Using explainable AI to identify disease-relevant and deep brain stimulation treatment-sensitive gait features in Parkinson's disease.

The analysis used a small existing dataset, so the model findings need testing in larger independent groups before they can guide practice. The study only validated internally with a held-out test set, not with an external cohort. Post-DBS gait was measured while participants were on medication and receiving stimulation, so the pre-DBS to post-DBS differences cannot be assigned to DBS alone. Age matching did not balance sex or other demographics, so some differences between Parkinson's disease and controls may reflect residual confounding. Highly correlated gait features were grouped before importance ranking, so the contribution of any single feature inside a group cannot be separated from the others. The univariate tests were exploratory because the 35 gait measures were correlated and the sample was small.

Do not read the post-DBS gait differences as proof that DBS alone improves gait. The post-DBS recordings were made with medication and stimulation together, and the study identifies candidate markers rather than tested treatment effects.