interpretable-ml-parkinsons-qsm-fmri

name	interpretable-ml-parkinsons-qsm-fmri
description	Interpretable machine learning methodology for predicting Parkinson's disease motor severity (MDS-UPDRS Part III) from neuroimaging features — Quantitative Susceptibility Mapping (QSM) MRI and multiband multiecho resting-state fMRI Regional Homogeneity (ReHo). Uses SVR, Elastic Net, Random Forest, XGBoost with nested CV and SHAP interpretability. Full multimodal model explains 45.4% variance. QSH+c clinical model achieves 75% within ±5 UPDRS points. Activation: Parkinson's prediction, QSM MRI, ReHo fMRI, MDS-UPDRS, motor severity prediction, SHAP neuroimaging, multiband multiecho fMRI, interpretable ML Parkinson, quantitative susceptibility mapping
metadata	{"arxiv_id":"2607.02553","published":"2026-06-26","authors":"Aixa X. Andrade","tags":["Parkinsons-disease","QSM","fMRI","ReHo","MDS-UPDRS","interpretable-ML","SHAP","biomarker"]}

Interpretable ML for Parkinson's Disease Severity Prediction

Overview

Predicts Parkinson's motor severity (MDS-UPDRS Part III) from QSM MRI and multiband multiecho fMRI-derived ReHo features using interpretable ML (SHAP).

Dataset

• 28 participants (24 PD, 4 controls)
• Features: regional QSM (structural, iron deposition) + ReHo (functional, local connectivity)

Experimental Design

13 feature-set experiments:

1. Imaging-only (QSM features)
2. Imaging-only (fMRI ReHo features)
3. Clinical-only
4. Full fMRI
5. Full QSM
6. Full fMRI + Full QSM + Clinical — best global fit (R² = 0.454)
7. Selected QSM + Clinical — best clinical proximity (75% within ±5 UPDRS points, lowest MAE)
8. Reduced, multimodal variants

Models

• Support Vector Regression (SVR)
• Elastic Net
• Random Forest
• XGBoost
• Nested cross-validation for all models

Key Results

Model	R²	Clinical Accuracy	Notes
Full fMRI + Full QSM + Clinical	0.454	—	Best global fit
Selected QSM + Clinical	—	75% within ±5	Best clinical proximity, lowest MAE
Imaging-only (QSM)	meaningful	—	Carries predictive signal
Imaging-only (fMRI ReHo)	meaningful	—	Carries predictive signal
Clinical-only	weak	—	Baseline

SHAP Feature Importance

Top features: cerebellar, thalamic, striatal, insular, and motor cortical regions.

Key Insight

Structural (QSM) and functional (ReHo) imaging contribute differently depending on the clinical prediction goal:

• Global fit → combine all modalities
• Clinical precision → selected QSM + clinical variables

Methodology Steps

1. Extract regional QSM features (iron deposition, structural)
2. Extract ReHo from multiband multiecho resting-state fMRI (local functional connectivity)
3. Design 13 feature-set experiments (imaging-only, clinical-only, multimodal, reduced)
4. Train 4 model types with nested CV
5. Evaluate: R², RMSE, MAE, Pearson r, permutation testing, within ±5 UPDRS accuracy
6. SHAP for interpretability

Pitfalls

• Small sample size: 28 participants limits generalizability — always report permutation tests
• Control imbalance: 24 PD vs 4 controls — may affect model calibration
• Nested CV essential: With small N, standard CV overestimates performance
• SHAP for interpretability: Essential for clinical adoption — black-box predictions insufficient

References

• Andrade, A.X. (2026). "Interpretable machine learning predicts Parkinson's disease severity using motion-corrected QSM MRI and multiband multiecho fMRI features" — arXiv:2607.02553