An ensemble machine learning classifier for Parkinson's disease diagnosis using optical coherence tomography...

Parkinson's disease (PD) is the fastest-growing neurodegenerative disorder worldwide, yet its early diagnosis remains a major challenge due to the absence of reliable biomarkers. Emerging evidence indicates that retinal microvascular alterations, detectable through Optical Coherence Tomography Angiography (OCTA), may serve as promising non-invasive biomarkers for PD. However, the lack of definitive diagnostic tests for early-stage PD underscores an urgent need for objective, non-invasive tools to facilitate timely detection and intervention. In this retrospective study, OCTA images were obtained from 53 PD patients and 39 healthy controls. Both the superficial vascular complex (SVC) and deep vascular complex (DVC) were segmented to extract 22 quantitative features, including foveal avascular zone (FAZ) descriptors and vascular density measures. A patient-based cross-validation strategy was employed to partition the dataset into training, validation, and independent test sets, ensuring that data from the same individual did not appear across multiple subsets. To reduce dimensionality and enhance generalizability, we applied a combined feature selection framework using Univariate Feature Selection, Recursive Feature Elimination, and Random Forest Feature Importance. Multiple machine learning algorithms were then trained and optimized, with the best-performing classifiers (XGBoost, Random Forest, and K-Nearest Neighbors) integrated into a weighted ensemble model. The ensemble approach outperformed individual classifiers, achieving an accuracy of 74.28%, sensitivity of 90%, specificity of 53.33%, and an AUC of 0.75 on the independent hold-out test set. Feature analysis revealed that both morphological descriptors (form factor, convexity, solidity, roundness) and vascular density parameters, including vessel area density (VAD) and vessel skeleton density (VSD) contributed strongly to model performance. A graphical user interface (PDAI - Parkinson's Disease Artificial Intelligence) was developed to facilitate clinical adoption, enabling interactive preprocessing, feature visualization, and automated prediction. Our findings provide a promising and non-invasive framework to support PD classification and screening, warranting further validation in larger and multi-center cohorts.