VoicePD: Parkinson's Disease Detection Using Vocal Biomarkers with a Hybrid EfficientNet-LSTM Model

Overview

Globally, over 10 million individuals are affected by Parkinson's disease and it is the second most common neurodegenerative disorder. Early diagnosis is extremely important as it assists in better symptom management and improving quality of life for patients. Vocal biomarkers including reduced pitch and subtle tremors can occur potentially years earlier than traditional symptoms like rigidity. Our project aims to detect potentially early Parkinson's disease from voice recordings of patients using machine learning. Our project is non-invasive, low-cost and accessible. Parkinson's is difficult to diagnose and typically diagnosed late due to several factors including human error and the absence of definitive tests to detect early Parkinson's. We acquired our data from the public dataset mPower. The dataset contains 40,000+ samples of voice recordings and metadata from both Parkinson's and non-Parkinson's patients. Since we were limited computationally we selected 15,000 samples to train our model on. We took 7,500 samples from Parkinson's patients and 7,500 from non-Parkinson's patients to reduce bias. We converted the raw audio files into mel spectrograms which were fed into a pre trained EfficientNet-B0 backbone + bidirectional LSTM. After training the model on 80% of our data (12,000 samples) we validated on the remaining 20% (3,000 samples). Using 5-fold stratified cross-validation, our revised model achieved an accuracy of 95.5% and an F1 score of 0.955. Our project proved that vocal biomarkers are viable for Parkinson's detection and may assist in future development. We met our criteria of reaching accuracy above 90%.