PD-VOST: Parkinson's Disease Voice Spectrogram Transformer

Abstract

Deep learning (DL) techniques are increasingly used for diagnosing Parkinsons Disease (PD) due to their non-invasive nature and accessibility. This study introduces PD-VOST (Parkinsons Disease Voice Spectrogram Transformer), an innovative DL model that employs the Transformer architecture for PD diagnosis using voice data collected via smartphones. Unlike conventional transfer learning approaches that typically use image-pretrained models for audio tasks, PD-VOST leverages a model pre-trained specifically on audio data. This specialization allows for more effective fine-tuning on voice recordings from individuals with PD and healthy controls. Our model achieved an average AUC of 95.89% and an average AUPRC of 87.11%, outperforming state-of-the-art methods. These results underscore the potential of employing audio-specific pre-trained models in advancing the early detection and management of PD using the voice as a biomarker.