Machine Learning for Prediction of Parkinson's Disease Progression Using DAT-SPECT Imaging and Multimodal MRI

Introduction: This study was conducted to develop a machine learning model for the prediction of Parkinson's disease progression using baseline dopamine transporter single photon emission computed tomography (DAT-SPECT) imaging and magnetic resonance imaging (MRI).

Methods: In this IRB-approved study, patients were initially selected from the Parkinson's Progress Markers Initiative (PPMI) database, an international multi-center observational study. Baseline and follow-up Movement Disorder Society Unified Parkinson's Disease Rating Scale (MDS-UPDRS) motor subscale scores were used to measure disease progression. Patients were then stratified into slow progression vs. fast progression groups. Using the PyRadiomics package, radiomic features were extracted from the caudate nucleus, putamen, and ventral striatum on baseline DAT-SPECT and midbrain on baseline MRI. Decision tree (DT), bagging classifier (BAG), multilayer perceptron (MLP), logistic regression (LR), AdaBoost (ADA), random forest (RF), and support vector machine (SVM) algorithms were trained and validated on datasets from the PPMI. Accuracy, sensitivity, specificity, area under the receiver operator curve (AUC), and area under the precision-recall curve (PR AUC) were used for model evaluation. For each modality, the classifier with the highest AUC on the validation set was used for subsequent ensemble models. The ensemble model achieving the highest AUC on the validation set was selected as the final model and evaluated on an internal test set from PPMI and an external test set from our institution.

Results: 197 patients were included in this study. Seventy-five of these patients were classified as fast progression. The final ensemble model consisted of an RF classifier trained on clinical features and DT classifiers trained on DAT-SPECT and T2WI features. The final model achieved an AUC of 0.90, PR AUC of 0.82, accuracy of 0.75 (95% CI: 0.53-0.89), sensitivity of 1.00 (95% CI: 0.60-1.00), and specificity of 0.62 (95% CI: 0.36-0.83) on the internal test set. The model outperformed single-modality models on the external test set, achieving an AUC of 0.72, PR AUC of 0.75, accuracy of 0.64 (95% CI: 0.43-0.81), sensitivity of 0.55 (95% CI: 0.28-0.79), and specificity of 0.73 (95% CI: 0.43-0.91).

Conclusions: We successfully constructed a machine learning prediction model for PD motor progression combining clinical information and radiomic features from multimodal imaging of specific regions of the nigrostriatal system. Our model achieved improved performance by integrating clinical information with DAT-SPECT and T2WI imaging features, highlighting the strength of multimodal imaging in model construction. Our model demonstrated the potential to predict PD progression at baseline, which is of great significance for early intervention and management. It also suggests an optimal combination of DAT-SPECT and T2WI for patient evaluation.

• Download figure
• Open in new tab
• Download powerpoint