Decoding Dystonia: unveiling neural patterns with interpretable EEG-Based Machine Learning

Dystonia has a multifaceted and complex pathogenesis. Current diagnostic proce-dures, which focus primarily on clinical signs, may lack accuracy due to the variable presentationsof different dystonia types. There is a need for objective, interpretable, and non-invasive diagnostictools. This study aims to develop an interpretable electroencephalography (EEG)-basedmachine learning (ML) and deep learning (DL) approach to distinguish between focal upper limbdystonia (ULD), cervical dystonia (CD), and healthy controls (HC). EEG data were recorded during resting-state, writing-from-memory, and finger-tapping tasks. The EEG signals were segmented into windows to generate connectivity matricesusing various pairwise correlation metrics. Machine learning models were trained to classify thegroups, with performance evaluated using accuracy and area under the curve (AUC) metrics. Our approach achieved accuracy and AUC scores close to 100%. Transfer entropyemerged as the most effective connectivity metric, revealing altered brain connections in dystonia.Complex network measures outperformed traditional EEG features, highlighting the relevance offunctional connectivity. Resting-state EEG showed the highest classification performance for ULD,suggesting strong diagnostic potential. Conclusions: This study provides the first machine learning-based comparison between differenttypes of dystonia, introduces novel cervical dystonia EEG data, and yields medically interpretableinsights into altered brain connectivity. The findings enhance our understanding of dystonia and support using EEG as alow-cost, interpretable tool for diagnosing and developing brain-machine interfaces.