correlation between the current feature and the label, expressed as mutual information:

[0110]

[0111] Among them, C is the feature corresponding label, f t The features currently searched.

[0112] It also includes the redundancy of the current feature and other features in the feature set, expressed as the sum of mutual information:

[0113]

[0114] Where F′ refers to the currently selected feature set.

[0115] It also includes the classification importance of the current feature in the benchmark classification model, expressed as the mean of the SHAP absolute value:

[0116]

[0117] Define state S as:

[0118]

[0119] Action A is defined as adding or not adding the feature to the feature set:

[0120] A t ={S(f t ),E(f t )}(3)

[0121] Among them, A t Denotes the action for the t-th feature, S(f t ) means adding the current feature f to the feature set t ,E(f t ) means not to add the current feature f to the feature set t .

[0122] The reward R is defined as the negative change in loss after adding the feature to the feature set:

[0123]

[0124] Among them, R t Indicates adding feature f to the feature set t Instant rewards, Indicates adding feature f t Before, the classification loss of the feature set in the classifier, Indicates adding feature f t Finally, the classification loss of the feature set in the classifier.

[0125] The transfer strategy P is defined as a finite frequency diffusion feature search method. The finite frequency diffusion feature search method is used to optimize the Markov decision process transfer strategy P. The method specifically includes:

[0126] 1) Set all Q tables to 0. For N types of features, sort the features based on the mutual information between each feature and the label to obtain N ordered feature sequences. Assume that the length of each sequence is the same as M, that is, each type of feature has M sub-frequency features.

[0127] 2) Select p highest mutual information features from N*M features to construct the basic feature set FB.

[0128] 3) Randomly select a sub-feature set F′ of length k from FB, input F′ into the classifier, and calculate the classifier loss The frequencies corresponding to the features in F′ are the sub-seed frequency set TF′.

[0129] 4) Among N features, search for feature f that is not in FB before and after k identical seed frequencies t , record the current F′ feature set state S t . Select A with a gradually decaying random rate ∈ t ①When rand()<∈, randomly decide whether to t Add F'; ② Otherwise, according to Q(S t ,A t ) Select the action that maximizes the Q value.

[0130] 5) If the current choice is to add f t , F′=F′∪{f t} Input into the machine learning classifier and calculate the classifier loss Calculate the addition f t The immediate reward R t , that is, the classifier loss changes negatively:

[0131] If R t >0, the feature f t Add the basic feature set FB and continue searching.

[0132] If the current choice is not to add f t or R t ≤0, then let R t =0, and stop searching for the feature in this frequency range.

[0133] 6) (S t ,A t ,R t ) is stored in memory.

[0134] 7) If, among the N types of features, the features added within the seed frequency range all increase the classifier loss, then the search for the seed frequency is terminated and the next seed frequency is searched sequentially.

[0135] 8) When the search for k seed frequencies is completed, update ∈ and proceed to the next round of search, that is, repeat steps (3-8) and record the current feature set F′ and its corresponding classifier loss

[0136] 9) After repeated search for n rounds, compare all the training results. smallest The corresponding feature subset is the optimal feature subset F * .

[0137] The embodiment of the present application utilizes a finite frequency diffusion feature search method to optimize the Markov decision process transfer strategy P, which can be further optimized by Figure 3 Understand in detail.

[0138] Efficient exploration of feature space and iteration of value functions through deep Q-network can include the following steps:

[0139] The update rule for the target Q value is defined as:

[0140]

[0141] in, Refers to the target network parameter θ - The maximum Q value of the next state is calculated, γ refers to the discount factor, Q target (S t ,A t ) indicates that in the current state S t Next, take action A t The ideal total return that can be obtained, R t Indicates adding feature f to the feature set t Instant rewards.

[0142] Construct Q value prediction network Q(S t ,A t ;θ) and the target network Q(S t ,A t θ - ), initialize the target network parameters θ - =θ;

[0143] To train the DQN neural network, the mean square error is used to calculate the loss of the Q network:

[0144]

[0145] Among them, L(θ) represents the error between the Q network prediction value and the target value, It represents the average of a batch of samples, and θ represents the network parameters.

[0146] During the training process, DQN updates the network parameters θ by minimizing the loss function L(θ);

[0147] Periodically synchronize the target network parameters, that is, update θ every C steps - ←θ, in the embodiment of the present application, C can be set to 10.

[0148] After determining the optimal subset for disease identification, machine learning-based feature importance assessment methods can be used to verify the accuracy of the determined disease-specific optimal subset. For example, an XGBoost machine learning classifier can be used to train the XGBoost classifier using the selected optimal feature subset of length m and a noise feature set outside the optimal subset of length 0.5m. After model training is complete, the feature importance is ranked. If more than 80% of the top m ranked features belong to the optimal subset, the screening is considered effective.

[0149] The determined optimal feature subset and its corresponding blood oxygen activity frequency can be considered as features and frequencies sensitive to the target population. For example, after feature screening, 40 features were selected into the optimal feature set, among which the sample entropy feature set of the low-frequency (0.01-0.05Hz) oxygenated hemoglobin concentration signal in the left frontal lobe was selected into the optimal feature subset. Furthermore, after feature importance ranking verification using the XGBoost model, 33 features in the optimal feature subset still appeared in the top 40 feature importance rankings despite the influence of 50% interference features. This preliminary conclusion indicates that the selected optimal feature subset can generally reflect changes in blood oxygen characteristics in stroke patients, and that changes in the sample entropy feature of the low-frequency oxygenated hemoglobin concentration signal in the left frontal lobe are concentrated in stroke patients. This provides a certain characteristic basis for clinical analysis of the condition.

[0150] According to the brain oxygen frequency domain feature screening method proposed in the embodiment of the present application, at least one frequency domain feature can be extracted from the acquired multi-position and multi-type brain oxygen time series data, and a Markov decision process based on the finite frequency diffusion feature search strategy is constructed based on the frequency domain feature. An approximate optimization strategy of the state-action strategy is performed through a deep Q network. Combined with the characteristics of the frequency concentration of brain oxygen signal features and the similar distribution of blood oxygen homologous features, the feature selection strategy is optimized according to the finite frequency diffusion feature search method to obtain the optimal feature subset for disease identification, thereby optimizing the feature selection logic, reducing the repeated selection of highly dependent features, reducing the time of model training, and providing effective feature guidance for disease analysis. Thus, the problem that the current feature screening method of near-infrared signals in the related art often ignores the key factors such as the interaction between features, the comprehensive characterization of the multidimensional features of brain oxygen signals, feature redundancy, and computational cost, resulting in significant deficiencies in efficiency, robustness, and interpretability, lacks the ability to dynamically optimize multidimensional brain oxygen data, and is difficult to determine disease features, etc.

[0151] Next, refer to the attached Figure 4 The present invention describes a brain oxygen frequency domain feature screening device proposed in an embodiment of the present application.

[0152] Figure 4 Schematic diagram of a brain oxygen frequency domain feature screening device according to an embodiment of the present application.

[0153] like Figure 4 As shown, the brain oxygen frequency domain feature screening device 10 includes: an acquisition module 100, which is used to respectively collect first near-infrared signal data of a target population that meets preset normal conditions and second near-infrared signal data of a target disease population; a processing module 200, which is used to preprocess the first near-infrared signal data and the second near-infrared signal data to obtain multi-position and multi-type brain oxygen time series data; a screening module 300, which is used to extract at least one frequency domain feature from the brain oxygen time series data, and based on the at least one frequency domain feature, combined with the frequency concentration of brain oxygen signal features and the similarity in the distribution of blood oxygen homologous features, establish a Markov decision process based on a finite frequency diffusion feature search strategy, and use a deep Q network to perform approximate optimization training of the state-action strategy to obtain the optimal feature subset for disease identification.

[0154] Optionally, in one embodiment of the present application, the screening device 10 of the embodiment of the present application further includes: a verification module.

[0155] Among them, the verification module is used to add the optimal feature subset and the interference feature subset to the pre-trained machine learning classification model to verify the rationality of the optimal feature subset.

[0156] Optionally, in one embodiment of the present application, the screening module 300 includes: a division unit, a calculation unit, and a generation unit.

[0157] Among them, the division unit is used to divide the brain oxygen time series data into multiple sub-frequency bands; the calculation unit is used to calculate the frequency domain characteristics of the blood oxygen signal of each type and / or each position based on the multiple sub-frequency bands; the generation unit is used to normalize the frequency domain characteristics of the blood oxygen signal of each type and / or each position to generate at least one frequency domain feature.

[0158] Optionally, in one embodiment of the present application, the brain oxygen frequency domain feature screening device 10 also includes: a first state component determination module, which is used to determine the first state component by the mutual information between the current feature and the label before obtaining the optimal feature subset for disease identification; a second state component determination module, which is used to determine the second state component by the sum of the mutual information between the current feature and other features before obtaining the optimal feature subset for disease identification; a third state component determination module, which is used to train a pre-built basic classification model based on all features before obtaining the optimal feature subset for disease identification, and calculate the static Shapley additive explanation value of each feature to obtain the absolute value mean, and determine the third state component based on the absolute value mean; a construction module, which is used to determine the current state based on the first state component, the second state component and the third state component before obtaining the optimal feature subset for disease identification, so as to determine the immediate reward and construct a Markov decision process.

[0159] Optionally, in one embodiment of the present application, the screening module 300 is further used to: construct an initial feature set based on the mutual information sorting of features and labels; select a seed feature set from the initial feature set; perform a dynamic pruning search within the homologous frequency band neighborhood of the seed features of the seed feature set, and after all seed feature searches are completed, perform the next round of seed feature selection and dynamic pruning search until the target number of rounds is reached, and output the feature subset corresponding to the minimum loss.

[0160] It should be noted that the aforementioned explanation of the embodiment of the brain oxygen frequency domain feature screening method is also applicable to the brain oxygen frequency domain feature screening device of this embodiment, and will not be repeated here.

[0161] According to the brain oxygen frequency domain feature screening device proposed in the embodiment of the present application, at least one frequency domain feature can be extracted from the acquired multi-location and multi-type brain oxygen time series data, and a Markov decision process based on the finite frequency diffusion feature search strategy is constructed based on the frequency domain feature. An approximate optimization strategy of the state-action strategy is performed through a deep Q network. Combined with the characteristics of the frequency concentration of brain oxygen signal features and the similar distribution of blood oxygen homologous features, the feature selection strategy is optimized according to the finite frequency diffusion feature search method to obtain the optimal feature subset for disease identification, thereby optimizing the feature selection logic, reducing the repeated selection of highly dependent features, reducing the time of model training, and providing effective feature guidance for disease analysis. Thus, the problem that the current feature screening method of near-infrared signals in the related art often ignores the key factors such as the interaction between features, the comprehensive characterization of the multidimensional features of brain oxygen signals, feature redundancy, and computational cost, resulting in significant deficiencies in efficiency, robustness, and interpretability, lacks the ability to dynamically optimize multidimensional brain oxygen data, and is difficult to determine disease features, etc.

[0162] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. The electronic device may include:

[0163] Memory 501 , processor 502 , and computer programs stored in the memory 501 and executable on the processor 502 .

[0164] When the processor 502 executes the program, the brain oxygen frequency domain feature screening method provided in the above embodiment is implemented.

[0165] Furthermore, the electronic device further includes:

[0166] The communication interface 503 is used for communication between the memory 501 and the processor 502 .

[0167] The memory 501 is used to store computer programs that can be run on the processor 502 .

[0168] The memory 501 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0169] If the memory 501, processor 502, and communication interface 503 are implemented independently, the communication interface 503, memory 501, and processor 502 can be interconnected via a bus and communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 5 Only one thick line is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0170] Optionally, in a specific implementation, if the memory 501, the processor 502 and the communication interface 503 are integrated on a chip, the memory 501, the processor 502 and the communication interface 503 can communicate with each other through an internal interface.

[0171] The processor 502 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0172] An embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned brain oxygen frequency domain feature screening method.

[0173] An embodiment of the present application also provides a computer program product, which stores a computer program, and when the program is executed by a processor, implements the above brain oxygen frequency domain feature screening method.

[0174] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0175] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of this application, "N" means at least two, for example, two, three, etc., unless otherwise specifically defined.

[0176] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or N executable instructions for implementing a custom logical function or process step, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may be performed in a different order than shown or discussed, including performing functions in a substantially simultaneous manner or in a reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present application pertain.

[0177] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or N wires (electronic devices), a portable computer disk cartridge (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and programmable read-only memory (EPROM or flash memory), fiber optic devices, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program can be obtained electronically by optically scanning the paper or other medium and then editing, interpreting or processing it in other suitable ways as necessary, and then storing it in a computer memory.

[0178] It should be understood that various parts of the present application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiment, the N steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. If implemented using hardware, as in another embodiment, it can be implemented using any one or a combination of the following technologies known in the art: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0179] Those skilled in the art will understand that all or part of the steps in the method of the above embodiment can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiment.

[0180] In addition, the functional units in the various embodiments of the present application may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into a module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0181] The storage medium mentioned above may be a read-only memory, a magnetic disk, or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present application. Persons skilled in the art may make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present application.

Claims

1. A method for screening brain oxygen frequency domain features, characterized in that: The following steps are involved: Collect first near-infrared signal data of a target population that meets preset normal conditions and second near-infrared signal data of a target disease population respectively; Preprocessing the first near-infrared signal data and the second near-infrared signal data to obtain multi-location and multi-type brain oxygen time series data; At least one frequency domain feature is extracted from the brain oxygen time series data. Based on the at least one frequency domain feature and combined with the similarity in the frequency concentration of brain oxygen signal features and the distribution of blood oxygen homologous features, a Markov decision process based on a finite frequency diffusion feature search strategy is established. Approximate optimization training of the state-action strategy is performed using a deep Q network to obtain the optimal feature subset for disease identification.

2. The method according to claim 1, characterized in that Also includes: The optimal feature subset and the interference feature subset are added to a pre-trained machine learning classification model to verify the rationality of the optimal feature subset.

3. The method according to claim 1, characterized in that The extracting at least one frequency domain feature from the brain oxygen time series data includes: Dividing the brain oxygen time series data into a plurality of sub-bands; Calculating frequency domain features of the blood oxygen signal for each type and / or each position based on the multiple sub-frequency bands; Normalization is performed on the frequency domain features of the blood oxygen signal of each type and / or each position to generate the at least one frequency domain feature.

4. The method according to claim 1, wherein The Markov decision process based on the finite frequency diffusion feature search strategy is established, including: Determine the first state component by the mutual information between the current feature and the label; Determine a second state component by summing the mutual information between the current feature and other features; Training a pre-built base classification model based on all features and calculating a static Shapley additive explanatory value for each feature to obtain an absolute value mean, and determining a third state component based on the absolute value mean; The current state is determined according to the first state component, the second state component, and the third state component to determine an immediate reward and construct the Markov decision process.

5. The method according to claim 4, characterized in that The expression of the current state is: Among them, S t is the current state; is the first state component; is the second state component; is the third state component; f t is the current feature; C is the label; I(C; f t ) is the current feature f t Mutual information with the label; f ′ is the feature set F ′ Other features in I(f t ,f ′ ) is the current feature f t With feature set F ′ Mutual information between other features in ∑ f′∈F′ I(f t ,f′) is the current feature f t The sum of the mutual information between the other features in the feature set F′; φ mean (f t ) is the Shapley additive explanatory value of the current feature in the benchmark classification model.

6. The method according to claim 1, characterized in that The finite frequency diffusion feature search strategy includes: Construct the initial feature set based on the mutual information ranking of features and labels; Selecting a seed feature set from the initial feature set; A dynamic pruning search is performed in the homologous frequency band neighborhood of the seed feature of the seed feature set, and feature addition decisions are made with the ε decay strategy. At the same time, the initial feature set is improved, and path pruning is performed based on the change in classifier loss after the candidate features are added. After all seed feature searches are completed, the next round of seed feature selection and dynamic pruning search is performed until the target number of rounds is reached, and the feature subset corresponding to the minimum loss is output.

7. A brain oxygen frequency domain feature screening device, characterized in that: include: An acquisition module is used to respectively acquire first near-infrared signal data of a target population meeting preset normal conditions and second near-infrared signal data of a target disease population; a processing module, configured to pre-process the first near-infrared signal data and the second near-infrared signal data to obtain multi-location and multi-type brain oxygen time series data; The screening module is used to extract at least one frequency domain feature from the brain oxygen time series data, and based on the at least one frequency domain feature, combined with the frequency concentration of brain oxygen signal features and the similar distribution of blood oxygen homologous features, establish a Markov decision process based on a finite frequency diffusion feature search strategy, and use a deep Q network to perform approximate optimization training of the state-action strategy to obtain the optimal feature subset for disease identification.

8. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the brain oxygen frequency domain feature screening method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: The program is executed by a processor to implement the brain oxygen frequency domain feature screening method according to any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that The computer program is executed to implement the brain oxygen frequency domain feature screening method according to any one of claims 1 to 6.

Images