Explainable Deep-Learning Prediction for Lower Extremity Motor Gains with Brain-Machine Interfaces After Stroke

Background and/or Objectives: Nowadays, many new rehabilitation training methods have developed. The closed-loop electroencephalogram (EEG)-based brain-computer interface (BCI) rehabilitation training has shown the potential to enhance a patient’s motor ability, but not every patient can receive the same gain in BCI training. This effect could be the difference between each patient’s original brain lesion region or size.

Design: All participants would ask to measure their EEG signal in two different states eye-close (EC) and eye-open (EO) resting states before rehabilitation training. The value based on the difference between FMA-LE before and after training to determine the efficiency that the gain from BMI rehabilitation is called the recovery rate. The thirty percentage of recovery rate is the threshold to separate patients into two groups: the good recovery group and the poor recovery group. With those two states' brain signals, we develop a multi-state fusion model to increase the accuracy for predicting the patients’ motor ability outcomes after the rehabilitation training.

Setting: 22 patients (M/F: 57%) with lower extremity impairments (median FMA: 17.02, IQR: 9-21)

Participants: Twenty-two stroke patients

Interventions: A 2-week Brain-Computer Interface rehabilitation training

Main Outcome Measures: The multi-state fusion network successfully classified good versus poor recovery with 82% accuracy.

Results: The majority contribution of the model in power spectrum density signal for both good and poor recovery groups, the highest-weight features are mostly presented in the delta and alpha bands. In the good recovery group, the low-frequency oscillation in frontal, central, and occipital contributes a high value of weight. In the poor recovery group, the Alpha band oscillation in the central parietal is a major contributor. The functional connectivity shows a similar result.

Conclusions: Our results demonstrated that multi-state EEG signals can increase the accuracy in predicting the motor recovery outcome after BCI training, and model explanation can reveal the underlying mechanisms of motor recovery in brain activity.