Vocal Feature Guided Detection of Parkinson’s Disease Using Machine Learning Algorithms

Parkinson's disease (PD) is a progressive neurodegenerative disorder that primarily affects the motor neurons of the brain and causes tremors, stiffness, and difficulty walking, balancing, and coordinating. Due to motor neuron loss, people with PD may suffer difficulties walking and talking as the disease develops. The changes in the vocal features in PD patients can be a significant metric for classifying Parkinson's disease for early prediction. Here we used some considerable vocal features such as jitter, shimmer, fundamental frequency parameters, Recurrence Period Density Entropy (RPDE), Noise/harmonic ratio (NHR), Pitch Period Entropy (PPE), Harmonics parameters, and Detrended Fluctuation Analysis (DFA) for Parkinson's disease prediction. In this paper, we applied ten different types of machine learning algorithms using vocal features for detecting PD and analyzed the machine learning performance among them. We applied XGBoost, LightGBM, Random Forest, Bagging, AdaBoost, Decision Tree, Logistic Regression, Support Vector Machine, K-Nearest Neighbor, and Naïve Bayes classifiers for detecting PD. Result: According to our overall analysis, LightGBM achieved the highest accuracy of 95%, AUC of 96%, sensitivity of 100%, and specificity of 93.33%, respectively. Moreover, LightGBM outperformed other machine learning algorithms and was determined as the proposed model in terms of accuracy and AUC score.