Parkinson’s Disease Detection from Bengali Voice Recordings Using Machine Learning Approach

Parkinson’s disease (PD) poses a growing health challenge globally, with Bangladesh experiencing a concerning rise in mortality rates attributed to the condition. Despite available diagnostic methods, early detection remains elusive and often entails cumbersome and costly procedures. This work proposes a creative strategy to deal with this problem by creating a model for early stage PD detection using Bengali voice datasets. Leveraging the power of acoustic analysis, the model aims to detect subtle vocal abnormalities associated with PD, offering an affordable and non-surgical diagnostic instrument. The study utilizes the Praat software for feature extraction and employs feature selection algorithms including LASSO, ReliefF, Recursive Feature Elimination (RFE) and Mann Whitney U test. The Extracted feature from ParkinsonVoiceBank dataset are included MFCC, Jitter, Shimmer, Formants analysis, Pitch and Pulse analysis, Intensity and Harmonicity analysis of voice. Various machine learning classifiers, including Logistic Regression, Random Forest, Decision Tree,Naive Byes, SVM, K-Nearest Neighbors (KNN), XGBoost and AdaBoost are employed for classification. Notably, using RFE feature selection in conjunction with XGBoost, the study achieves a remarkable accuracy of 91.3%, f1 score of 93.2% and precision of 93.7%.