Decoding Parkinson’s diagnosis: An OCT-based explainable AI with SHAP/LIME transparency from the Persian Cohort Study

Abstract

Background

Parkinson’s disease (PD) diagnosis remains challenging due to subjective clinical assessments and late-stage symptom manifestation. Retinal optical coherence tomography (OCT) biomarkers, reflecting neurodegenerative changes, offer a non-invasive diagnostic avenue. This study integrates retinal OCT with explainable artificial intelligence (XAI) to address PD diagnostic uncertainties.

Methods

Leveraging data from the Persian Cohort Study (202 PD patients, 972 controls), we developed a 6-layer deep neural network (DNN) combining OCT biomarkers (foveal thickness and volume) and clinical variables (motor scores, olfactory dysfunction). Synthetic Minority Oversampling (SMOTE) mitigated class imbalance (PD:Healthy ≈ 1:5). Model interpretability was ensured via SHAP (global feature importance) and LIME (local explanations).

Results

This study developed an explainable AI framework integrating retinal OCT biomarkers and clinical data to diagnose Parkinson’s disease (PD) with 95.3 % accuracy and 0.98 AUC-ROC. Using data from the Persian Cohort (1176 participants), the model identified SUPERIOR4 thickness (<120 µm) and foveal volume expansion (>0.15 mm³) as key biomarkers, alongside motor and olfactory deficits. SHAP/LIME provided interpretable thresholds (e.g., SUPERIOR4 <120 µm = high risk), while SMOTE mitigated class imbalance, reducing false negatives by 12 % without compromising specificity (94.8 %).

Conclusion

This study pioneers a transparent, OCT-based AI framework for PD diagnosis, emphasizing early detection through retinal neurodegeneration patterns. The integration of multimodal data, explainability, and imbalance robustness positions it as a scalable tool for resource-limited settings. Future work should validate biomarkers across diverse populations and standardize OCT protocols.