A Machine Learning Approach to Voice-Based Parkinson Disease Screening Using Multiview Spectrogram and Speech Recognition Features: Diagnostic Study.

BACKGROUND: Parkinson disease frequently manifests early vocal impairment, motivating the development of noninvasive and scalable digital screening tools. OBJECTIVE: This study proposes a multiview spectrogram-based deep learning framework integrating recognition-aware context for Parkinson disease detection from voice recordings. METHODS: Voice recordings from 203 participants (121 with Parkinson disease and 82 healthy controls) were collected prospectively. Three spectrogram representations (Mel, short-time Fourier transform, and constant-Q transform) were extracted and processed through parallel convolutional neural network branches. A recognition ratio (RR) feature vector derived from automatic speech recognition transcript agreement was optionally fused with spectrogram embeddings. Models were evaluated using strict subject-wise 5-fold cross-validation. RESULTS: Multiview spectrogram recognition-aware Parkinson detection network achieved a mean test accuracy of 86.9% (SD 25.2%) using 3-view spectrogram fusion, improving to 97.4% (SD 5.7%) when incorporating the RR feature. RR integration reduced the false negative rate by approximately 84.5%, substantially improving sensitivity in screening-oriented settings. CONCLUSIONS: Combining multiview spectrogram learning with recognition-aware context significantly enhances voice-based Parkinson disease classification under leakage-free evaluation. These findings support the potential of this approach for noninvasive screening in structured recording settings, while further validation in diverse real-world environments is needed.