Speech-Based Parkinson's Disease Screening: A Deep Learning Approach Using Acoustic Biomarkers and Invertible Neural Networks

Parkinson's disease (PD) affects over 10 million people worldwide, with speech-related impairments reported to be present in 90% of patients with PD and typically occurring before any motor symptoms (Dorsey et al., 2018; Su et al. 2025). Therefore, non-invasive and accessible screening methods are required for population-based early detection, especially in resource-limited settings where access to specialists is limited (Ngo et al., 2022; Xu et al. 2025). This Paper is a novel deep learning architecture that combines deep residual encoders and squeeze-and-excitation attention mechanisms with invertible normalizing flows for speech-based PD screening. On the Telemonitoring dataset, the model achieved an AUC of 0.979, 100% specificity, 72.4% sensitivity, and 79.5% accuracy. The Multiple Sound Recording dataset demonstrated balanced classification with 71.1% accuracy, an AUC of 0.806, and symmetric sensitivity 71.0% and specificity 71.2%. Misclassified samples had approximately four times more prediction variance, supporting the uncertainty quantification capabilities of the architecture (Xu et al. 2024). A major clinical advancement for population screening purposes was observed when 100% specificity of telemonitoring speech data was attained, as this removes the issue of false positives. Using invertible normalizing flows enables exact density estimation and principled uncertainty quantification; this technology aids in the creation of ambiguity detections (cases that would require further clinical investigation). Therefore, this technology supports telemedicine applications related to smartphone remote screening of Parkinson's disease.