Al-based Parkinson's disease diagnosis with word level speech data using advanced signal processing techniques

Abstract

Parkinson's Disease (PD) is a progressive neurodegenerative disorder that profoundly compromises patients' quality of life. Early and accurate diagnosis remains a clinical challenge, with voice signal analysis emerging as a promising non-invasive biomarker. Unlike the classical vowel sounds used in existing studies, we present a new PD sound dataset that includes a Turkish word "gofret"vocalized by the PD and control groups. The collected dataset was converted into images using Mel - Frequency Cepstral Coefficients (MFCCs), spectrograms, chromograms and tempograms in order to develop vision-based deep learning models from audio recordings. To mitigate data scarcity and enhance model generalizability, a suite of data augmentation strategies including frequency masking, time stretching, shifting and masking were systematically applied. In this study, alternative deep learnnig architectures were developed and their performances were compared. Quantitative evaluations, employing rigorous cross-validation protocols, demonstrated superior classification performance of spectrogram-based models with 85.6% accuracy in PD diagnosis, underscoring their robustness in capturing pathological vocal characteristics. The findings advocate for the integration of advanced augmentation techniques and multifaceted acoustic representations to bolster automated PD detection efficacy from voice data.