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

Typing dynamics are being investigated as a digital behavioral marker for Parkinson's disease (PD), but the cross-dataset reliability of typing-based classi ers remains uncertain. We evaluated four public datasets (475 par ticipants; 270 with PD and 205 controls) across 12 transfer directions using split conformal prediction targeting 90% marginal coverage. In all three pri mary directions, coverage was below the 0.90 target and the one-sided tests remained signi cant after Holm correction: 0.598 (95% CI, 0.4830.704) for Tappy→MIT, 0.850 (0.7880.899) for MIT→Tappy, and 0.732 (0.6840.777) for clinical→self-reported transfer. Discrimination and coverage did not con sistently change together: LightGBM had lower AUC than logistic regression in all 12 directions but higher coverage in 11, with larger prediction sets. Class-conditional coverage varied by target cohort: PD-conditional cover age was 0.214 for Tappy→MIT, whereas healthy-control coverage was 0.700 for MIT→Tappy and 0.627 for clinical→self-reported transfer. Recalibration with 9 and 18 labeled target participants raised the lower condence bound above 0.80 in the latter two directions; in Tappy→MIT, it remained below 0.80 even with 25 labels. Weighted conformal prediction yielded coverage of 0.9181.000 without target labels but returned the full label set for 86% 100% of participants. Stress tests did not reproduce the observed coverage patterns, and black-box shift estimation did not support pure label shift as the dominant explanation for Tappy→MIT. Evaluating typing dynamics as a digital marker for PD requires cross-dataset assessment of discrimination, marginal and class-conditional coverage, and recovery costs. Clinical screen ing utility has not been established.