Personalized Prediction of Regional Brain Atrophy in Parkinson’s Disease through Multimodal Longitudinal AI Modeling
Tara Najafi1, Yusen Wu3, Phuong Nguyen3, Tatjana Rundek2, Yelena Yesha3, Ihtsham Haq1
1University of Miami Miller School of Medicine, 2Neurology, University of Miami Miller School of Medicine, 3Computer Science, University of Miami
Objective:
Novel spatio-temporal deep learning model that integrates with multi-modal patient data to forecast regional brain atrophy, offering clinically meaningful predictions and personalized care.
Background:
Predicting Parkinson's disease (PD) progression remains a challenge, with traditional tools struggling to capture individualized neurodegeneration patterns. AI advances offer solutions by integrating multimodal data, including MRI, clinical trials, and biomarkers, into a model that enables personalized forecasts. However, current models fail to capture temporal dynamics and focus on brain volume loss without accounting for region-specific degeneration.
Design/Methods:
We analyzed data from 197 PD patients with 424 records, each with 2-4 visits over 2-5 years. At each visit, T1-weighted 3D MRI scans and clinical assessments were completed, including cognitive scores, motor evaluations, CSF biomarkers, etc. The long short-term memory units (LSTM) based model performed regression to forecast future regional brain volumes and classification to identify regions at high risk of atrophy. Model performance was evaluated through cross-validation, with mean absolute error (MAE) and AUROC as primary outcomes. Interpretability was assessed by using SHAP feature importance.
Results:
5 brain regions achieved AUROC over 0.9, 11 regions are over 0.8, with the mean sensitivity exceeding 0.85, and several folds surpassing 0.9. The LSTM accurately predicts annualized regional brain atrophy across 15 regions. In all regions, MAE is lower than the average observed change, showing the model efficiently captures true change. Regression results for the putamen, brainstem, and hippocampus achieved remarkable MAE±SD [0.8%-1.5%]±[0.59%-1.40%]. Strong correlations existed between 94 predicted regional volumes and 2 critical clinical measures. 17 regions demonstrated strong correlations with UPDRS3 and MoCA, aligning with established clinical results.
Conclusions:
Regional brain atrophy in PD can be forecast with clinically useful accuracy with the strongest performance observed in 15 brain regions. These results highlight the potential for personalized prognostic tools integrating longitudinal 3D imaging with clinical data.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.