A Lightweight Interpretable Machine Learning Framework for Parkinson Disease Detection with Feature Selection Technique

A degenerative neurological condition called Parkinson disease (PD) that evolves progressively, making detection difficult. A neurologist requires a clear healthcare history from the patients, as well as periodic scans, to make the diagnostic. In recent years, AI-based computer-aided diagnostic (CAD) programs have outperformed simpler approaches mainly because of their capacity to predict irregularities in healthcare data. Despite, the intricacy of AI models frequently leads to their employment as “black boxes” that may cause distrust among physicians owing to an absence of transparency regarding decision-making. This study introduces an interpretable machine learning approach to solve these difficulties. This approach offers both regional and worldwide insights for the auxiliary diagnostic of PD while preserving excellent prediction accuracy. This investigation used 894 healthcare instances contained several optimized characteristics. We used a two-stage data preparation strategy to manage extremes and equalize the data while preventing biased outcomes. We simulated multiple state-of-art ML models named boosting, voting and stacking with three features selectors such as mRMR, LDA, and PCA. Among these features selectors and models, the stacking + LDA approach provided the greatest accuracy of 100%. After that, two interpretable AI models named Local Interpretable Model-agnostic Explanations (LIME) and SHapely Adaptive Explanations (SHAP) are implemented for feature interpretability. This feature interpretability makes the proposed approach as a suitable candidate in medical sector.