Abstract
•Machine learning identified predictors of early and late NAION progression•Early visual loss linked to OSA, diastolic pressure, and fellow-eye NAION•Later decline associated with metabolic and vascular dysregulation•SHAP analysis revealed systemic and ocular factors driving visual loss risk
To determine whether combinations of modifiable clinical/systemic risk factors and structured trial variables predict early disease progression in acute NAION, as a clinical-feature benchmark, using machine learning for multivariable analysis.
Secondary analysis of a multicenter, double-masked, sham-controlled, randomized clinical trial.
We analyzed 589 study eyes with acute NAION from 729 participants prospectively enrolled in the QRK207 trial who had separate Screening and Day 1 evaluations. Progression was evaluated at screening, Baseline, and Month 2. Only pre-treatment and placebo-group participants were included.
Visual loss was modeled using best-corrected visual acuity (BCVA), defined as ≥10- or ≥15-letter loss on the Early Treatment Diabetic Retinopathy Study (ETDRS) scale, and standardized automated perimetry (SAP) using censored average total deviation (avgTD). Logistic regression, random forest, XGBoost, and support vector machine classifiers were trained with 5-fold cross-validation. Performance (AUROC, PR-AUC, accuracy, sensitivity, specificity, F1-score) and SHapley Additive exPlanations (SHAP) identified systemic and ocular predictors of visual deterioration. No features were extracted from raw OCT scans, fundus photographs, or raw visual field images; analyses were limited to structured clinical/systemic and trial-captured variables (including numeric ophthalmic measures when available).
Model performance for visual function progression, and the clinical features contributing most to predicted risk.
Models showed modest performance (AUROC 0.59–0.77; PR-AUC up to 0.60 varied by endpoint and prevalence). Early decline was associated with fellow-eye NAION, obstructive sleep apnea, and higher diastolic pressure, while later progression reflected metabolic and vascular stress (elevated LDH, triglycerides, blood pressure, BMI). Preserved RNFL thickness, normal renal indices, and diabetes medication use were linked to lower risk.
Machine-learning models achieved modest discrimination but identified clinically relevant features distinguishing early from later NAION progression, supporting future biomarker-based and longitudinal modeling efforts. Findings suggest that improved prediction will likely require richer ophthalmic biomarkers (e.g., OCT/VF-derived features), multimodal models, and longitudinal approaches.