Comprehensive Summary
This peer-reviewed single-cohort retrospective study used the Parkinson’s Progression Marker Initiative (PPMI) dataset: 150 participants (55 PD, 45 prodromal, 50 controls) with T1 MRI, SPECT, CSF proteins, and clinical scores (UPDRS, MoCA) from their most recent visit. A 1D-CNN classifier ingested 121 features (radiomics, SBR values, CSF markers, clinical scales, and five engineered ratios) after a 70/30 stratified train-test split with SMOTE applied only to the training set. The best configuration (all modalities and ratio features) achieved 93.7% accuracy for three-class classification, with improved precision and F1 versus reduced feature sets; AUROC, calibration, and confidence intervals were not reported. The CNN’s SHAP/LIME feature scores fed into a fine-tuned transformer LLM (ChatGPT-4o Mini), which generated patient-specific explanations, Q&A, and narrative reports within a cloud-based web portal.
Outcomes and Implications
This AI system has the ability to make a substantial impact in care for patients with Parkinson’s disease. As it synthesizes imaging, CSF biomarkers, and clinical scores, it may further reduce variability in interpretation by specialists and aid in moving towards more standardized decision making. The generated natural language explanations may further improve communication in clinical settings as they give neurologists a way to explain complex findings in more patient-friendly language. This system could serve as a single-point reference tool across radiology, neurology, and research teams in multidisciplinary centers like specialized movement disorder centers. This would accelerate diagnosis timelines and reduce repetitive testing.