Consciousness disorder patient classification method based on multi-feature fusion and transfer learning

Abstract

The invention discloses a classification method for patients with disturbance of consciousness based on multi-feature fusion and transfer learning. The classification method is specifically implemented according to the following steps: step 1, data preprocessing; step 2, multi-dimensional feature extraction and fusion; step 3, feature input and model training; step 4, performing minimum error entropy optimization; and step 5, model strategy training optimization. According to the method, the multi-dimensional features are fused, the minimum error entropy is adopted to replace a traditional mean square error to serve as a cost function, the method can better adapt to electroencephalogram signal features of different patients, the problem of data covariable offset caused by transfer learning and individual differences is relieved, and therefore the classification accuracy and the system robustness are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing and machine learning methods, and specifically relates to a method for classifying patients with consciousness disorders based on multi-feature fusion and transfer learning. Background Art

[0002] Disorders of consciousness (DOC) refer to abnormalities in the patient's consciousness state caused by brain injury or disease. Common types include unresponsive wakefulness syndrome (UWS) and minimally conscious state (MCS). In the treatment and prognosis prediction of DOC patients, the following key issues are faced: (1) Traditional behavioral scale monitoring can have a misdiagnosis rate of up to 40% when the patient's speech or non-verbal function is lost, especially when the patient's activities are limited or intermittent, and the accuracy of existing evaluation methods is low; (2) Existing imaging methods are accurate in predicting adverse outcomes (such as brain death), but are not sensitive enough in identifying patients whose consciousness is gradually recovering. At the same time, patients with disorders of consciousness are inconvenient to move, and frequent collection of imaging data is difficult; (3) The patient's clinical data (such as age, time since injury, etiology, etc.) is difficult to provide sufficient basis for treatment plan formulation and prognosis prediction. There is an urgent need to develop more sensitive consciousness level assessment models and personalized treatment plans.

[0003] Electroencephalogram (EEG) signals are important electrophysiological indicators that directly reflect brain activity. They are widely used in clinical fields such as epilepsy and sleep staging, and are important bedside monitoring tools in intensive care units (ICUs). EEG can monitor the brain's electrical activity in real time, helping doctors assess patients' neurological functions. In recent years, it has also been increasingly used in the diagnosis and assessment of patients with impaired consciousness.

[0004] In practical applications, EEG signals are susceptible to interference from non-Gaussian noise (such as power frequency interference, equipment noise, electromyographic artifacts, sweating or poor electrode contact, etc.), which can significantly reduce signal quality and thus affect the accuracy of feature extraction. There is often a significant covariate shift in the source and target domain data of transfer learning. In addition, due to individual differences and the diversity of disease conditions, EEG signals vary significantly between different patients, resulting in a decrease in the generalization ability of the classification model. Traditional methods usually rely on simple feature extraction and classification algorithms, which are difficult to effectively address these challenges, especially when the amount of data is limited. Summary of the invention

[0005] The purpose of the present invention is to provide a method for classifying patients with disorders of consciousness based on multi-feature fusion and transfer learning, which solves the problem of poor accuracy and robustness in the classification of patients with chronic disorders of consciousness in the prior art.

[0006] The technical solution adopted by the present invention is a method for classifying patients with impaired consciousness based on multi-feature fusion and transfer learning, which is specifically implemented in the following steps: Step 1: perform eye movement artifact correction, power frequency noise elimination, bandpass filtering, downsampling, re-reference and bad segment removal on the target domain data, and perform the same preprocessing steps on the source domain data as the target domain data; Step 2: Multi-dimensional feature extraction and fusion; Step 3: Feature input and model training; Step 4: Minimum error entropy optimization; Step 5: Model strategy training and optimization.

[0007] The technical solution of the present invention is also characterized in that: Step 1 is implemented according to the following steps: Step 1.1: Filtering and downsampling The raw EEG signal data is filtered to remove low-frequency drift and high-frequency noise; then down-sampling is performed to reduce redundant data and retain key signal information; Step 1.2: Segmentation Cut the continuous EEG signal into multiple time segments, each containing data of a fixed time length; Step 1.3: Re-reference and bad segment removal The mean value of each electrode signal was used as a reference to reduce the differences between different electrodes. The EEG data was observed through an interactive interface, and useless segments containing artifacts or irrelevant noise were manually removed to obtain the cleaned preprocessed EEG signal data.

[0008] Step 2 is: extract multidimensional features from the time domain, frequency domain and nonlinear dynamic features, extract corresponding features from the source domain and target domain data respectively, and construct a feature vector that can effectively characterize the state of the EEG signal through normalization and three-dimensional feature fusion.

[0009] In the visual state space module, the input features are processed by two parallel branches: in the main branch, the input features are first expanded by a linear layer, and then stacked by depthwise separable convolution, SiLU activation function, selective scanning module and layer normalization; in the second branch, the linear layer is used to implement channel expansion and then activated by SiLU. Finally, the inputs of the two branches are combined by element-wise multiplication, and the linear layer is used to adjust the channel to be the same as the input channel. Step 2 is implemented according to the following steps: Step 2.1: Nonlinear dynamic feature extraction For each 30-second EEG signal segment, nonlinear dynamic features used to characterize signal complexity were extracted, including approximate entropy, permutation entropy, sample entropy, and Lempel-Ziv complexity. Step 2.2: Time domain feature extraction Extract time domain features from each EEG signal segment, including standard deviation, mean, and kurtosis; Step 2.3: Normalize eigenvalues Map feature data to a specified range through linear changes to ensure scale consistency of different features, thereby avoiding interference caused by different scales; Step 2.4: 3D feature fusion The multi-dimensional features of each electrode channel at each time step are fused to construct a three-dimensional feature vector, thereby enhancing the model's ability to represent complex signals.

[0010] Step 3 is: input the fused feature vector into the pre-trained temporal convolutional neural network model. The model captures temporal dependencies through causal convolution and dilated convolution structures, uses residual connections to alleviate the gradient vanishing problem, and introduces a Dropout layer to prevent overfitting. During the training process, the source domain data and the target domain data are input into the network together for training.

[0011] Step 3 is implemented as follows: Step 3.1: Input feature vector The three-dimensional feature vector fused in step 2 is input into the neural network model, the number of channels represents the number of electrode channels of the EEG signal, and the time step represents the time series length of the signal; Step 3.2: Core feature extraction unit A residual block containing 128 filters is constructed as the core feature extraction unit, and a compact 3×1 convolution kernel is used to capture local features while ensuring computational efficiency. Step 3.3: Dilated convolution and causal convolution Each residual block contains two dilated causal convolutional layers to capture short-term and long-term temporal dependencies, respectively; Step 3.4: Skip Connection and Dropout Using skip connections to add the input directly to the convolutional layer output alleviates the gradient vanishing problem and allows the network to learn constant mappings. Each residual block also includes a Dropout layer. Step 3.5: Pre-training model Pre-training is performed through a temporal convolutional neural network. The model training should be at least 200 cycles, and the learning rate is set to 10 -4 , the batch size is 64, 10% of the training data is used for validation, and the best weights are saved based on the validation error to mitigate overfitting.

[0012] Step 4 is: in the model fine-tuning and linear detection stages, the minimum error entropy is used as the cost function, the kernel width is adjusted through the minimum error entropy, non-Gaussian noise is dynamically adapted, and the covariate shift problem between the source domain and the target domain is effectively suppressed through the entropy minimization strategy.

[0013] Step 4 is implemented according to the following steps: Step 4.1: Non-Gaussian noise suppression The kernel width is adjusted using minimum error entropy to dynamically adapt to non-Gaussian noise, and outliers in input and output are optimized by minimum error entropy to improve signal quality; Step 4.2: Covariate shift suppression Reduce the covariate shift caused by individual differences. There are significant individual differences in EEG signals between different patients. Construct a feature extraction layer shared across patients. Use the minimum error entropy as the loss function in the transfer learning fine-tuning and linear detection stages to retain the error distribution characteristics of the source domain. Step 4.3: Transfer learning and minimum error entropy collaborative optimization Knowledge transfer is achieved through the pre-training-fine-tuning paradigm. The model is pre-trained on the source domain data and robust basic feature representation is learned using minimum error entropy; fine-tuning and linear detection are performed in the target domain. Step 4.4: Model parameter update Use minimum error entropy to update feature extraction parameters and regression parameters to ensure the accuracy of the model in the source domain and target domain; Step 4.5: Logistic regression classification In the process of EEG signal classification and prediction of patients with consciousness disorders, a logistic regression model is used to establish the relationship between EEG signal features and the patient's consciousness state.

[0014] Step 5 is: Use cross-validation and transfer learning performance indicators to evaluate the model, and further optimize the convergence speed and generalization ability of the model by dynamically adjusting the learning rate to ensure that the model has good predictive performance on new data.

[0015] Step 5 is implemented according to the following steps: Step 5.1, learning rate adjustment Design learning rate scheduling strategies, including learning rate decay and dynamic adjustment methods, to balance the convergence speed and performance of the model; Step 5.2: Regularization method Introduce regularization methods to prevent overfitting and improve the generalization ability of the model; Step 5.3: Model evaluation The model performance was evaluated through cross-validation, and the model parameters and hyperparameters were adjusted through Bayesian optimization to optimize the model performance.

[0016] The beneficial effects of the present invention are: The present invention is based on a method for classifying patients with impaired consciousness based on multi-feature fusion and transfer learning. By fusing multidimensional features and using minimum error entropy instead of the traditional mean square error as the cost function, it can better adapt to the EEG signal characteristics of different patients, alleviate the data covariate offset problem caused by transfer learning and individual differences, and thus effectively improve the classification accuracy and system robustness. The present invention can be widely used in clinical auxiliary diagnosis, disease monitoring and rehabilitation evaluation of patients with impaired consciousness, and provide reliable data support for personalized treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is the overall framework diagram of the present invention; Figure 2 is a flow chart of EEG data preprocessing and feature extraction of the present invention; Figure 3 This is the architecture diagram of the temporal convolutional neural network of the present invention. DETAILED DESCRIPTION

[0018] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] Example 1 The present invention is based on a method for classifying patients with consciousness disorders based on multi-feature fusion and transfer learning. The overall framework is as follows: Figure 1 As shown, the specific implementation steps are as follows: Step 1: Data preprocessing; Step 2: Multi-dimensional feature extraction and fusion; Step 3: Feature input and model training; Step 4: Minimum error entropy optimization; Step 5: Model strategy training and optimization.

[0020] Example 2 A method for classifying patients with disorders of consciousness based on multi-feature fusion and transfer learning, wherein step 1 is to perform eye movement artifact correction, power frequency noise elimination, bandpass filtering, downsampling, re-referencing, and bad segment removal on the source domain data (the EEG signals of patients under anesthesia) and the target domain data (the EEG signals of patients with disorders of consciousness) to improve the signal quality, ensure the consistency of the source domain and target domain data in the feature space, and reduce the impact of preprocessing differences on model training.

[0021] Step 1 is implemented according to the following steps: Step 1.1: Filtering and downsampling Filter the source domain (patient EEG signals) and target domain (EEG signals of patients with impaired consciousness) data (0.5-100 Hz) to remove low-frequency drift and high-frequency noise; then perform downsampling to reduce redundant data and retain key signal information; Step 1.2: Segmentation Cut the continuous EEG signal into multiple time segments, each containing data of a fixed time length (30 seconds), which helps convert long time series data into small segments for subsequent processing and analysis; Step 1.3: Re-reference and bad segment removal The mean value of each electrode signal is used as a reference to reduce the differences between different electrodes. The EEG data is observed through an interactive interface (such as EEGLAB), and useless segments containing artifacts or irrelevant noise are manually removed to obtain the cleaned preprocessed EEG signal data.

[0022] Example 3 A classification method for patients with consciousness disorders based on multi-feature fusion and transfer learning, in which step 2 is: extracting multidimensional features from the time domain (mean, standard deviation, kurtosis), frequency domain (power spectrum density) and nonlinear dynamic features (approximate entropy, permutation entropy, sample entropy), extracting corresponding features from the source domain and target domain data respectively, and fusing them through normalization and three-dimensional features, such as Figure 2 As shown, a feature vector is constructed that can effectively characterize the state of the EEG signal and enhance the feature representation ability of the model.

[0023] Step 2 is implemented as follows: Step 2.1: Nonlinear dynamic feature extraction For each 30-s EEG signal segment, nonlinear dynamic features were extracted to characterize signal complexity, including approximate entropy (m=2, r=0.2), permutation entropy (m=3, r=1), sample entropy (m=2, r=0.2), and Lempel-Ziv complexity; Step 2.2: Time domain feature extraction Extract time domain features from each EEG signal segment, including standard deviation, mean, kurtosis, etc. These features can reflect the basic statistical characteristics of EEG signals and help capture the amplitude changes and distribution characteristics of the signals; Step 2.3: Normalize eigenvalues The feature data is mapped to [0,1] through linear changes to ensure the scale consistency of different features, thereby avoiding interference caused by different scales.

[0024] Step 2.4: 3D feature fusion The multi-dimensional features of each electrode channel at each time step are fused to construct a three-dimensional feature vector of (channel × time step × feature (8 × 30 × 7)) to enhance the model's ability to represent complex signals.

[0025] Example 4 A method for classifying patients with impaired consciousness based on multi-feature fusion and transfer learning, wherein step 3 is: inputting the fused feature vector into a pre-trained temporal convolutional neural network (TCN) model. Figure 3 As shown in the figure, the TCN model captures temporal dependencies through causal convolution and dilated convolution structures, uses residual connections to alleviate the gradient vanishing problem, and introduces a Dropout layer to prevent overfitting. During the training process, the source domain data (the EEG data of anesthetized patients) and the target domain data (the EEG data of patients with impaired consciousness) are input into the network for training. The model can gradually adjust parameters through transfer learning to improve the classification performance of the target domain data.

[0026] Step 3 is implemented as follows: Step 3.1: Input feature vector The three-dimensional feature vector fused in step 2 is input into the neural network model, the number of channels represents the number of electrode channels of the EEG signal, and the time step represents the time series length of the signal; Step 3.2: Core feature extraction unit A residual block containing 128 filters is constructed as the core feature extraction unit, and a compact 3×1 convolution kernel is used to capture local features while ensuring computational efficiency.

[0027] Step 3.3: Dilated convolution and causal convolution Each residual block contains two dilated causal convolutional layers, which are used to capture short-term (dilation rate 1) and long-term (dilation rate 2) temporal dependencies respectively; Step 3.4: Skip Connection and Dropout A skip connection is used to add the input directly to the output of the convolutional layer to alleviate the gradient vanishing problem and allow the network to learn the identity mapping. Each residual block also includes a Dropout layer (Dropout rate is 0.1), that is, 10% of the neurons are randomly discarded during training to prevent overfitting.

[0028] Step 3.5: Pre-training model Pre-training is performed through a temporal convolutional neural network. The model training should be at least 200 cycles, and the learning rate is set to 10 -4 , the batch size is 64, 10% of the training data is used for validation, and the best weights are saved based on the validation error to mitigate overfitting.

[0029] Example 5 A classification method for patients with disorders of consciousness based on multi-feature fusion and transfer learning, wherein step 4 is: in the model fine-tuning and linear detection stages, the minimum error entropy is used instead of the traditional mean square error as the cost function, the kernel width is adjusted by the minimum error entropy, non-Gaussian noise is dynamically adapted, and the covariate shift problem between the source domain and the target domain is effectively suppressed through the entropy minimization strategy, which significantly improves the robustness of the model to the distribution difference between the source domain and the target domain.

[0030] Step 4 is implemented according to the following steps: Step 4.1: Non-Gaussian noise suppression The kernel width is adjusted using minimum error entropy (MEE), which dynamically adapts to non-Gaussian noise (such as electromyographic artifacts), and outliers in input and output are optimized by minimum error entropy to improve signal quality. Step 4.2: Covariate shift suppression Reduce the covariate shift caused by individual differences. There are significant individual differences in EEG signals of different patients. Construct a feature extraction layer shared across patients. Use the minimum error entropy as the loss function in the transfer learning fine-tuning and linear detection stages to retain the error distribution characteristics of the source domain (the EEG data of patients under anesthesia). Step 4.3: Transfer learning and minimum error entropy collaborative optimization Knowledge transfer is achieved through the pre-training-fine-tuning paradigm. The model is pre-trained on the source domain data (patient anesthesia data) and robust basic feature representation is learned using minimum error entropy; fine-tuning and linear detection are performed in the target domain (disordered consciousness patient data); Step 4.4: Model parameter update Use minimum error entropy to update feature extraction parameters and regression parameters, and dynamically adjust learning rate and error correction to ensure the accuracy of the model in the source domain and target domain; Step 4.5: Logistic regression classification In the process of EEG signal classification and prediction of patients with consciousness disorders, a logistic regression model is used to establish the relationship between EEG signal features and the patient's consciousness state, especially to distinguish between "minimally conscious state" and "unresponsive wakefulness syndrome / vegetative state".

[0031] Example 6 Classification method for patients with impaired consciousness based on multi-feature fusion and transfer learning, where step 5 is: evaluate the model using 10-fold cross validation and transfer learning performance indicator classification accuracy, and adjust the learning rate to 10 -4 , optimize the convergence speed and generalization ability of the model, and ensure that the model has good predictive performance on new data.

[0032] Step 5 is implemented according to the following steps: Step 5.1, learning rate adjustment Design learning rate scheduling strategies, including learning rate decay (when the validation set loss does not decrease for 5 consecutive epochs, the learning rate is halved) and dynamic adjustment (hot restart), to balance the convergence speed and performance of the model; Step 5.2: Regularization method Introduce regularization methods (weight decay) to prevent overfitting and improve the generalization ability of the model; Step 5.3: Model evaluation The model performance was evaluated using 10-fold cross validation, and the model parameters and hyperparameters were adjusted through Bayesian optimization to optimize the model performance.

[0033] The present invention is based on a method for classifying patients with impaired consciousness based on multi-feature fusion and transfer learning. By fusing and extracting multi-dimensional features, the features are input into a pre-trained model in a temporal convolutional neural network, and the minimum error entropy is used as the cost function. The pre-trained model is fine-tuned and linearly detected, which significantly improves the robustness of the EEG signal classification model to non-Gaussian noise and covariate shift problems in transfer learning, and improves the accuracy of EEG signal classification in patients with impaired consciousness. This method can better adapt to classification tasks under different patients and disease conditions, and has strong versatility and clinical application potential.

Claims

1. A method for classifying patients with impaired consciousness based on multi-feature fusion and transfer learning, characterized in that: Follow the steps below to implement it: Step 1: perform eye movement artifact correction, power frequency noise elimination, bandpass filtering, downsampling, re-reference and bad segment removal on the target domain data, and perform the same preprocessing steps on the source domain data as the target domain data; Step 2: Multi-dimensional feature extraction and fusion; Step 3: Feature input and model training; Step 4: Minimum error entropy optimization; Step 5: Model strategy training and optimization.

2. The method for classifying patients with impaired consciousness based on multi-feature fusion and transfer learning according to claim 1, characterized in that: The step 1 is specifically implemented according to the following steps: Step 1.1: Filtering and downsampling The raw EEG signal data is filtered to remove low-frequency drift and high-frequency noise; then down-sampling is performed to reduce redundant data and retain key signal information; Step 1.2: Segmentation Cut the continuous EEG signal into multiple time segments, each containing data of a fixed time length; Step 1.3: Re-reference and bad segment removal The mean value of each electrode signal was used as a reference to reduce the differences between different electrodes. The EEG data was observed through an interactive interface, and useless segments containing artifacts or irrelevant noise were manually removed to obtain the cleaned preprocessed EEG signal data.

3. The method for classifying patients with impaired consciousness based on multi-feature fusion and transfer learning according to claim 1, characterized in that: The step 2 is: extracting multidimensional features from the time domain, frequency domain and nonlinear dynamic features, extracting corresponding features from the source domain and target domain data respectively, and constructing a feature vector that can effectively characterize the state of the EEG signal through normalization and three-dimensional feature fusion.

4. The method for classifying patients with impaired consciousness based on multi-feature fusion and transfer learning according to claim 3 is characterized in that: The step 2 is specifically implemented according to the following steps: Step 2.1: Nonlinear dynamic feature extraction For each 30-second EEG signal segment, nonlinear dynamic features used to characterize signal complexity were extracted, including approximate entropy, permutation entropy, sample entropy, and Lempel-Ziv complexity. Step 2.2: Time domain feature extraction Extract time domain features from each EEG signal segment, including standard deviation, mean, and kurtosis; Step 2.3: Normalize eigenvalues Map feature data to a specified range through linear changes to ensure scale consistency of different features, thereby avoiding interference caused by different scales; Step 2.4: 3D feature fusion The multi-dimensional features of each electrode channel at each time step are fused to construct a three-dimensional feature vector, thereby enhancing the model's ability to represent complex signals.

5. The method for classifying patients with impaired consciousness based on multi-feature fusion and transfer learning according to claim 1, characterized in that: The step 3 is: inputting the fused feature vector into the pre-trained temporal convolutional neural network model. The model captures temporal dependencies through causal convolution and dilated convolution structures, uses residual connections to alleviate the gradient vanishing problem, and introduces a Dropout layer to prevent overfitting. During the training process, the source domain data and the target domain data are input into the network together for training.

6. The method for classifying patients with impaired consciousness based on multi-feature fusion and transfer learning according to claim 5, characterized in that: The step 3 is specifically implemented according to the following steps: Step 3.1: Input feature vector The three-dimensional feature vector fused in step 2 is input into the neural network model, the number of channels represents the number of electrode channels of the EEG signal, and the time step represents the time series length of the signal; Step 3.2: Core feature extraction unit A residual block containing 128 filters is constructed as the core feature extraction unit, and a compact 3×1 convolution kernel is used to capture local features while ensuring computational efficiency. Step 3.3: Dilated convolution and causal convolution Each residual block contains two dilated causal convolutional layers to capture short-term and long-term temporal dependencies, respectively; Step 3.4: Skip Connection and Dropout Using skip connections to add the input directly to the output of the convolutional layer alleviates the gradient vanishing problem and allows the network to learn constant mappings. Each residual block also includes a Dropout layer. Step 3.5: Pre-training model The model is pre-trained through a temporal convolutional neural network. The model training is at least 200 cycles, and the learning rate is set to 10 -4 , the batch size is 64, 10% of the training data is used for validation, and the best weights are saved based on the validation error to mitigate overfitting.

7. The method for classifying patients with impaired consciousness based on multi-feature fusion and transfer learning according to claim 1, characterized in that: The step 4 is: in the model fine-tuning and linear detection stage, the minimum error entropy is used as the cost function, the kernel width is adjusted by the minimum error entropy, the non-Gaussian noise is dynamically adapted, and the covariate shift problem between the source domain and the target domain is effectively suppressed by the entropy minimization strategy.

8. The method for classifying patients with impaired consciousness based on multi-feature fusion and transfer learning according to claim 7, characterized in that: The step 4 is specifically implemented according to the following steps: Step 4.1: Non-Gaussian noise suppression The kernel width is adjusted using minimum error entropy to dynamically adapt to non-Gaussian noise, and outliers in input and output are optimized by minimum error entropy to improve signal quality; Step 4.2: Covariate shift suppression Reduce the covariate shift caused by individual differences. There are significant individual differences in EEG signals between different patients. Construct a feature extraction layer shared across patients. Use the minimum error entropy as the loss function in the transfer learning fine-tuning and linear detection stages to retain the error distribution characteristics of the source domain. Step 4.3: Transfer learning and minimum error entropy collaborative optimization Knowledge transfer is achieved through the pre-training-fine-tuning paradigm, where the model is pre-trained on the source domain data and robust basic feature representation is learned using minimum error entropy; Fine-tuning and linear probing in the target domain; Step 4.4: Model parameter update Use minimum error entropy to update feature extraction parameters and regression parameters to ensure the accuracy of the model in the source domain and target domain; Step 4.5: Logistic regression classification In the process of EEG signal classification and prediction of patients with consciousness disorders, a logistic regression model is used to establish the relationship between EEG signal features and the patient's consciousness state.

9. The method for classifying patients with impaired consciousness based on multi-feature fusion and transfer learning according to claim 1, characterized in that: The step 5 is: using cross-validation and transfer learning performance indicators to evaluate the model, and further optimizing the convergence speed and generalization ability of the model by dynamically adjusting the learning rate to ensure that the model has good prediction performance on new data.

10. The method for classifying patients with impaired consciousness based on multi-feature fusion and transfer learning according to claim 9, characterized in that: The step 5 is specifically implemented according to the following steps: Step 5.1, learning rate adjustment Design learning rate scheduling strategies, including learning rate decay and dynamic adjustment methods, to balance the convergence speed and performance of the model; Step 5.2: Regularization method Introduce regularization methods to prevent overfitting and improve the generalization ability of the model; Step 5.3: Model evaluation The model performance was evaluated through cross-validation, and the model parameters and hyperparameters were adjusted through Bayesian optimization to optimize the model performance.

Images