Optimizing Parkinson’s disease detection: A deep dive into ensemble classifiers with feature selection, hyperparameter tuning, and integrating lime analysis

The degenerative neurological disorder Parkinson’s disease (PD) mostly damages older individuals’ neurological system and mobility. Despite increasing research into treating the disease’s underlying causes, Parkinson’s disease (PD) detection remains a major concern since there are currently no recognized laboratory-based tests for the condition. Speech problems and other abnormal speech patterns are common in Parkinson’s disease patients. Although early detection is crucial, there are still obstacles due to the absence of medical resources and expensive testing methods. Telemedicine devices and speech pattern analysis provide a practical, non-invasive diagnosis method. This study aims to diagnose PD by utilizing patient audio data to evaluate and implement many ensemble classification algorithms. We have used both grid and randomized search with cross-validation to adjust the hyperparameters of the AdaBoost, LightGBM, XGBoost, CatBoost, and random forest ensemble classifiers. Our suggested framework combines feature selection methods and dimensionality reduction method, such as Principal Component Analysis (PCA), Forward Selection (FS), Maximum Relevance Minimum Redundancy (mRMR), and Recursive Feature Elimination (RFE), to find important features that are crucial to the evaluation of PD. After employing mRMR for feature selection and utilizing grid search and random search cross-validation, the experimental results indicated that the LightGBM model outperformed other machine learning models, reaching an accuracy of 97.44%. Initially, before hyperparameter tuning, the accuracy of LightGBM was 92.31%, demonstrating an improvement of 5.13% following the tuning process. To achieve explainability in model predictions, we finally combined a LIME-based explainable model with the best-performing LightGBM ensemble model. This interpretable ensemble model makes it possible to identify patients early and treat them in a way that makes sense to medical practitioners.