Voice classification in Parkinson’s disease: A deep learning approach using transformers and error rate metrics

Alterations in voice and speech affect up to 90% of individuals with Parkinson’s disease, significantly impacting their quality of life. This study investigates the potential of speech-related error metrics and deep learning models for classifying Parkinsonian and healthy individuals and assessing disease severity. Word Error Rate and Character Error Rate were analyzed across groups and in relation to the Movement Disorder Society - Unified Parkinson’s Disease Rating Scale scores. The Vision Transformer and Audio Spectrogram Transformer were trained on vocal recordings using five-fold cross-validation and achieved classification performance up to 73% accuracy, 75% precision, 70% recall, and 72% F1-score. Error rate metrics were significantly higher in patients compared to healthy controls across multiple datasets, increasing with the severity of symptoms. This association suggests that greater vocal production errors reflect progression of the disorder. These findings highlight the potential of deep learning methods combined with speech analysis as non-invasive tools for early diagnosis and continuous monitoring of Parkinson’s disease, including remote assessment through mobile voice recordings. • Speech Error Metrics for Parkinson’s Disease- PD: The study demonstrates significant association of Word and Character Error Rates with disease severity. • Deep Learning Models for Classification and Severity Assessment: ViT and AST models demontrated to be sensitive to disease progression. • Generalizabily of the proposed models: Multiple datasets have been integrated, to validate models generalizability.