Electroencephalogram classification and prognosis prediction method based on sparse parameter sharing multi-task learning

Abstract

The application discloses a multi-task learning electroencephalogram classification and prognosis prediction method based on sparse parameter sharing, and is specifically implemented according to the following steps: step 1, data acquisition; step 2, electroencephalogram data preprocessing; step 3, multi-dimensional brain activity feature extraction and fusion; and step 4, multi-task learning model training.The application significantly improves the prediction accuracy by adopting a model with high coding capacity and combining prior information to constrain the model state space.Even in an over-parameterized neural network, a sub-network can be found through random initialization and training, and the sub-network can achieve the same prediction performance as the original network under the same number of training times.

Description

Technical Field

[0001] This invention belongs to the field of medical artificial intelligence technology, specifically involving a multi-task learning electroencephalogram classification and prognosis prediction method based on sparse parameter sharing. Background Technology

[0002] With the rapid development of emergency medical techniques and critical care medicine, the success rate of treating patients with severe brain injuries has significantly improved, leading to a continuous increase in the number of long-term surviving patients with impaired consciousness. Impaired consciousness refers to an abnormal state of consciousness caused by brain injury or disease; common types include unresponsive awakening syndrome (UWS) and minimally conscious state (MCS). Although various treatments have been attempted to treat patients with impaired consciousness, none have been able to consistently and effectively promote recovery. Some studies have reported the benefits of drug treatments (such as amantadine and zolpidem) for patients with impaired consciousness, but conflicting results exist in different studies, the fundamental mechanisms by which these drugs affect consciousness improvement remain unclear, and accurate and reliable prognostic prediction methods are lacking.

[0003] Because DOC patients present with varying clinical conditions and often have coexisting brain damage of different degrees, drugs such as amantadine and zolpidem are frequently used clinically for trial treatment. However, due to the specific nature of the patient population, the sample size for single-drug therapy is small, which makes it difficult to establish accurate efficacy / prognostic prediction models for specific drugs. UWS and MCS patients have different clinical characteristics, but the differences between them are extremely subtle, making differential diagnosis using conventional clinical methods very difficult. MCS patients are often misdiagnosed as UWS. The prognoses of the two also differ, with MCS having a relatively better prognosis. Previous studies have found that not only does the patient's state of consciousness have predictive value for prognosis, but the EEG responsiveness to drug treatment also has the ability to differentiate patient prognosis. The combination of these two factors yields even better predictive results.

[0004] The main problems with existing technologies include: high misdiagnosis rate: current behavioral scales and imaging assessment methods are difficult to accurately assess patients with progressively recovering consciousness, and their accuracy is limited. Insufficient sample data: for the efficacy assessment of specific drugs, the existing sample size is small, and traditional machine learning methods are prone to overfitting, resulting in insufficient generalization ability of the predictive models.

[0005] Therefore, there is an urgent need to develop a more accurate, sensitive, and generalizable assessment method to solve the above-mentioned technical problems, especially in the areas of classification of patients with disorders of consciousness, evaluation of drug treatment effects, and prognosis prediction. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-task learning EEG classification and prognostic prediction method based on sparse parameter sharing, which improves the classification accuracy of patients with consciousness disorders.

[0007] The technical solution adopted in this invention is a multi-task learning EEG classification and prognostic prediction method based on sparse parameter sharing, which is implemented according to the following steps:

[0008] Step 1: Data Collection;

[0009] Step 2: EEG data preprocessing;

[0010] Step 3: Extraction and fusion of multi-dimensional brain activity features;

[0011] Step 4: Training the multi-task learning model.

[0012] The technical solution of this invention is also characterized by:

[0013] In step 1, we collect the datasets used to train the base network and the datasets used to train the task sub-networks.

[0014] The dataset used to train the base network includes historical EEG data from conscious or unconscious patients, with prognostic records from a one-year follow-up and data from a healthy control group; the dataset used to train the task sub-network includes treatment response data for specific drugs and prognostic outcomes such as whether the patient was conscious six months later.

[0015] Step 2 is implemented in the following steps:

[0016] Step 2.1, Filtering and Downsampling

[0017] The EEG data is filtered to remove low-frequency drift and high-frequency noise, and then downsampled to reduce data redundancy.

[0018] Step 2.2: Artifact Removal Processing

[0019] Eye movement artifacts were removed using independent component analysis.

[0020] Step 2.3, Re-reference and Segmentation

[0021] Electrode rereference is performed based on the acquisition data, and the continuous EEG signal is divided into multiple time periods, each containing data of a fixed duration, to facilitate subsequent analysis and processing.

[0022] Step 2.4, Data Cleaning

[0023] Cleaned EEG signal data is obtained by manually or automatically removing useless fragments containing artifacts or irrelevant noise.

[0024] Step 3 is implemented in the following steps:

[0025] Step 3.1: Power spectral density extraction

[0026] Power spectral density is extracted from EEG signals to reflect the activity state of the brain at different frequency bands;

[0027] Step 3.2: Time-frequency domain feature generation

[0028] By combining Hilbert Huang transform and zero-crossing point estimation, the fractional Fourier transform order is dynamically and adaptively adjusted through signal decomposition, frequency estimation, and dynamic adjustment of the FRFT order, effectively capturing the time-frequency characteristics of non-stationary signals.

[0029] Step 3.3: Connectivity Feature Extraction

[0030] Functional connectivity features were extracted, including mutual information between leads, phase synchronization, graph theory indicators, brain region aggregation and separation coefficients, to reflect the activity correlation between different brain regions.

[0031] Step 3.4: Information Theory Feature Extraction

[0032] Calculating approximate entropy, permutation entropy, and information theory features of Korotkoff complexity can help to further identify the nonlinear dynamic features of the relevant circuits of human brain consciousness;

[0033] Step 3.5: Feature Normalization and Fusion

[0034] By normalizing, all features are mapped to a uniform scale to avoid interference from features of different scales to model training. Then, the extracted features are fused in three dimensions to generate a comprehensive feature vector, which enhances the model’s ability to represent complex brain states.

[0035] Step 3.6: After obtaining the EEG markers, the feature expression layer generates the feature vector of the training model. This vector is composed of the difference between a single EEG marker and two EEG markers.

[0036] The signal decomposition and frequency estimation in step 3.2 are as follows: the signal is decomposed into multiple intrinsic mode functions using empirical mode decomposition to handle the time-varying frequency characteristics in the EEG; zero-crossing points are calculated for each IMF, and the instantaneous frequency is estimated by the interval between the zero-crossing points.

[0037] The specific steps for dynamically adjusting the FRFT order in step 3.2 are as follows: by calculating the rate of change of the zero-crossing interval, the dynamic changes in the signal frequency are captured; the FRFT order is dynamically adjusted according to the rate of change of the signal frequency to ensure that the frequency changes of the signal in different time periods are accurately captured.

[0038] Step 4 is implemented in the following steps:

[0039] Step 4.1, Basic Network Design

[0040] A multi-layer fully connected neural network is used as the base network. The input is the fusion feature extracted in step 3, and the output is the results of tasks such as consciousness state classification, drug efficacy prediction and prognosis analysis.

[0041] Step 4.2, Sparse Parameter Sharing Mechanism

[0042] Subnetworks are generated using an iterative magnitude pruning strategy, with shared weights between tasks not exceeding 30%, and the remaining weights being task-independent. By dynamically optimizing the shared layers, network complexity is reduced and overfitting is avoided.

[0043] Step 4.3, Training Strategy

[0044] The weighted cross-entropy loss function is used for training, which is particularly suitable for optimization in the context of small sample data, ensuring that the model can fit well for different tasks.

[0045] The beneficial effects of this invention are:

[0046] This invention presents a multi-task learning-based EEG classification and prognostic prediction method based on sparse parameter sharing. By employing a model with high encoding capacity and constraining the model's state space with prior information, it significantly improves prediction accuracy. Even in overparameterized neural networks, through random initialization and training, a subnetwork can be found that achieves the same predictive performance as the original network with the same number of training iterations. This demonstrates that despite using a large number of parameters, only a small percentage (less than 10%) ultimately play a crucial role in the prediction result. Through effective feature encoding, a low-complexity network can achieve the same performance as a complex network, reducing data requirements. Attached Figure Description

[0047] Figure 1 This is the overall framework diagram of the present invention. Detailed Implementation

[0048] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0049] Example 1

[0050] This invention relates to a multi-task learning EEG classification and prognostic prediction method based on sparse parameter sharing, used for the classification, drug efficacy prediction, and prognostic analysis of patients with disorders of consciousness (DOC). The method is implemented according to the following steps:

[0051] Step 1: Data Collection;

[0052] Step 2: EEG data preprocessing;

[0053] Step 3: Extraction and fusion of multi-dimensional brain activity features;

[0054] Step 4: Training the multi-task learning model.

[0055] like Figure 1 The diagram shown illustrates the overall framework of this invention, demonstrating the weight-sharing mechanism between the base network and subtasks. Dashed boxes represent shared weights, while solid boxes represent task-independent weights. In the design of the base network, relevant features are combined using prior knowledge or nodes are automatically connected according to preset criteria to reduce model complexity and enhance feature representation capabilities.

[0056] Table 1 below lists three types of EEG markers that are effective in distinguishing states of consciousness: information theory markers, spectral markers, and connectivity markers. The effectiveness of these markers in indexing consciousness has been verified in previous studies.

[0057] Table 1

[0058]

[0059] Despite limited training data, this invention optimizes the model and improves prediction accuracy by utilizing long-term resting-state clinical data and EEG data after drug treatment. Patients with good prognoses share common EEG characteristics, but different drugs modulate consciousness-related neural circuits differently; therefore, prognostic prediction is designed as a loosely coupled task sharing a portion of the network. The sparse shared structure allows for iterative optimization of network parameters in a consistent direction, reducing the required model parameters and avoiding performance degradation in task-independent scenarios. Patients with poor prognoses, however, are affected by multiple factors, resulting in complex and variable EEG signal characteristics. Through multi-feature fusion, this invention provides common and targeted feature encoding schemes for prognostic prediction tasks across different datasets, optimizing model performance and improving prediction accuracy.

[0060] To address the issue of small sample data, this invention performs joint learning on multiple resting-state EEG datasets and constrains model complexity through prior knowledge and information from other datasets. Since the correlation between different tasks is unclear, a sparse sharing mechanism that automatically adapts to varying correlations is employed to ensure shared subnets across tasks and parallel training. By introducing prior information, model complexity is reduced, and the connection methods between hidden layer nodes and upper-layer nodes can be enhanced by combining relevant features based on prior knowledge or through automatic connections, thereby improving feature representation and simplifying the training process.

[0061] Example 2

[0062] A multi-task learning method for EEG classification and prognosis prediction based on sparse parameter sharing is proposed. In step 1, two types of data are used: the dataset used to train the base network and the dataset used to train the task sub-network.

[0063] The dataset used to train the basic network includes: historical EEG data from conscious / unconscious patients. This dataset contains clinical data from thousands of DOC patients accumulated in the hospital over many years, with continuous EEG monitoring for several hours and prognostic records after one year of follow-up; publicly available multicenter EEG data from DOC patients; and data from healthy control groups, including historical interictal EEG data from epilepsy patients with normal consciousness and good background brain activity in neurology departments.

[0064] The dataset used for training the sub-network included: After enrollment, participants underwent weekly behavioral assessments and initial EEG data collection based on the Coma Recovery Scale-Revised (CRS-R). Treatment response data for two specific drugs (10mg amantadine administered via nasogastric tube and zolpidem) were also collected, specifically baseline resting-state EEG data one hour before medication and EEG data five to six hours after medication. After data collection, the patients' sensory recovery was assessed again. In clinical practice, if significant recovery of consciousness was observed after the first dose, medication was continued. Each patient underwent two assessments using the Modified Glasgow Outcome Scale (GOS) by an experienced clinician. Scores 1 and 2 were defined as no sensory recovery, while scores 3, 4, 5, and the MCS were defined as sensory recovery. Patients who died after recovery due to other causes were also classified as having sensory recovery. Finally, based on the GOS and CRS-R behavioral assessments, patients were followed up for six months to determine whether the predicted outcome for those with significant improvement was "awake".

[0065] Example 3

[0066] A multi-task learning-based EEG classification and prognostic prediction method based on sparse parameter sharing, wherein step 2 is implemented according to the following steps:

[0067] Step 2.1, Filtering and Downsampling: Filter the EEG data to remove low-frequency drift (e.g., 0.5-100 Hz) and high-frequency noise, and then perform downsampling to reduce data redundancy;

[0068] Step 2.2, Artifact Removal: Independent Component Analysis (ICA) was used to remove eye-tracking artifacts;

[0069] Step 2.3, Rereference and Segmentation: Based on the acquisition data, perform electrode rereference (such as average reference, unipolar reference, bipolar reference) to divide the continuous EEG signal into multiple time segments, each containing data of a fixed duration (such as 30 seconds) to facilitate subsequent analysis and processing.

[0070] Step 2.4, Data Cleaning: Useless segments containing artifacts or irrelevant noise are removed using manual or automated tools to obtain cleaned EEG signal data.

[0071] Example 4

[0072] A multi-task learning-based EEG classification and prognostic prediction method based on sparse parameter sharing, wherein step 3 is implemented according to the following steps:

[0073] Step 3.1, Power spectral density extraction: Extract the power spectral density (such as Alpha, Beta, Delta, Gamma, and Theta band power) from the EEG signal to reflect the activity state of the brain in different frequency bands;

[0074] Step 3.2, Time-Frequency Domain Feature Generation: Combining Hilbert Huang Transform (HHT) and zero-crossing point estimation, the fractional Fourier Transform (FRFT) order is dynamically and adaptively adjusted through signal decomposition and frequency estimation and dynamic adjustment of the FRFT order, thereby effectively capturing the time-frequency characteristics of non-stationary signals;

[0075] Step 3.3, Connectivity Feature Extraction: Extract functional connectivity features, including mutual information between leads, phase synchronization, graph theory indicators, brain region aggregation and separation coefficients, etc., to reflect the activity correlation between different brain regions and reflect the functional connectivity of the brain network;

[0076] Step 3.4, Information Theory Feature Extraction: Calculate information theory features such as approximate entropy, permutation entropy, and Korotkoff complexity. These features can measure the complexity and orderliness of brain networks and help to further identify the nonlinear dynamic characteristics of human brain consciousness-related circuits.

[0077] Step 3.5, Feature Normalization and Fusion: Through normalization (such as standardization or max-min normalization), all features are mapped to a uniform scale to avoid interference from features of different scales during model training. Then, the extracted features are fused using three-dimensional feature fusion to generate a comprehensive feature vector, enhancing the model's ability to represent complex brain states.

[0078] Step 3.6: After obtaining the EEG markers, the feature expression layer generates the feature vector of the training model. This vector is composed of the difference between a single EEG marker and two EEG markers.

[0079] Example 5

[0080] A multi-task learning-based EEG classification and prognostic prediction method based on sparse parameter sharing, wherein step 3.2 involves signal decomposition and frequency estimation:

[0081] Signal decomposition: Empirical mode decomposition (EMD) is used to decompose the signal into multiple intrinsic mode functions (IMFs) to handle the time-varying frequency characteristics in EEG;

[0082] Zero-crossing point estimation: Zero-crossing points are calculated for each IMF, and the instantaneous frequency is estimated by the interval of the zero-crossing points, simplifying the traditional frequency estimation method.

[0083] Dynamically adjust the FRFT order

[0084] Frequency change rate estimation: Capture the dynamic changes in signal frequency by calculating the rate of change of the zero-crossing interval;

[0085] Adaptive adjustment of FRFT order: The FRFT order is dynamically adjusted according to the frequency change rate of the signal to ensure that the frequency changes of the signal in different time periods are accurately captured, thus better adapting to the non-stationary characteristics of EEG signals;

[0086] The formula for adjusting the order is:

[0087] in, This is the default order; 0.1 is commonly chosen for EEG data. It is an adjustment factor. To estimate the rate of change of the zero-crossing interval, a higher rate of frequency change (faster change of the zero-crossing interval) corresponds to a higher FRFT order; conversely, a slower rate of frequency change uses a lower order.

[0088] Example 6

[0089] A multi-task learning-based EEG classification and prognostic prediction method based on sparse parameter sharing, wherein step 4 is implemented according to the following steps:

[0090] Step 4.1, Basic Network Design: A multi-layer fully connected neural network is used as the basic network. The input is the fusion feature extracted in Step 3, and the output is the results of tasks such as consciousness state classification, drug efficacy prediction and prognosis analysis.

[0091] Step 4.2, Sparse Parameter Sharing Mechanism: Subnetworks are generated using an Iterative Amplitude Pruning (IMP) strategy. The shared weights between tasks do not exceed 30%, with the remaining weights being task-independent. By dynamically optimizing the shared layers, network complexity is reduced, and overfitting is avoided.

[0092] Step 4.3 Training Strategy: The weighted cross-entropy loss function is used for training, which is particularly suitable for optimization in the context of small sample data, ensuring that the model can fit well for different tasks (classification of consciousness state, prediction of drug response, etc.).

[0093] Multi-task neural networks use feature representations as input to predict patient prognosis. A multi-task sparse-shared neural network consists of an input layer, two sparse-shared layers, and a decision layer. Different tasks share the same input layer and have independent decision layers. In the shared sparse layers, the weights for each task come from the same network, called the base network. Each task's subnet is composed of a subset of the weights from the base network.

[0094] Optimization of sparse shared networks involves two main steps, performed iteratively, until the model converges on the test set:

[0095] Generate subnetworks: Update the mask and generate subnetworks for each task;

[0096] The subnetwork is generated using the Iterative Amplitude Clipping (IMP) method, as follows:

[0097] (1) Initialize model parameters;

[0098] (2) Initialize the mask for each task t;

[0099] (3) Select a subset from the data of task t and train the model multiple times;

[0100] (4) Sort the parameters according to their amplitude, select a certain proportion of the parameters and set their mask values ​​to 0;

[0101] (5) If the parameter density is too high, reset and return to the training step; otherwise, the subnetwork generation ends.

[0102] Multi-task sub-network joint training: Optimize the basic network parameters and train multiple sub-networks.

[0103] The task sub-networks are trained alternately using stochastic gradient descent until convergence. The process is as follows:

[0104] (1) Select task t in sequence for training.

[0105] (2) Select a random mini-batch from the data of task t for training.

[0106] (3) Input the mini-batch into the sub-network and use gradient descent to perform one round of training until all sub-networks converge.

[0107] Through these steps, model parameters are shared across tasks and continuously updated during training, ultimately enabling the network to achieve consistency across multiple tasks. Differences between tasks are represented by non-shared parameters.

[0108] This invention presents a multi-task learning method for EEG classification and prognostic prediction based on sparse parameter sharing, employing a sparse parameter sharing multi-task network model incorporating prior information. In this architecture, parameters are partially shared across tasks, allowing for automatic extraction of subnetworks for each task and automatic selection of shared weight sets through pruning strategies. Introducing prior information into the network structure design reduces model complexity. Furthermore, this invention combines HHT transform and zero-crossing point estimation to extract features, estimating instantaneous frequencies through zero-crossing points and adjusting the order of the FRFT based on frequency changes. This method simplifies instantaneous frequency calculation while preserving the rich time-frequency characteristics of EEG signals, making it particularly suitable for the analysis of non-stationary signals. This invention is applicable to precision medicine for patients with altered consciousness and has high clinical application value.

Claims

1. A multi-task learning-based EEG classification and prognostic prediction method based on sparse parameter sharing, characterized in that, The specific steps are as follows: Step 1: Data Collection; Step 2: EEG data preprocessing; Step 3: Extraction and fusion of multi-dimensional brain activity features; Step 4: Training the multi-task learning model; Step 4 is implemented in the following steps: Step 4.1, Basic Network Design A multi-layer fully connected neural network is used as the base network. The input is the fusion feature extracted in step 3, and the output is the result of the consciousness state classification, drug efficacy prediction and prognosis analysis task. Step 4.2, Sparse Parameter Sharing Mechanism Subnetworks are generated using an iterative magnitude pruning strategy, with shared weights between tasks not exceeding 30%, and the remaining weights being task-independent. By dynamically optimizing the shared layers, network complexity is reduced and overfitting is avoided. Step 4.3, Training Strategy The weighted cross-entropy loss function is used for training, which is suitable for optimization in the context of small sample data and ensures that the model can fit well for different tasks. Multi-task neural networks use feature representations as input to predict patient prognosis. Multi-task sparse shared neural networks consist of an input layer, two sparse shared layers, and a decision layer. Different tasks have the same input layer and independent decision layers. In the shared sparse layer, the weights of each task come from the same network, called the base network. Each task is composed of a subset of the weights of the base network to form a subnet.

2. The multi-task learning EEG classification and prognostic prediction method based on sparse parameter sharing according to claim 1, characterized in that, In step 1, the datasets used to train the base network and the datasets used to train the task sub-networks are collected.

3. The multi-task learning EEG classification and prognostic prediction method based on sparse parameter sharing according to claim 2, characterized in that, The dataset used for training the base network includes historical EEG data from conscious or unconscious patients, prognostic records with a 1-year follow-up, and data from a healthy control group; the dataset used for training the task sub-network includes treatment response data for specific drugs, and prognostic outcomes such as whether the patient is conscious after 6 months.

4. The multi-task learning EEG classification and prognostic prediction method based on sparse parameter sharing according to claim 1, characterized in that, Step 2 is implemented in the following steps: Step 2.1, Filtering and Downsampling The EEG data is filtered to remove low-frequency drift and high-frequency noise, and then downsampled to reduce data redundancy. Step 2.2: Artifact Removal Processing Eye movement artifacts were removed using independent component analysis. Step 2.3, Re-reference and Segmentation Electrode rereference is performed based on the acquisition data, and the continuous EEG signal is divided into multiple time periods, each containing data of a fixed duration, to facilitate subsequent analysis and processing. Step 2.4, Data Cleaning Cleaned EEG signal data is obtained by manually or automatically removing useless fragments containing artifacts or irrelevant noise.

5. The multi-task learning EEG classification and prognostic prediction method based on sparse parameter sharing according to claim 1, characterized in that, Step 3 is implemented in the following steps: Step 3.1: Power spectral density extraction Power spectral density is extracted from EEG signals to reflect the activity state of the brain at different frequency bands; Step 3.2: Time-frequency domain feature generation By combining Hilbert Huang transform and zero-crossing point estimation, the fractional Fourier transform order is dynamically and adaptively adjusted through signal decomposition, frequency estimation, and dynamic adjustment of the FRFT order, effectively capturing the time-frequency characteristics of non-stationary signals. Step 3.3: Connectivity Feature Extraction Extract functional connectivity features, including mutual information between leads, phase synchronization, graph theory indicators, brain region aggregation and separation coefficients, to reflect the activity correlation between different brain regions and reflect the functional connectivity of the brain network; Step 3.4: Information Theory Feature Extraction Calculating approximate entropy, permutation entropy, and information theory features of Korotkoff complexity can help to further identify the nonlinear dynamic features of the relevant circuits of human brain consciousness; Step 3.5: Feature Normalization and Fusion By normalizing, all features are mapped to a uniform scale to avoid interference from features of different scales to model training. Then, the extracted features are fused in three dimensions to generate a comprehensive feature vector, which enhances the model’s ability to represent complex brain states. Step 3.6: After obtaining the EEG markers, the feature expression layer generates the feature vector of the training model. This vector is composed of the difference between a single EEG marker and two EEG markers.

6. The multi-task learning EEG classification and prognostic prediction method based on sparse parameter sharing according to claim 5, characterized in that, The signal decomposition and frequency estimation in step 3.2 are as follows: the signal is decomposed into multiple intrinsic mode functions using empirical mode decomposition to handle the time-varying frequency characteristics in the EEG; zero-crossing points are calculated for each IMF, and the instantaneous frequency is estimated by the interval between the zero-crossing points.

7. The multi-task learning EEG classification and prognostic prediction method based on sparse parameter sharing according to claim 5, characterized in that, The specific steps in step 3.2 for dynamically adjusting the FRFT order are as follows: by calculating the rate of change of the zero-crossing interval, the dynamic changes in the signal frequency are captured; the FRFT order is dynamically adjusted according to the rate of change of the signal frequency to ensure that the frequency changes of the signal in different time periods are accurately captured.

Images