Limited Target Data Calibration Improves Cross-Dataset Transfer for Parkinson's Disease Detection from Speech

This is the author manuscript corresponding to the article published online in the National High School Journal of Science on May 15, 2026. Published article: https://nhsjs.com/2026/limited-target-data-calibration-improves-cross-dataset-transfer-for-parkinsons-disease-detection-from-speech/ Automated detection of Parkinson's Disease (PD) from speech presents a low-cost, non-invasive screening tool, but its reliability across languages and recording conditions is uncertain. This study evaluated whether speech-based PD detection could transfer across diverse datasets that differ in language, recording conditions, and tasks under limited exposure to the target dataset. We investigated model performance in-dataset and in cross-dataset zero-shot testing, whether adding a small, speaker-balanced fraction of the target dataset during training improved performance, and if fairness (ΔFNRF−M, difference in the False Negative Rates between female and male speakers) depended on training alone or on threshold selection. Models were developed using public datasets, including NeuroVoz (Castilian Spanish), EWA-DB (Slovak), IPVS (Italian), MDVR-KCL (English-UK), and English-US sustained “Ah” sounds, with varied speech tasks and recording environments. A frozen Wav2Vec2 backbone with two task-specific classification heads (vowel and read/other) was evaluated using AUROC (Area Under the Receiver Operating Characteristic Curve), sex-stratified FNR, and ablation analyses. The results showed that while in-dataset testing yielded reliable performance (e.g., AUROC ≈ 0.73 for D2), zero-shot cross-dataset testing did not. Further, combining different datasets into a multilingual model was inadequate (e.g., AUROC ≈ 0.41 for multilingual model transfer to D2), but including a small portion of target data led to performance improvements (achieving AUROC ≈ 0.67). Overall, these findings suggested that speech-based PD detection may transfer better across datasets using limited, carefully sampled, target data. The results of this study should be interpreted as investigative evidence rather than demonstration of clinical readiness.