A multimodal explainable artificial intelligence framework for interpretable Parkinson's disease prediction.

Parkinson’s disease (PD) is a complex neurological disorder that is difficult to diagnose at an early stage. Early diagnosis and prediction of the disease can help improve the prognosis of patients and therapeutic interventions. Although machine learning (ML) has shown strong potential for PD diagnosis and prediction, its clinical use is limited by the lack of model interpretability. In this study, advanced but interpretable framework is proposed for Parkinson’s disease diagnosis by incorporating explainable artificial intelligence (XAI) and machine learning methods. The proposed framework estimates the probability of PD with concise and clinically relevant explanations using heterogeneous data, including neuroimaging, motor and non-motor symptoms, and clinical features. Several ML classifiers, including support vector machines, random forests, k-nearest neighbors, and decision trees, are equipped with XAI methods, SHAP and LIME, to provide better interpretability. Under the proposed framework, the ML models demonstrated better predictive performance than classic ML models. AdaBoost achieved the best performance with an accuracy of 93%, precision of 90%, recall of 90%, a F1-score of 90%, and an AUC of 0.95, representing an improvement of approximately %0.75 in accuracy over the best-performing baseline model. In addition, local and global explanations are provided by LIME, SHAP, and ELI5 to identify essential neuroimaging and clinical features. In summary, this study provides clinical validation for AI by achieving a trade-off between predictive accuracy and interpretability, to aid early diagnosis and personalized treatment of Parkinson’s Disease.