Lesion-specific EEG signatures in stroke: a multi-scale framework integrating oscillations, connectivity, and asymmetry for machine learning decoding.

The lesion subgroups were small, with 18 to 22 patients per category, so many comparisons were too small to detect real differences and became null after multiple-test correction. The EEG was recorded only at admission, while behavioral scores were also collected at discharge, so the study did not track how EEG changed over recovery. Sixteen patients stopped EEG early because of physical discomfort, which may affect who was represented in the final sample. Lesion categories were coarse, combining fronto-temporal and centrum semiovale lesions, and ischemic and hemorrhagic strokes were analysed together. Scalp EEG is an indirect measure of brain sources, and the authors state that sensor-level patterns should not be read as precise neuroanatomical localisation. The machine-learning results used leave-one-subject-out validation within the same cohort, so they show internal separability, not readiness for clinical use. No EEG feature remained associated with motor, cognitive, or daily living scores after multiple-test correction.

Funding came from the National Natural Science Foundation of China, the Gusu District Health Talent Training Project, the Shenzhen Science and Technology Program, the Guangdong Basic and Applied Basic Research Foundation, and the Shenzhen Medical Research Foundation. The supplied text does not state author conflicts of interest.

Do not read the machine-learning accuracy as evidence that EEG can diagnose lesion location in practice. The model was validated only within this small exploratory cohort, and several lesion-specific EEG differences did not survive correction.