Multimodal Behavioral Data Predict Stroke Patient's Response to BCI Treatment Through Explainable AI.

Brain-computer interface (BCI)-based neurorehabilitation holds promise in enhancing motor recovery after stroke. However, recent research has reported heterogeneous results, indicating both responders and non-responders to BCI therapy. Using explainable artificial intelligence (XAI) methods, this study aims to investigate the independent and combined importance of multimodal behavioral data to predict patients' response to BCI therapy. Forty-two subacute stroke patients with lower-limb motor impairment underwent behavioral assessments, and received two-week BCI rehabilitation training. Linear regression, elastic net and artificial neural network models were developed to predict response to BCI therapy. Two XAI techniques, the stepwise method and Shapley additive explanation, were used to interpret model outcomes. The multivariate model (R ${}^{\mathbf {{2}}} =0.852$ , P<0.001) that combines an optimal subset of multimodal behavioral data outperformed the univariate model (R ${}^{\mathbf {{2}}} =0.758$ , P<0.001) trained on a single variable. Elastic net and artificial neural network models both demonstrated high prediction performance, as indicated by classification accuracies of 0.810 and 0.762, and areas under the receiver operating characteristic curve of 0.782 and 0.771. Our results revealed that multimodal behavioral data, including demographic, clinical, and biomechanical characteristics, provided unique and complementary information for interpreting the response of subacute patients to BCI therapy. Particularly, baseline motor impairment, muscle spasticity and balance function were primary predictors. Our findings highlight the core role of XAI methods towards precision medicine, which can help clinicians to identify individual recovery potentials and plan optimal treatment strategies.