Deep Learning and Explainable Benchmarking for Early Parkinson’s Disease Detection Using Speech Signals in the United States

Abstract

Early-stage Parkinson’s disease (Early PD) detection using speech analysis has emerged as a promising and non-invasive approach for improving neurological healthcare in the United States. However, existing studies remain difficult to compare due to variations in datasets, speech tasks, languages, evaluation strategies, and definitions of Early PD. To address these limitations, this study proposes a comprehensive benchmark framework for speech-based Early PD detection using speaker-independent evaluation protocols to ensure fair, reproducible, and clinically reliable comparisons. The proposed benchmark evaluates multiple speech tasks under different training-resource settings and provides multidimensional performance analysis based on dataset characteristics, gender, aggregation level, and disease severity. Experimental findings offer actionable insights into the robustness and generalizability of speech-based Parkinson’s detection systems. The proposed benchmark establishes a reliable reference framework for advancing explainable, scalable, and clinically meaningful Early PD detection technologies within modern U.S. healthcare and neurological diagnostic systems.