An Explainable Ensemble and Deep Learning Framework for Accurate and Interpretable Parkinson's Disease Detection from...

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Computational Linguistics

Background:

• Parkinson's disease (PD) is a neurodegenerative disorder impacting motor and speech functions, necessitating early diagnosis for improved patient outcomes.
• Current diagnostic methods may lack the sensitivity for early detection, highlighting the need for advanced analytical tools.
• Voice analysis offers a non-invasive approach to identify subtle changes associated with PD.

Purpose of the Study:

• To develop and validate a unified, explainable AI framework for the early detection of Parkinson's disease using voice data.
• To integrate ensemble and deep learning models with interpretable AI techniques for robust PD identification.
• To assess the performance of various machine learning and deep learning models in classifying PD based on acoustic features.

Main Methods:

• Extraction of acoustic features from the Parkinson's Voice Disorder Dataset.
• Evaluation of diverse machine learning models, including traditional classifiers, ensemble methods (Random Forest, LightGBM), and neural networks (CNN, LSTM, GAN).
• Application of Explainable AI (XAI) techniques to identify key acoustic biomarkers predictive of PD.

Main Results:

• Ensemble methods, particularly LightGBM and Random Forest, achieved state-of-the-art accuracy (98.01%) and ROC-AUC (0.9914).
• Deep learning models demonstrated capability in capturing complex voice patterns, yielding competitive results.
• XAI analysis identified nonlinear acoustic biomarkers (e.g., spread2, PPE, RPDE) as significant predictors, aligning with clinical observations of dysphonia in PD.

Conclusions:

• The proposed AI framework offers a scalable, non-invasive, and clinically relevant solution for early Parkinson's disease detection.
• The study highlights a strong balance between high predictive accuracy and model interpretability.
• The findings support the use of voice analysis combined with advanced AI for objective PD diagnosis.