Multi-task Heterogeneous Ensemble Learning-based Cross-subject EEG Classification in Stroke Patients

Medical Laboratory Science Information Science Artificial Intelligence Neurology Medicine
stroke electroencephalography motor imagery cross-subject training multi-task heterogeneous ensemble learning

Research Process Schematic Diagram

Background Introduction

Motor Imagery (MI) refers to performing activities through imagination without actual muscle movement. This paradigm is widely used in Brain-Computer Interface (BCI) to decode brain activities into control commands for external devices. Specifically, Electroencephalography (EEG) is widely used in BCI due to its relative affordability, mobility, and higher temporal resolution compared to other neuroimaging tools. Additionally, this paradigm can help stroke patients with neural rehabilitation. Research has shown that robot-assisted BCI training can improve motor rehabilitation outcomes for stroke patients (see papers [5] and [6]). This is because the neural pathways activated during MI are similar to those of actual Motor Execution (ME), hence, imagination can also activate neural pathways in sensory-motor regions, aiding post-stroke motor rehabilitation.

Research Origin

This paper was written by Minji Lee et al., with authors from Catholic University of Korea, Kyungpook National University, New York University Abu Dhabi, Hallym University, and Mingje Chun Creek Rehabilitation Hospital. It was published in the IEEE Transactions on Neural Systems and Rehabilitation Engineering, Vol. 32, 2024. The research was supported by Chungbuk National University BrainKorea21 Program (2023) and the Korean government Ministry of Science and ICT (MSIT).

Research Details

a) Research Process

The main research process includes data collection, experimental protocol, data preprocessing, Multi-task Heterogeneous Ensemble Learning (MEEG-HEL) framework, and performance evaluation. The specific steps are as follows:

  1. Participants

    • A total of nine chronic ischemic stroke patients participated in this study. The average age of the participants was 55 years (±5.36), including 3 women. All patients underwent cognitive function assessment (MMSE) and upper limb motor function assessment (FMA_UL) to determine their suitability for the experiment.

  2. Experimental Protocol

    • Subjects performed a sequential finger tapping task with both hands, divided into ME and MI paradigms. The tasks were designed in blocks, with each task containing 20 blocks. In each block, when a dot (cue) appeared on the screen, patients tapped sequentially from the ring finger to the little finger.

  3. Data Collection and Preprocessing

    • EEG data were measured using the NeuroPraX EEG system (NeuroConn GmbH, Germany) at a sampling rate of 4000 Hz. Signals were collected using Ag/AgCl surface electrodes and preprocessed using the EEGLAB toolbox. Finally, 27 channels were selected, and the EEG data were segmented.

  4. MEEG-HEL Framework Design

    • Common Spatial Pattern (CSP) was used for feature extraction, Sequential Forward Floating Selection (SFFS) was used for feature selection, and a heterogeneous ensemble learner was constructed using seven benchmark models, including Shallow Neural Network (SNN), Kernel Support Vector Machine (KSVM), and Subspace Discriminant Classifier (SDC), etc. Classification was ultimately carried out using a weighted method.

b) Main Results

  1. Direction Recognition Task (DR)

    • In the ME and MI paradigms, classification performance under the ME paradigm was 0.7061 (±0.1270), while the MI paradigm’s classification performance was 0.7419 (±0.0811). The Confusion Matrix results showed that the MI paradigm had higher accuracy, especially in distinguishing between the right hand (RT) and the left hand (LT).

  2. Motor Assessment Task (MA)

    • The classification performances in the ME and MI paradigms were 0.6791 (±0.1253) and 0.7457 (±0.1317), respectively. In the MI paradigm, the TP and FN values were higher, indicating the model’s higher accuracy in predicting the healthy hand.

  3. CSP Results

    • In classification tasks, CSP pattern experiences show that activity regions under the MI paradigm mainly concentrate in the sensorimotor cortex, whereas, under the ME paradigm, right-hand activities mainly activated the left sensorimotor cortex.

c) Research Conclusion

This study shows that the MEEG-HEL framework performs excellently in multi-task classification for stroke patients, significantly outperforming other benchmark models. By classifying direction recognition and motor assessment tasks, it can effectively aid rehabilitation training for stroke patients. This framework not only improves classification accuracy but also enhances the model’s applicability to new patients, facilitating the promotion of neural rehabilitation technology in clinical settings.

d) Research Highlights

  1. The multi-task learning method combines direction recognition and motor assessment tasks, improving the model’s classification performance.
  2. A heterogeneous ensemble learning method is proposed, achieving good classification results even with small samples.
  3. CSP feature extraction and SFFS feature selection greatly enhance the generalization ability and accuracy of the model.

Summary

This study employs a multi-task heterogeneous ensemble learning method for EEG classification in stroke patients, effectively improving classification performance and demonstrating the potential of this framework in neural rehabilitation. By identifying direction and motor assessment tasks, it can provide more accurate rehabilitation training programs for stroke patients and improve their quality of life. The innovation of this study lies in considering the characteristics and actual needs of stroke patients, promising greater breakthroughs in future clinical applications.