Smartphone-based prediction of dopaminergic deficit in prodromal and manifest Parkinson's disease

Dopamine transporter (DaT) SPECT can confirm dopaminergic deficiency in Parkinson's disease (PD) but remains costly and inaccessible. We investigated whether brief smartphone-based motor assessments could predict DaT scan results as a scalable alternative. Data from Oxford and Genoa cohorts included individuals with iRBD, PD, and controls. Machine learning models trained on smartphone-derived features classified DaT scan status and predicted striatal binding ratios, compared with MDS-UPDRS-III benchmarks. Among 100 DaT scans, the smartphone-only XGBoost model achieved AUC = 0.80, improving to 0.82 when combined with MDS-UPDRS-III (AUC's gender-corrected). A simpler logistic regression model performed better with MDS-UPDRS-III alone (AUC = 0.83) versus smartphone features, with slightly higher performance when combined (AUC = 0.85). Regression models predicted binding ratios with modest error (RMSE = 0.49, R² = 0.56). Gait, tremor, and dexterity features were most predictive. These findings support smartphone-based assessments complementing clinical evaluations, though larger independent validation remains essential.