Demographic Confounding in Voice-Based Parkinson Disease Screening: Methodological Analysis of the Bridge2AI Voice Dataset

Abstract Background Voice-based deep learning models for Parkinson disease (PD) and dementia screening report areas under the curve (AUCs) of 0.85‐0.97, but rarely audit demographic confounding. Because speech changes substantially with age, case-control age imbalance alone can produce high classification performance independent of disease. Objective We audited the Bridge to Artificial Intelligence (Bridge2AI) Voice Dataset v3.0.0 with three objectives: (1) quantify age and site confounding in voice screening for PD and dementia, (2) evaluate whether a disease-specific acoustic signal persists after demographic adjustment, and (3) propose minimum reporting standards. Methods We fine-tuned an audio spectrogram transformer (AST; 86.4 million parameters) using 5-fold participant-level cross-validation for PD (n=253) and dementia (n=221). Logistic regression on age and sex provided a demographic-only baseline under identical splits. Confounding was assessed by (1) restricting evaluation to ages 60‐80 years, (2) restricting to US participants because all Canadian PD (n=62) and all Canadian dementia (n=70) participants were cases with no Canadian controls, (3) retraining AST from scratch on the age-restricted subgroup, (4) 1:1 nearest-neighbor propensity-score matching on age and sex in the US age-restricted PD subgroup, and (5) applying v3.0.0-trained models to v2.0.1 spectrograms of the same participants. Results On the full cohort, age alone was statistically indistinguishable from the AST (PD: AUC 0.875 vs 0.843, DeLong P =.36; dementia: 0.905 vs 0.895, P =.75), indicating a demographic shortcut; because both cohorts share the same 148-person control group, this parallel pattern is one dataset-level artifact, not 2 independent confirmations. Within ages 60‐80 years, the AST exceeded age-only for both conditions (PD: 0.787 vs 0.568, ΔAUC +0.225, 95% CI +0.089 to +0.361; P =.001, n=129; dementia: 0.809 vs 0.598, ΔAUC +0.215, 95% CI +0.056 to +0.375; P =.008, n=95), though the dementia result reflects residual site confounding. Three convergent estimates of PD-specific signal, post hoc age-restricted (0.787), de novo retrained (0.798), and propensity-matched (US, 28 pairs; 0.760; P =.02 vs age-only), all fell substantially below the 0.843 full-cohort AUC. Under the most conservative adjustment (US-only, ages 60‐80 years), AST AUC was 0.726, which we consider the least confounded estimate. Cross-version preprocessing testing produced AUC 0.618, indicating brittle representations. Of 13 published voice-PD studies, only 7 reported case and control age distributions, and none reported a demographic-only baseline or an age-restricted evaluation. Conclusions Demographic confounding dominates full-cohort voice-screening metrics on the Bridge2AI Voice Dataset for both PD and dementia. After age restriction, site control, and propensity matching, residual AST performance for PD (AUC 0.726‐0.798) remained significantly above demographic baselines, supporting a disease-specific acoustic signal substantially weaker than full-cohort metrics suggest; for dementia, insufficient US cases precluded similar estimates. Voice biomarker studies should report case and control demographic distributions, demographic-only baselines under identical splits, and age-restricted performance alongside full-cohort metrics.