🧬 biology

Interpretable machine learning predicts Parkinson's disease severity using motion-corrected QSM MRI and multiband multiecho fMRI features

This study demonstrates that interpretable machine learning models integrating motion-corrected QSM MRI and multiband multiecho fMRI features can effectively predict Parkinson's disease motor severity, with combined structural and clinical variables yielding the most accurate and clinically relevant predictions.

Original authors: Aixa X. Andrade

Published 2026-07-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Aixa X. Andrade

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine trying to guess how shaky a person's hands are (a key sign of Parkinson's disease) just by looking at them. Doctors do this every day using a checklist called the MDS-UPDRS. But sometimes, the checklist misses the subtle, invisible changes happening inside the brain.

This paper is like a team of detectives trying to build a better "crystal ball" to predict that shakiness. Instead of just looking at the patient, they used two special types of brain scans as their clues:

The "Iron Detector" (QSM): Think of this as a metal detector for the brain. Parkinson's is linked to too much iron building up in specific areas. This scan measures that iron density, acting like a structural map of the brain's "rust."
The "Activity Map" (fMRI): This is like a thermal camera that shows which parts of the brain are "warming up" and talking to each other while the person is resting. It measures how synchronized the brain's local neighborhoods are.

The Experiment: A Cooking Competition
The researchers gathered a small group of 28 people (24 with Parkinson's and 4 healthy controls). They wanted to see if they could use these brain scans to predict the patient's shakiness score.

They set up a "cooking competition" with 13 different recipes (models). Some recipes used only the Iron Detector, some used only the Activity Map, some used only the doctor's notes (like age and medication dosage), and some mixed them all together. They also tried "reducing the ingredients," picking only the most important brain regions instead of looking at the whole brain, to see if that made the recipe tastier (more accurate).

The Results: What Worked Best?
Here is what they found, using simple terms:

The Scans Beat the Notes: If they tried to guess the shakiness using only the doctor's notes (age, medication), the prediction was weak. It was like trying to guess the weather by only looking at a calendar. However, the brain scans alone were much better at predicting the shakiness.
The "All-In" Mix was the Strongest: The recipe that combined both the Iron Detector and the Activity Map, plus the doctor's notes, gave the best overall picture of the brain's condition. It explained about 45% of the differences in shakiness among the group.
The "Targeted" Mix was the Most Accurate for Individuals: While the "All-In" mix was great for the group average, a different recipe worked better for guessing the score of a specific person. This recipe used a smaller, targeted list of brain regions from the Iron Detector (QSM) plus the doctor's notes. It got the score right within a very small margin (±5 points) for 75% of the people.
Less is Sometimes More: Surprisingly, looking at fewer brain regions (the "reduced" sets) often worked better than looking at the whole brain. It's like trying to find a needle in a haystack; if you only search the most likely spots, you find it faster and with less confusion.

The "Why": What the Brain Told Them
The researchers didn't just want a black box that gave a number; they wanted to know why it worked. They used a tool called SHAP (which acts like a highlighter) to see which clues mattered most.

The highlighter pointed to specific brain areas known to be involved in movement, such as the cerebellum (balance), thalamus (relay station), striatum (movement control), and insular cortex. This confirmed that the computer wasn't just guessing randomly; it was actually "reading" the parts of the brain that Parkinson's affects.

The Bottom Line
This study shows that combining a structural scan (iron levels) and a functional scan (brain activity) gives a clearer picture of Parkinson's severity than looking at either one alone or just using a checklist.

The Iron Detector tells us about the brain's physical structure.
The Activity Map tells us how the brain is currently functioning.
The Doctor's Notes add helpful context.

When you put them together, you get a much better prediction of how severe a patient's symptoms are. The study suggests that for the most accurate individual predictions, focusing on a specific, smaller set of brain regions related to movement is the winning strategy.

Technical Summary: Interpretable Machine Learning Predicts Parkinson's Disease Severity Using Motion-Corrected QSM MRI and Multiband Multi-Echo fMRI Features

Problem Statement
Current clinical assessment of Parkinson's disease (PD) motor severity relies on the clinician-rated MDS-UPDRS Part III scale, which captures observable symptoms but offers limited insight into the underlying neurobiological variation. While structural and functional MRI can complement clinical exams, prior studies have not systematically evaluated the combined value of Quantitative Susceptibility Mapping (QSM)—a structural marker sensitive to iron deposition—and multiband multi-echo resting-state functional MRI (fMRI) derived Regional Homogeneity (ReHo) features for predicting current motor severity. Furthermore, existing literature often focuses on total UPDRS scores, disease progression, or diagnostic status rather than the specific prediction of current clinician-rated motor examination scores using complementary structural and functional biomarkers.

Methodology
The study utilized a cohort of 28 participants (24 PD patients in the ON medication state and 4 healthy controls) imaged at UT Southwestern Medical Center. The analysis pipeline involved:

Feature Extraction:
- QSM: Mean magnetic susceptibility was extracted from 206 bilateral atlas-defined regions.
- fMRI: Regional Homogeneity (ReHo) values, quantifying local BOLD signal synchronization, were extracted from 135 regions using the Schaefer 2018 atlas.
- Clinical: Six variables were encoded, including sex, handedness, age, Levodopa Equivalent Daily Dose (LEDD), time since last dose, and diagnosis.
Experimental Design:
- Thirteen distinct feature-set experiments were conducted, comparing imaging-only, clinical-only, imaging-plus-clinical, full (all features), reduced (biologically selected features from motor-circuit regions), and multimodal inputs.
- Reduced Sets: Included 89 QSM and 47 fMRI features selected based on literature evidence of motor-circuit involvement.
Machine Learning Pipeline:
- Four interpretable models were evaluated: Support Vector Regression (SVR), Elastic Net, Random Forest, and XGBoost.
- Validation: A nested stratified cross-validation approach was used (5 outer folds for held-out evaluation, 4 inner folds for hyperparameter optimization) to prevent overfitting in this small-sample, high-dimensional setting.
- Evaluation Metrics: Performance was assessed via pooled held-out $R^2$ , RMSE, MAE, Pearson correlation, and a clinically interpretable metric: the proportion of participants predicted within ±5 points of the observed MDS-UPDRS Part III score (the Clinically Important Difference).
- Significance Testing: Full-pipeline permutation testing (200 permutations) was performed to ensure performance exceeded chance levels.
- Interpretability: SHAP (SHapley Additive exPlanations) analysis was applied to identify the most influential features driving predictions.

Key Results

Predictive Signal: Imaging-derived features carried meaningful predictive signal. Imaging-only models (both reduced and full) outperformed the clinical-only model ( $R^2 = 0.023$ ), indicating predictions were not driven solely by clinical variables.
Global Fit vs. Clinical Closeness:
- Global Fit: The strongest global fit ( $R^2 = 0.454$ ) was achieved by the full multimodal model combining 135 fMRI features, 206 QSM features, and clinical variables (Expt 6).
- Clinical Closeness: The most clinically accurate predictions were achieved by a reduced model combining 89 QSM features and clinical variables (Expt 7). This model achieved the highest proportion of predictions within ±5 points (75.0%) and the lowest MAE (3.997) among top performers, despite a lower $R^2$ (0.374) than the full multimodal model.
Modality Comparison: QSM and fMRI provided non-redundant information. While full QSM showed stronger global fit than full fMRI, reduced QSM combined with clinical variables yielded superior clinical closeness compared to reduced fMRI.
Feature Reduction: Biologically informed feature reduction improved performance in single-modality models (e.g., reducing fMRI from 135 to 47 regions increased $R^2$ from 0.138 to 0.294). However, for multimodal models, the full feature set often retained stronger global performance.
Model Performance: In this small-sample setting, margin-based SVR and regularized linear Elastic Net models performed more consistently than tree-based methods (Random Forest and XGBoost).
Biological Plausibility: SHAP analysis highlighted top predictors in biologically plausible regions, including the cerebellum, thalamus, striatum, insula, and motor cortex. Clinical variables like LEDD and time since dose were influential but did not dominate the feature rankings.

Significance and Claims
The authors claim that this study demonstrates that structural (QSM) and functional (multiband multi-echo fMRI ReHo) imaging capture distinct, interpretable dimensions of PD motor severity. The findings suggest that:

Non-Redundancy: Structural and functional imaging provide complementary, non-redundant predictive information rather than one modality being universally superior.
Goal-Dependent Optimization: There is no single "best" biomarker; the optimal feature set depends on the prediction goal. Full multimodal integration yields the best global fit ( $R^2$ ), while selected QSM plus clinical variables yields the most clinically close individual predictions.
Interpretability: The use of interpretable machine learning allows for the identification of neurobiologically plausible brain regions associated with motor impairment, moving beyond "black box" predictions.
Clinical Context: Clinical variables add contextual value, particularly when combined with reduced, biologically informed imaging feature sets, but are insufficient on their own for accurate severity prediction.

The paper concludes that QSM and multiband multi-echo fMRI-derived ReHo are viable, non-redundant biomarkers for estimating current PD motor severity, supporting a framework where structural and functional features are used to both predict severity and identify interpretable brain regions associated with individual differences in motor impairment. The authors maintain modesty regarding generalizability due to the small sample size and call for future validation in larger, multisite cohorts.

More like this