Explainable multimodal retinal optical imaging for clinical parkinson's disease classification using fundus photography and OCT-derived features.

Research overview

The text below is the source article’s original abstract under a verified reuse license. cc by

BackgroundRetinal optical imaging offers a non-invasive photodiagnostic window into neurodegenerative disease. We evaluated whether bilateral fundus photographs, Optical Coherence Tomography (OCT)-derived quantitative retinal features, and leakage-controlled structured variables can support participant-level Parkinson's disease classification.MethodsDe-identified bilateral fundus photographs, OCT-derived quantitative retinal features, and structured tabular data were obtained from the Persian Cohort at Mashhad University of Medical Sciences. Diagnosis, medication status, UPDRS (Unified Parkinson's Disease Rating Scale) total scores, and all UPDRS item/severity-related variables were excluded from all predictive inputs before modeling. Three participant-level branches were evaluated: tabular-only, image-only Siamese EfficientNetV2-B1, and Fundus + Tabular FiLM (Feature-wise Linear Modulation) fusion. Evaluation used subject-level splitting, out-of-fold calibration, frozen thresholds, and additional shuffled-conditioning and image-ablated controls to assess the FiLM pathway.ResultsIn branch-specific held-out evaluation, the leakage-guarded tabular-only branch achieved ROC-AUC = 0.728 and PR-AUC = 0.194, the image-only fundus branch achieved ROC-AUC = 0.826 and PR-AUC = 0.527, and the Fundus + Tabular FiLM fusion branch achieved ROC-AUC = 0.845 and PR-AUC = 0.702 with calibrated Brier score = 0.030. At the frozen fusion threshold (0.143), sensitivity was 0.795 and specificity was 0.800. Shuffled-conditioning and image-ablated/FiLM-only controls reduced performance (ROC-AUC = 0.781 and 0.612, respectively), supporting dependence on aligned image-context integration.ConclusionsBilateral fundus images, OCT-derived quantitative retinal features, and leakage-controlled structured tabular variables supported participant-level PD classification in this single-center case-control cohort. The findings should be interpreted as preliminary classification results, not prospective risk stratification, and require age-controlled, externally validated evaluation before clinical translation.

Related topics