An explainable EEG connectome framework for frequency-resolved analysis of functional connectivity alterations in Parkinson’s disease

Abstract Parkinson’s disease (PD) is a neurodegenerative disorder characterized by the degeneration of dopaminergic neurons, leading to a variety of motor and non-motor symptoms. Analyzing brain functional connectivity is crucial for understanding PD, with electroencephalography (EEG) serving as a valuable, non-invasive tool for this purpose. Integrating complex network analysis with explainable artificial intelligence (XAI) provides deeper insights into the brain’s neural mechanisms affected by PD. Personalized connectome analysis allows for more accurate diagnosis and treatment, as it considers individual variations in brain networks, enabling the identification of specific patterns associated with the disease. This study analyzed electroencephalographic signals from 31 subjects to create complex networks. In our study, we decomposed the spectral content of the EEG signal into different frequency sub-bands, creating a connectome for each sub-band. Key network features, such as degree and centrality measures, were extracted and fed into the XGBoost algorithm for classification and employed explainable artificial intelligence (XAI) models. These models offer a structured and organized framework for EEG data analysis, enabling full utilization of the information contained in different frequency bands. The overall model achieved an AUC of $$0.85\pm 0.03$$ , precision of $$0.84\pm 0.07$$ , and accuracy of $$0.80\pm 0.05$$ . Each connectome of the five EEG rhythms was analyzed separately, with the theta band connectome showing the best performance (AUC $$0.82\pm 0.03$$ , precision $$0.76\pm 0.04$$ , accuracy $$0.75\pm 0.03$$ ). The multi-connectome, considering all EEG rhythm connectome, was also analyzed obtaining $$0.72\pm 0.04$$ as AUC, $$0.67\pm 0.06$$ as accuracy and $$0.68\pm 0.07$$ as precision score. The SHAP algorithm was used to evaluate feature contributions. By leveraging this integration, the model seeks to capture both band-specific, multi-band and global patterns in brain activity, enhancing the accuracy and interpretability of the final predictions. This study offers new perspectives on brain connectivity and the neural mechanisms involved in Parkinson’s disease. It underscores the importance of EEG connectivity analysis across different frequency bands and the crucial role of XAI models in achieving a more comprehensive understanding of the neural connections, highlighting the mechanisms underlying Parkinson’s disease. Our contributions include the use of a frequency-specific connectome model, integration of multiple complex network features with an XGBoost classifier, and the application of SHAP-based XAI to enable interpretable, personalized analysis of PD-related brain connectivity changes.