A Deep Learning-Based Method for Multi-Angle Feature Extraction and Selection of Magnetocardiogram Signals

Abstract

This invention relates to a deep learning-based method for multi-angle feature extraction and selection of magnetocardiogram (MCG) signals. The method includes: preprocessing the MCG signal; designing a temporal feature extraction method based on learnable positional encoding and a CNN-Transformer deep learning model; a frequency domain feature extraction method based on HHT and a CNN deep learning model; and a multi-modal feature extraction method based on adaptive VMD and center frequency weighting. A feature selection model based on adaptive LASSO is used to select the subset of MCG features that has the greatest impact on disease classification. Compared with traditional MCG feature extraction and selection methods, the proposed feature extraction and selection method can automatically and more accurately extract multi-angle MCG features, and the extracted features have more angles, better robustness, higher reliability, and a wider range of applicable disease classifications.

Description

Technical Field

[0001] This invention belongs to the field of biomedical signal analysis and relates to a method for multi-angle feature extraction and selection of magnetic heart signals based on deep learning. Background Technology

[0002] With the continuous development of biomedical signal analysis technology, magnetocardiography (MCG), as an emerging non-invasive detection method, has gradually attracted researchers' attention due to its advantages such as high sensitivity, non-contact detection, no radiation, immunity to media interference, and short detection time. Magnetic cardiography signals can reflect the electrophysiological activity of the heart, exhibiting high temporal and spatial resolution, providing important evidence for the early diagnosis and monitoring of heart diseases. However, traditional methods for classifying magnetic cardiography signals mainly rely on manual feature extraction and time-frequency domain analysis. These methods are typically difficult to implement, require domain expert knowledge, have low feature robustness, and are easily affected by noise and signal variations.

[0003] Traditional time-domain feature extraction utilizes statistical methods to calculate basic statistical characteristics of a signal, such as mean, variance, peak value, and waveform width. Frequency-domain feature extraction uses methods like Fourier transform to convert the signal to the frequency domain and extract frequency components, such as power spectral density and dominant frequency. Time-frequency domain feature extraction combines information from both the time and frequency domains, using methods like wavelet transform or short-time Fourier transform to extract features of the signal at different times and frequencies. While these traditional methods are effective to some extent, they often rely on the knowledge of medical experts, leading to a high degree of subjectivity in the feature extraction process and making it difficult to adapt to different signal characteristics. Furthermore, traditional feature extraction methods have limited feature representation capabilities, typically only capturing local features of the signal and failing to fully reflect its complexity.

[0004] In recent years, the rapid development of deep learning technology has provided new solutions for magnetocardiogram (MCC) signal analysis. By constructing multi-layered neural networks, deep learning can automatically learn multi-faceted features of signals, reducing reliance on expert knowledge and improving the objectivity and consistency of feature extraction. Deep learning methods also possess powerful feature representation capabilities, capturing complex nonlinear relationships in signals, fully mining multi-level feature information, and enhancing the model's expressive power. Furthermore, deep learning models exhibit strong robustness in handling noise and interference, effectively improving the accuracy and reliability of classification. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, this invention provides a method for multi-angle feature extraction and selection of magnetocardiogram (MCG) signals based on deep learning. This method fully considers the time-domain, frequency-domain, and multimodal features of MCG signals, and establishes a comprehensive feature extraction framework through deep learning to achieve the goal of extracting features containing more information. Different feature subsets are selected for different classification tasks, achieving good classification results in the diagnosis of various diseases. Through automated feature extraction and classification mechanisms, the method improves the automatic disease classification capability, thereby providing a more effective tool for clinical diagnosis.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for multi-angle feature extraction and selection of magnetocardiogram signals based on deep learning, the method comprising the following steps:

[0008] Step S1 involves preprocessing the raw magnetic field signals based on the matrix arrangement characteristics of the magnetic field signals acquired using a 36-channel atomic magnetometer. This preprocessing includes handling outliers, scaling data, and input dimension processing.

[0009] Step S2 involves extracting multi-angle magnetic field features from the preprocessed magnetocardiogram (MCC) data obtained in Step S1. This includes extracting temporal features based on learnable positional encoding and a CNN-Transformer deep learning model, extracting frequency features based on HHT and a CNN deep learning model, and extracting temporal features based on adaptive VMD and a center frequency weighting method.

[0010] Step S3: Linearly concatenate the time-domain features, frequency-domain features, and multimodal features of the magnetic field of heart (MPH) extracted in step S2 to obtain multi-angle MPH signal features. Use an adaptive LASSO-based feature selection model to select multi-angle MPH signal features for different disease classification tasks.

[0011] The specific process for handling outliers, scaling data, and input dimension data as described in step S1 is as follows:

[0012] The data outlier processing refers to performing missing value filling and outlier scaling operations on all acquired raw magnetocardiogram data, so that each sample of magnetocardiogram data constitutes 1×36000 data points.

[0013] The data scaling process refers to the standardization operation of all 1×36000 magnetocardiogram data after outlier processing, using the min-max standardization method to make the mean 0 and the variance 1, thereby eliminating the error caused by individual differences among different subjects.

[0014] The data input dimension processing refers to designing two different data input dimensions for the magnetocardiogram (MCC) data after outlier processing and scaling, to accommodate different feature extraction methods in subsequent steps. For the MCC temporal feature extraction method based on learnable positional encoding and CNN-Transformer deep learning model, the MCC data dimension of each sample needs to be adjusted to 6×6×1000 to fully simulate the arrangement characteristics of the MCC signal sensor and extract the correlation features between each channel. For the MCC frequency domain feature extraction method based on HHT and CNN deep learning model and the MCC multimodal feature extraction method based on adaptive VMD and center frequency weighting, the MCC data dimension of each sample needs to be adjusted to 36×1000, and then each channel is transformed and feature extracted sequentially.

[0015] The method for extracting temporal features of magnetocardiography based on learnable positional encoding and CNN-Transformer deep learning model described in step S2 specifically refers to:

[0016] Building upon the traditional Transformer architecture, a method was designed for the encoder to form positional codes by calculating sine and cosine values ​​from angle values. This generates a CNN-Transformer model with learnable positional codes. After encoding, a multi-head attention mechanism and a multilayer perceptron are added, assigning different weights to features. The resulting integrated model yields the temporal features of the magnetocardiogram (MCC) signal. This method for extracting MCC temporal features based on learnable positional codes and a CNN-Transformer deep learning model can focus on different parts of the signal, dynamically adjust the degree of attention to different features, capture the relationships and interactions between multiple features, and thus extract more representative dynamic global temporal features.

[0017] The CNN model refers to a Convolutional Neural Network (CNN), which can automatically extract features from input data without the need for manually designed feature extractors.

[0018] The Transformer model refers to a deep learning model that assigns different weights to different features according to their importance and discriminative power, focusing on different parts of the signal and capturing the relationships and interactions between multiple features. It typically consists of a positional encoder, encoder, decoder, and self-attention mechanism. In this way, the model can dynamically adjust the degree of attention given to different features, thereby extracting more representative dynamic global temporal features.

[0019] The method of calculating sine and cosine values ​​using angle values ​​to form position codes refers to defining a position index for each position in the input sequence, then calculating a frequency factor based on the embedding dimension, combining the position index with the frequency using a formula to generate angle values, and finally calculating these angle values ​​using sine and cosine functions to form position codes, where even-numbered dimensions use sine values ​​and odd-numbered dimensions use cosine values.

[0020] The method for extracting magnetocardiographic frequency domain features based on HHT and CNN deep learning models described in step S2 specifically refers to:

[0021] A feature extraction model combining the Hilbert-Huang Transform (HHT) and CNN was designed. 36 channels of magnetocardiogram (MCC) signal data were sequentially input into the model. First, the HHT was applied to obtain 36×9 Hilbert spectra. Then, a CNN model consisting of convolutional layers, pooling layers, and dense layers further extracted deep-level features from the Hilbert spectra. The resulting output was the frequency domain features of the MCC signal. This MCC frequency domain feature extraction method based on the HHT and CNN deep learning model can not only automatically learn frequency domain features but also extract local features through convolutional layers and reduce feature dimensionality through pooling layers, thereby capturing important dynamic changes in the MCC signal. This process significantly improves the accuracy of feature extraction, reduces reliance on expert knowledge, and provides a solid data foundation for subsequent analysis.

[0022] The Hilbert-Huang transform is a method applicable to the analysis of nonlinear and non-stationary signals. It decomposes the signal into several eigenmode functions through empirical mode decomposition, and then performs Hilbert transform on the eigenfunctions to effectively extract the instantaneous frequency and amplitude information of the signal.

[0023] Empirical Mode Decomposition (EMD) is a signal processing method used to analyze nonlinear and non-stationary signals. It can extract intrinsic mode functions from signals, thereby revealing the local features and time-varying characteristics of signals. It is widely used in fields such as fault diagnosis and medical signal processing.

[0024] The Hilbert transform is a method that transforms a signal through convolution operations to obtain the phase difference between the original signal and the transformed signal. The Hilbert transform is an analytical method that combines the original signal with the Hilbert transform to form a complex signal. Through the Hilbert transform, the instantaneous phase and frequency of the signal can be obtained, helping to analyze the frequency variation characteristics of the signal.

[0025] The method for extracting multimodal features of magnetic cardiomyography based on adaptive VMD and center frequency weighting, as described in step S2, specifically refers to:

[0026] An adaptive mode number optimization method based on the sparsity index of magnetocardiogram (MCC) signals is designed on the basis of the traditional variational mode decomposition (VMD) method. The optimal mode number is used to decompose the MCC signal and extract features. Then, each mode is weighted by its center frequency, and a weighted average of the features is performed to obtain the multimodal features of the MCC signal. This is achieved by setting VMD parameters and obtaining the adaptive VMD mode number. The magnetocardiogram (MCC) signal is decomposed using Virtual Mode Decomposition (VMD), which breaks down the complex signal into K modes with specific frequency characteristics. Each mode is then weighted by its center frequency, and a feature-weighted average is performed to extract the signal's latent patterns and trends, ensuring that the extracted multimodal features are both comprehensive and representative.

[0027] The adaptive mode number optimization method for the sparsity index of the magnetic heart signal refers to determining the optimal decomposition mode number based on the strategy of calculating sparsity based on the marginal spectrum, and then performing decomposition using this mode number.

[0028] The VMD parameters include: penalty factor, noise tolerance, DC component, initialization method, and convergence tolerance.

[0029] The multimodal features include the time-domain features of each mode: mean, variance, kurtosis and skewness; and the frequency-domain features: spectral amplitude, total energy, center frequency, spectral bandwidth and maximum frequency.

[0030] The adaptive LASSO feature selection model mentioned in step S3 specifically refers to:

[0031] The temporal features of magnetocardiography (MCG) extracted in step S2 using the method based on learnable positional encoding and the CNN-Transformer deep learning model, the frequency features of MCG extracted using the method based on HHT and the CNN deep learning model, and the multimodal features of MCG extracted using the method based on adaptive VMD and center frequency weighting are linearly concatenated and used as input to the adaptive Least Absolute Shrinkage and Selection Operator (LASSO) model. First, an adaptive LASSO regression model is constructed, aiming to minimize the loss function including the L1 regularization term to control model complexity. Then, the optimal regularization parameter is selected through cross-validation, and the model is trained using an optimization algorithm to solve for the feature coefficients. Finally, during training, adaptive LASSO compresses the coefficients of some features to zero, thereby achieving feature selection. Adaptive LASSO can automatically remove irrelevant or redundant features while retaining important features, which not only improves the interpretability of the model but also reduces the risk of overfitting, making the final classification model more concise and efficient. By performing adaptive LASSO regression analysis on the extracted multi-angle features of the magnetic field, the feature set can be further optimized, providing more accurate input for subsequent classification tasks and ensuring effectiveness and reliability in practical applications.

[0032] The subsequent classification task refers to classifying the features extracted and selected by the feature extraction and selection model using a Support Vector Machine (SVM) model in machine learning. The feature extraction and selection method of this invention can be applied to various heart-related diseases, with a focus on achieving high-precision classification between healthy individuals and patients with coronary artery stenosis. The classification model assigns different labels to healthy individuals and disease patients, and classifies these labels based on the different features extracted from these samples, identifying the feature differences between samples with different labels. During the classification process, all samples are randomly shuffled and divided into training and test sets in an 8:2 ratio. The SVM model is trained using the sample data from the training set, and then used to predict the sample data from the test set, calculating various evaluation metrics, including accuracy, sensitivity, specificity, F1 score, and AUC value.

[0033] The beneficial effects of this invention are as follows:

[0034] This invention employs a temporal feature extraction method for magnetocardiography based on learnable positional encoding and a CNN-Transformer deep learning model. Different features are assigned different weights according to their importance and discriminative power to focus on different parts of the signal and capture the relationships and interactions between various features. In this way, the model can dynamically adjust the degree of attention given to different features, thereby extracting more representative dynamic global temporal features.

[0035] This paper employs a magnetocardiogram (MCG) frequency domain feature extraction method based on Hidden Sensory Hormone Extraction (HHT) and CNN deep learning models. HHT effectively handles nonlinear and non-stationary signals by adaptively decomposing the signal into multiple intrinsic mode functions, thereby capturing more detailed frequency domain information. Furthermore, CNN possesses powerful feature extraction and classification capabilities, automatically learning important features in the data and performing effective pattern recognition. Combining these two approaches, HHT provides high-quality input features for CNN, making the model more accurate and robust when processing complex signals.

[0036] A multimodal feature extraction method based on adaptive VMD and center frequency weighting for magnetocardiography can adaptively decompose a signal into multiple mode functions with different frequency components. These modes can not only better capture the instantaneous characteristics of the signal, but also effectively suppress the influence of noise. Unlike wavelet transform, which relies on fixed basis functions and the window length selection of short-time Fourier transform, VMD achieves fine decomposition of the signal through optimization algorithms, avoiding the uncertainty caused by manual parameter settings.

[0037] By constructing an adaptive LASSO regression model, the loss function including L1 regularization is minimized to control model complexity. Then, the optimal regularization parameter is selected through cross-validation, and the model is trained using an optimization algorithm to solve for feature coefficients. Finally, during training, adaptive LASSO compresses the coefficients of some features to zero, thus achieving feature selection. Adaptive LASSO can automatically remove irrelevant or redundant features while retaining important ones, improving model interpretability and reducing the risk of overfitting, resulting in a more concise and efficient classification model. By performing adaptive LASSO regression analysis on extracted multi-angle magnetocardiogram features, the feature set can be further optimized, providing more accurate input for subsequent classification tasks and ensuring effectiveness and reliability in practical applications. Furthermore, adaptive LASSO possesses strong interpretability, providing a concise and practically meaningful model suitable for feature selection and modeling needs in real-world applications. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the multi-angle feature extraction and selection method for magnetocardiogram signals based on deep learning, as described in this invention.

[0039] Figure 2 This is a flowchart of the preprocessing operation.

[0040] Figure 3 This is an architecture for a magnetocardiogram temporal feature extraction method based on learnable positional encoding and a CNN-Transformer deep learning model;

[0041] Figure 4 This is an architecture for a magnetocardiogram frequency domain feature extraction method based on HHT and CNN deep learning models;

[0042] Figure 5 This is the architecture of a multimodal feature extraction method for magnetocardiography based on adaptive VMD and center frequency weighting;

[0043] Figure 6 This is the overall architecture of the multi-angle feature extraction method of the present invention. Detailed Implementation

[0044] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings and examples.

[0045] like Figure 1 As shown, the deep learning-based multi-angle feature extraction and selection method for magnetocardiogram signals of the present invention includes the following three steps:

[0046] Step S1: Based on the matrix arrangement characteristics of the magnetocardiogram (MCC) signals acquired using a 36-channel atomic magnetometer, perform more comprehensive preprocessing operations on the MCC signals. These preprocessing operations include first acquiring the MCC signals, then sequentially handling outlier values ​​and scaling the data, and finally processing the input dimension of the data.

[0047] The preprocessing of the magnetic field signal described in step S1 is as follows: Figure 2 As shown, the process includes outlier handling, scaling, and input dimension processing of the raw magnetocardiogram data. The specific steps are as follows:

[0048] The data outlier processing refers to performing missing value filling and outlier scaling operations on all acquired raw magnetocardiogram data, so that each sample of magnetocardiogram data constitutes 1×36000 data points.

[0049] The data scaling process refers to the standardization operation of all 1×36000 magnetocardiogram data after outlier processing, using the min-max standardization method to make the mean 0 and the variance 1, thereby eliminating the error caused by individual differences among different subjects.

[0050] The data input dimension processing refers to designing two different data input dimensions for the magnetocardiogram (MCC) data after outlier processing and scaling, to accommodate different feature extraction methods in subsequent steps. For the MCC temporal feature extraction method based on learnable positional encoding and CNN-Transformer deep learning model, the MCC data dimension of each sample is adjusted to 6×6×1000 to fully simulate the arrangement characteristics of the MCC signal sensor and extract the correlation features between each channel. For the MCC frequency domain feature extraction method based on HHT and CNN deep learning model and the MCC multimodal feature extraction method based on adaptive VMD and center frequency weighting, the MCC data dimension of each sample is adjusted to 36×1000, and then each channel is transformed and feature extracted sequentially.

[0051] Step S2: Based on the preprocessed magnetocardiogram (MCC) data obtained in Step S1, multi-angle MCC feature extraction is performed. Three methods were designed: a temporal MCC feature extraction method based on learnable positional encoding and a CNN-Transformer deep learning model; a frequency domain MCC feature extraction method based on HHT and a CNN deep learning model; and a multimodal MCC feature extraction method based on adaptive VMD and center frequency weighting.

[0052] Step S2.1: Extract temporal features of the magnetocardiogram (MCC) signal using a method based on learnable positional encoding and a CNN-Transformer deep learning model. This model, building upon the traditional Transformer architecture, designs a CNN-Transformer model for the encoder that uses sine and cosine values ​​calculated from angle values ​​to form positional encodings, generating learnable positional encodings. A multi-head attention mechanism is added after encoding, utilizing the CNN model to mine and extract features. These extracted features are then input into the Transformer, assigned different weights, and integrated to obtain the temporal features of the MCC signal. The specific model architecture is as follows: Figure 3 As shown, the model first uses the sample data with adjusted dimensions in step S1 to deeply mine and extract features through a CNN model composed of convolutional and pooling layers with time distribution. After flattening, the data is input into a Transformer. Then, different weights are assigned to the features, and the output is generated through Dropout and fully connected layers. Finally, 512 magnetocardiogram time-domain features are output.

[0053] The magnetocardiogram temporal feature extraction method based on learnable positional encoding and CNN-Transformer deep learning model can focus on different parts of the signal, dynamically adjust the degree of attention to different features, capture the relationship and interaction between multiple features, and thus extract more representative dynamic global temporal features.

[0054] The CNN model refers to a Convolutional Neural Network (CNN), which can automatically extract features from input data without the need for manually designed feature extractors.

[0055] The convolutional layers, pooling layers, dropout layers, and fully connected layers are all key components of a CNN. Convolutional layers extract features from the input data through convolution operations, and then use a sliding window to calculate the weighted sum of local regions, thereby generating features. Pooling layers typically use max pooling or average pooling operations to downsample the feature maps. Dropout layers randomly "drop out" a portion of neurons during training to reduce overfitting. Fully connected layers connect all neurons from the previous layer to every neuron in the current layer, mapping features to the final output space through linear transformations and activation functions.

[0056] The Transformer model refers to a deep learning model that assigns different weights to different features according to their importance and discriminative power, focusing on different parts of the signal and capturing the relationships and interactions between multiple features. It typically consists of a positional encoder, encoder, decoder, and self-attention mechanism. In this way, the model can dynamically adjust the degree of attention given to different features, thereby extracting more representative dynamic global temporal features.

[0057] The method of calculating sine and cosine values ​​using angle values ​​to form position codes refers to defining a position index for each position in the input sequence, then calculating a frequency factor based on the embedding dimension, then combining the position index with the frequency using a formula to generate angle values, and finally calculating these angle values ​​using sine and cosine functions to form position codes, where even-numbered dimensions use sine values ​​and odd-numbered dimensions use cosine values.

[0058] The calculation of the angle value includes:

[0059] For each position in the input sequence Define a location index This indicates the position of the element in the sequence. A frequency factor is defined, typically related to the embedding dimension. Relevant, specifically:

[0060] ,

[0061] in, This indicates the embedded dimension index.

[0062] For each position and frequency Calculate the angle value:

[0063] ,

[0064] The formation of position codes refers to generating position codes using calculated angle values ​​through sine and cosine functions:

[0065] ,

[0066] ,

[0067] This approach ensures that positional encoding alternates between sine and cosine functions in even and odd dimensions, thus giving the encoding between adjacent positions a certain degree of smoothness and periodicity.

[0068] Step S2.2: Extract frequency domain features of the magnetocardiogram (MCC) signal using a method based on HHT and a CNN deep learning model. The 36-channel MCC signal data are sequentially input into the model. First, HHT is used to obtain 36×9 Hilbert spectra. Then, a CNN model consisting of convolutional layers, pooling layers, and Dense layers further extracts deeper features from the Hilbert spectra. After processing and outputting, the frequency domain features of the MCC signal are obtained. The specific model architecture is as follows... Figure 4 As shown, the model first iteratively performs Hilbert-Huang Transform on the sample data after dimensionality adjustment in step S1. This involves using Empirical Mode Decomposition (EMD) to extract several intrinsic mode functions (EMFs), performing a Hilbert Transform on each EMF, calculating the instantaneous amplitude, instantaneous phase, and instantaneous frequency, and storing these values ​​in a four-dimensional feature matrix. The flattened feature matrix is ​​then input into a CNN model, which includes convolutional layers, pooling layers, and fully connected layers. It employs ReLU activation, L2 regularization, Batch Normalization, and Dropout, ultimately outputting 256 magnetic field features.

[0069] The method based on HHT and CNN deep learning models can not only automatically learn frequency domain features, but also extract local features through convolutional layers and reduce feature dimensionality through pooling layers, thereby capturing important dynamic changes in magnetocardiogram signals. This process significantly improves the accuracy of feature extraction, reduces reliance on expert knowledge, and provides a solid data foundation for subsequent analysis.

[0070] The Hilbert-Huang transform is a method applicable to the analysis of nonlinear and non-stationary signals. It decomposes the signal into several intrinsic mode functions through empirical mode decomposition, which can effectively extract the instantaneous frequency and amplitude information of the signal.

[0071] The empirical mode decomposition (EMD) is used to analyze nonlinear and non-stationary signals. It can extract the intrinsic mode functions (EMFs) from the signal and then perform Hilbert transform on the EMFs to effectively extract the instantaneous frequency and amplitude information of the signal, thereby revealing the local features and time-varying characteristics of the signal. It is widely used in fields such as fault diagnosis and medical signal processing.

[0072] The Hilbert transform is a method that transforms a signal through convolution operations to obtain the phase difference between the original signal and the transformed signal. The Hilbert transform is an analytical method that combines the original signal with the Hilbert transform to form a complex signal. Through the Hilbert transform, the instantaneous phase and frequency of the signal can be obtained, helping to analyze the frequency variation characteristics of the signal.

[0073] The ReLU activation function, L2 regularization, and Batch Normalization are all key components of CNNs. The ReLU activation function introduces non-linearity, enabling the neural network to learn complex patterns and features; L2 regularization prevents overfitting by penalizing model complexity to improve generalization ability; and Batch Normalization accelerates training, improves model stability and performance, makes the training process more stable, and allows for larger learning rates.

[0074] Step S2.3: Extract multimodal features of the magnetic field signal using an adaptive VMD-based and center frequency weighted method. This method involves designing an adaptive mode number optimization method based on the sparsity index of the magnetic field signal on top of the traditional VMD method. The optimal mode number is used to decompose the magnetic field signal and extract features. Then, each mode is weighted by its center frequency, and a weighted average of the features is performed to obtain the multimodal features of the magnetic field signal. The specific model architecture is as follows... Figure 5 As shown, by setting VMD parameters and obtaining the adaptive number of VMD modes... VMD decomposition was performed on the magnetic field signal to decompose the complex signal into... By analyzing a modality with specific frequency characteristics, more detailed multimodal features can be obtained. Through weighted averaging, the potential patterns and trends of the signal can be extracted, ensuring that the extracted multimodal features are both comprehensive and representative. Finally, 1728 magnetic field multimodal features are output.

[0075] The adaptive mode number optimization method for the sparsity index of the magnetic heart signal refers to determining the optimal decomposition mode number based on the strategy of calculating sparsity based on the marginal spectrum, and then performing decomposition using this mode number.

[0076] The VMD parameters include: penalty factor, noise tolerance, DC component, initialization method, and convergence tolerance.

[0077] The multimodal features include the time-domain features of each mode: mean, variance, kurtosis and skewness; and the frequency-domain features: spectral amplitude, total energy, frequency center, spectral bandwidth and maximum frequency.

[0078] Step S3: Linearly concatenate the temporal, frequency, and multimodal features of the magnetic field signal (MFS) extracted in Step S2 to obtain multi-angle MFS signal features. Use an adaptive LASSO-based feature selection model to select these multi-angle MFS signal features for different disease classification tasks. Linearly concatenate the temporal features extracted in Step S2 using a method based on learnable positional encoding and a CNN-Transformer deep learning model, the frequency features extracted using a method based on HHT and a CNN deep learning model, and the multimodal features extracted using an adaptive VMD and center frequency weighting method. This concatenates these features as input to the adaptive LASSO model. First, an adaptive LASSO regression model is constructed, aiming to minimize the loss function including an L1 regularization term to control model complexity. Then, cross-validation is used to select the optimal regularization parameter, and an optimization algorithm is used to train the model and solve for the feature coefficients. Finally, during training, adaptive LASSO compresses the coefficients of some features to zero, thus achieving feature selection.

[0079] Adaptive LASSO can automatically remove irrelevant or redundant features while retaining important features, which not only improves the interpretability of the model but also reduces the risk of overfitting, making the final classification model more concise and efficient. By performing adaptive LASSO regression analysis on the extracted multi-angle magnetocardiogram features, the feature set can be further optimized, providing more accurate input for subsequent classification tasks and ensuring effectiveness and reliability in practical applications.

[0080] The subsequent classification task refers to classifying the features extracted and selected by the feature extraction and selection model using an SVM (Signal Vector Machine) in machine learning. The feature extraction and selection method of this invention can be applied to various heart-related diseases, with a focus on achieving high-precision binary classification between healthy individuals and patients with coronary artery stenosis. It also achieves good results in binary classification tasks based on the degree of stenosis. The classification model assigns different labels to healthy individuals and patients with diseases, and classifies these labels based on the different features extracted from these samples, identifying the feature differences between samples with different labels. During the classification process, all samples are randomly shuffled and divided into training and test sets in an 8:2 ratio. The SVM model is trained using the sample data from the training set, and then this model is used to predict the sample data from the test set, calculating various evaluation metrics, including accuracy, sensitivity, specificity, F1 score, and AUC value.

[0081] In summary, the deep learning-based multi-angle feature extraction and selection method for magnetocardiogram (MCG) signals described in this invention preprocesses the MCG signals and then designs MCG temporal feature extraction methods based on learnable positional encoding and CNN-Transformer deep learning models, MCG frequency domain feature extraction methods based on HHT and CNN deep learning models, and MCG multimodal feature extraction methods based on adaptive VMD and center frequency weighting. The overall model architecture is as follows: Figure 6 As shown in the diagram. Finally, the subset of multi-angle cardiac features with the greatest impact on classification is selected using an adaptive LASSO-based feature selection model. For a binary classification example involving healthy individuals and patients with coronary artery stenosis, the selected features achieve an accuracy, sensitivity, specificity, and F1 score of over 95%, as well as an AUC value of over 0.98, demonstrating excellent classification performance. Compared to traditional manual feature extraction methods and time-frequency domain analysis feature extraction methods, the method proposed in this invention has the advantages of automatic feature extraction, a wide range of extracted feature angles, a large number of features, high reliability, and broad applicability to disease classification. Therefore, this invention effectively overcomes the shortcomings of existing technologies and has high medical application value.

[0082] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for multi-angle feature extraction and selection of magnetocardiogram signals based on deep learning, characterized in that, The method includes the following steps: Step S1: Based on the matrix arrangement characteristics of the magnetocardiogram signals acquired using a 36-channel atomic magnetometer, preprocess the raw magnetocardiogram signals, including outlier processing, scaling, and input dimension processing. Step S2: Extract multi-angle magnetic field features from the preprocessed raw magnetic field data, including extracting temporal features based on learnable positional encoding and CNN-Transformer deep learning model, extracting frequency features based on Hilbert-Huang Transform (HHT) and CNN deep learning model, and extracting multimodal features based on adaptive VMD and center frequency weighting method. This includes: decomposing the magnetic field signal into the optimal number of modes and extracting statistical features using an adaptive mode number optimization method based on the sparsity index of the magnetic field signal, then weighting each modal feature with the center frequency value, performing feature weighted averaging, and finally sorting and outputting the multimodal features of the magnetic field signal. Step S3: Linearly concatenate the extracted time-domain features, frequency-domain features, and multimodal features of the magnetic field to obtain multi-angle magnetic field signal features. Use an adaptive LASSO-based feature selection model to select multi-angle magnetic field signal features for different disease classification tasks.

2. The method for multi-angle feature extraction and selection of magnetocardiogram signals based on deep learning according to claim 1, characterized in that, In step S1: The outlier processing includes filling missing values ​​and scaling outliers in the original magnetocardiogram data, so that each sample of magnetocardiogram data consists of 1×36000 data points. The data scaling process includes standardizing all the magnetocardiogram data after outlier processing using the min-max standardization method to make its mean 0 and variance 1. The data input dimension processing includes designing two different data input dimensions for the magnetocardiogram (MCC) data after data scaling. For the MCC data based on learnable location encoding and CNN-Transformer deep learning model, the MCC data dimension of each sample is adjusted to 6×6×1000. For the MCC data based on HHT and CNN deep learning model and based on adaptive VMD and center frequency weighting method, the MCC data dimension of each sample is adjusted to 36×1000.

3. The method for multi-angle feature extraction and selection of magnetocardiogram signals based on deep learning according to claim 1, characterized in that, The step S2, which involves extracting temporal features of the magnetic field based on learnable positional encoding and a CNN-Transformer deep learning model, includes: using a CNN model to mine and extract local temporal features of the magnetic field, then inputting the extracted temporal features into a Transformer, calculating sine and cosine values ​​through angle values ​​to form the positional encoding of the Transformer, and adding a multi-head attention mechanism and a multilayer perceptron after the positional encoding, and finally processing the output to obtain the temporal features of the magnetic field signal.

4. The method for multi-angle feature extraction and selection of magnetocardiogram signals based on deep learning according to claim 1, characterized in that, The step S2, which involves extracting the frequency domain features of the magnetic field signal based on the HHT and CNN deep learning model, includes: inputting the preprocessed 36-channel magnetic field signal data sequentially, obtaining 36×9 Hilbert spectra through HHT, and then extracting the deep magnetic field features from each Hilbert spectra through a CNN model composed of convolutional layers, pooling layers, and Dense layers, and finally processing and outputting the frequency domain features of the magnetic field signal.

Images