Integrating task-state electroencephalography and motor function for intelligent assessment of recovery status from prolonged disorders of consciousness

Accurately assessing recovery in patients with prolonged disorders of consciousness (PDoC) is essential for evaluating the therapeutic efficacy and optimizing clinical care. However, substantial individual variability and poorly understood recovery mechanisms render this task exceptionally challenging. We collected task-state electroencephalography and clinical records from 78 PDoC patients to develop an intelligent assessment framework. Power spectral densities (PSDs) and coherence values (CVs) derived from event-related potentials (ERPs) were extracted and integrated with motor function scores to construct a multidimensional feature set. Linear support vector machine (SVM) and logistic regression (LR) models were built and rigorously evaluated via leave-one-out cross-validation (LOOCV). SHapley Additive exPlanations (SHAP) was utilized for visual explanation. The LOOCV indicated that both LR and SVM achieved optimal, similar performance in integrating alpha-band features with motor function (Macro-Average F1-score [Ma-F]: 76.79% versus 75.54%). The Receiver Operating Characteristic curves (with Area Under the Curve values of 77.57% for LR and 72.82% for SVM) and Precision-Recall curves (with Average Precision values of 68.87% for LR and 59.85% for SVM) suggested that the framework performs stably. The reliability curves indicated that both models demonstrate good calibration and are reliable for assessment. The study provides preliminary evidence that alpha-band PSDs and CVs derived from ERPs potentially enhance the motor function assessment performance. This integrated paradigm provides a valuable indicator. Overall, our findings suggest that combining neuroscience with intelligent technology holds promise for improving personalized diagnosis and optimizing the PDoC management.