Causal Analysis of Parkinson's Motor Symptoms Using Structured Smartphone Accelerometer Data

Remote smartphone sensing enables longitudinal assessment of Parkinson’s disease (PD) motor symptoms. However, most existing studies rely on associational models that do not account for confounding. In this study, we use structured accelerometer data with automated data-quality control within a causally informed evaluation framework for smartphone-based PD classification. Using tremor and walking tasks, we first examined whether differences between PD and control participants remain after adjusting for age and gender. After stratified analysis and outlier removal, we did not detect significant differences in feature distributions at the p < 0.05 level. We then investigated identity confounding at the recording level. Participant identity could be predicted from sensor features far above chance, indicating strong subject-specific structure. Diagnosis prediction showed higher and more stable accuracy under record-wise cross-validation, and lower accuracy with larger variation under subject-wise cross-validation, consistent with identity-related bias and distribution shift. To reduce this effect, we introduced identity binning and a sample-weighted loss based on conditional probabilities. This produced a lower but more reliable estimate of diagnostic accuracy without requiring identity information at prediction time. These results show that reliable interpretation of smartphone-based PD classification requires both demographic and identity-aware deconfounding.