Machine Learning Decoding of Dystonic States from EEG-LFP Signals in GPi-DBS: A Multimodal Case Study

Background: Programming Deep Brain Stimulation (DBS) for pediatric dystonia remains largely empirical, as no validated electrophysiological biomarker currently guides parameter selection, and most studies treat each signal modality separately rather than exploiting their joint structure. Methods: In a single pediatric patient with generalized dystonia and bilateral globus pallidus internus (GPi) DBS implantation, we synchronized electroencephalography (EEG), local field potential (LFP), and electromyography (EMG) recordings across seven sessions spanning stimulator-OFF and multiple stimulation intensities. Dystonic muscle activity labels were derived from a validated quantitative EMG pipeline. Regularized logistic regression classifiers, trained on EEG-only, LFP-only, and combined EEG-LFP feature sets, were evaluated under four validation schemes: within-OFF cross-validation, OFF-to-ON generalization, a mixed leave-one-recording-out (LORO) scheme, and generalization to unseen stimulation intensities. Results: The Combined model performed best in the mixed LORO scenario (balanced accuracy 0.74, sensitivity 0.78, specificity 0.70), inheriting a favorable trade-off from the complementary biases of EEG (sensitivity-dominated) and LFP (specificity-dominated). LFP-derived Hjorth mobility, the beta/gamma power ratio, EEG central alpha power, and central-temporal gamma coherence ranked among the most consistent candidate predictive features across scenarios. Generalization to unseen stimulation intensities preserved a similar, moderately degraded performance profile (balanced accuracy 0.66–0.70). Conclusions: These findings provide proof-of-concept evidence that multimodal EEG-LFP decoding can identify epochs of dystonic muscle activity in paediatric dystonia.