Parkinson’s disease detection using voice recordings and Deep Learning

Abstract

Parkinson’s disease (PD) is a progressive neurodegenerative disease with significant social and economic impact. PD is characterized by motor and non-motor symptoms, with vocal disorders such as dysarthria, dysphonia, and hypophonia preceding common motor symptoms in approximately 90% of PD patients at the early stages of the disease. PD’s detection and diagnosis during the early stages are considered crucial for the progression and management of the disease. TabNet is a canonical DNN architecture that uses sequential attention and DNN building blocks to implement a Decision Tree-like output manifold with soft, instant-wise feature selection giving the model local and global interpretability. TabNet was trained with a vocal features dataset available from the UCI Machine Learning repository. The dataset includes extracted features from voice recordings gathered from 252 subjects (188 PD – 64 control), with each subject providing three samples of the sustained phonation of the vowel /a/. Due to the multiple samples per subject and the unbalanced number of healthy individuals and patients with PD in the dataset, ensemble learning and the Leave-One-Subject-Out cross-validation were used to train TabNet in order to keep it unbiased. TabNet outperforms the classifiers in the literature with average and maximum accuracy for every sample reaching 94.5% and 95.2%, respectively. The average and maximum values for F1-score and MCC were 96.3%, 96.8%, and 86%, 87.8%, respectively. When doing majority voting of the three samples to make a final prediction of every subject, the average and maximum observed results were 95.9% and 97.2% for accuracy, 97.3% and 98.1% for F1-score, 89.4% and 92.8% for MCC metric.