ML predicts Parkinson's severity with 75% accuracy from QSM and fMRI

A new study from Aixa X. Andrade leverages interpretable machine learning to predict Parkinson's disease motor severity from structural and functional MRI data. Using QSM (quantitative susceptibility mapping) and multiband multi-echo resting-state fMRI-derived ReHo features from 28 participants (24 with Parkinson's, 4 controls), the team trained four models: support vector regression, Elastic Net, Random Forest, and XGBoost. Nested cross-validation with pooling produced held-out R², RMSE, MAE, and permutation testing. The best global fit came from combining full fMRI, full QSM, and clinical variables, explaining 45.4% of motor severity variance. Remarkably, a reduced model using only selected QSM features plus clinical variables delivered the most clinically useful predictions—75% of participants predicted within ±5 points on MDS-UPDRS Part III, with the lowest MAE among top performers.

The interpretability aspect is key: SHAP (SHapley Additive exPlanations) highlighted distinct contributions from cerebellar, thalamic, striatal, insular, and motor cortical regions. This demonstrates that structural iron deposition (QSM) and functional connectivity (ReHo from fMRI) capture complementary dimensions of Parkinson's severity. The authors emphasize that imaging-only models carried meaningful predictive signal, while clinical-only models performed weakly. This work moves toward objective, non-invasive biomarkers for Parkinson's progression tracking, potentially enabling earlier interventions and personalized treatment adjustments. The code and data are not yet released, but the paper on arXiv (2607.02553) provides full methodological details.