A Cross-Dataset Audit of Discrimination and Conformal Prediction Coverage in Keystroke-Based Parkinson's Disease Classi cation

Split conformal prediction for keystroke-based Parkinson’s disease classification targets 90% marginal coverage but achieves substantially lower coverage in primary transfer directions, including 0.598 for Tappy[->]MIT.

• Why it matters: Cross-dataset reliability is unclear for typing-based classifiers as digital markers, and without coverage guarantees clinical screening utility cannot be established.
• What they did: Four public datasets totaling 475 participants (270 with PD, 205 controls) were evaluated across 12 transfer directions using split conformal prediction with 90% marginal coverage and Holm-corrected one-sided tests.
• The result: Coverage was below target in primary directions (0.598, 0.850, 0.732), discrimination and coverage varied independently, and recalibration with 9–18 target labels raised lower bounds above 0.80 in two directions but not Tappy[->]MIT.

AI summary of the abstract · may contain errors · not medical advice · read the paper.