Detection of candidate acoustic biomarkers for Parkinson's disease using an explainable machine learning composite spectrogram approach

Introduction Speech abnormalities are common in Parkinson's disease (PD), and machine learning (ML) offers a potential framework for objective and scalable speech-based diagnostics. This study introduces an explainable ML pipeline that leverages a novel vowel articulation-based composite input to detect PD and identify candidate phoneme-level biomarkers. Methods Two publicly available datasets of PD speech recordings were analyzed. Sustained vowel articulations were converted into log-mel spectrograms either individually or as a composite image by vertically concatenating a set of vowels per subject. Processed spectrograms were used to train ML models, with performance assessed using repeated random subsampling with five participant-level 80:20 splits and bootstrapped area under the receiver operating characteristic curve (AUROC). A support vector machine, logistic regression, and random forest trained on conventional acoustic features served as baselines. Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to quantify model attention across vowel regions. Results A total of 150 participants (49.3% PD) were included. After adjustment for age, sex, and dataset, cepstral peak prominence and harmonics-to-noise ratio were significantly higher in PD, particularly for the vowel /u/. The ML model achieved an average AUROC of 0.895 using individual vowels and 0.890 using the composite input ( p = 0.726). When trained on one dataset and tested on the other, the composite model achieved near-chance AUROCs of 0.654 and 0.557. The composite model achieved a higher AUROC than acoustic-feature baselines (0.746 to 0.803), significantly so for the support vector machine and random forest. Grad-CAM activation differed across vowels ( p < 0.001) and was highest for the vowel /u/ regardless of vowel order, consistent with acoustic findings. Conclusion The proposed explainable composite spectrogram approach enabled identification of a candidate vowel-level biomarker while retaining high classification performance. Concordance between ML and acoustic analyses suggests the potential of explainable ML in PD speech assessment, pending validation in larger cohorts and other languages.