Vision-Focused Nomogram for Assessing the Risk of Depression in Diabetic Retinopathy: Integrating Visual Function,...

To develop and validate a comprehensive nomogram integrating visual function, metabolic factors, and psychosocial measures for assessing the risk of depression risk in patients with type 2 diabetic retinopathy (DR), using a robust machine learning-based feature selection approach.

This cross-sectional study enrolled 490 DR patients from January 2020 to March 2025. Participants were randomly allocated to a training cohort (n=343) and an internal validation cohort (n=147). We assessed nineteen candidate variables. To identify the most significant and consistent predictors, a multi-step feature selection was employed: 1) LASSO regression with 10-fold cross-validation, 2) Stepwise Forward Logistic Regression, and 3) Random Forest importance ranking. The intersection of key variables from all three methods was used to build the final nomogram via multivariate logistic regression. Model performance was evaluated using ROC curves, calibration plots, 10-fold cross-validation, decision curve analysis, and SHAP for interpretability. Temporal validation was performed on an independent cohort (n=104) from the same center (June 2025-December 2025).

The overall prevalence of depressive symptoms was 29.18% (143/490). The integrative feature selection identified four robust independent predictors: glycated hemoglobin (HbA1c) (OR=1.716, 95% CI: 1.438-2.049), sleep quality index (PSQI score) (OR=1.306, 95% CI: 1.146-1.488), social support score (SSRS score) (OR=0.883, 95% CI: 0.824-0.947), and vision-related quality of life (VRQoL score) (OR=1.138, 95% CI: 1.094-1.183). The nomogram demonstrated excellent discrimination in the training cohort (AUC=0.918, 95% CI: 0.886-0.950) and internal validation cohort (AUC=0.867, 95% CI: 0.797-0.936), with good calibration. The 10-fold cross-validation yielded a stable average AUC of 0.891. SHAP analysis confirmed VRQoL as the most influential predictor. In temporal validation, the model maintained good performance (AUC=0.875, 95% CI: 0.805-0.946).

This vision-focused nomogram, developed through a robust multi-method feature selection process, provides an effective and interpretable tool for depression risk stratification in DR patients. It integrates key clinical and patient-reported outcomes, demonstrating good and stable discriminative performance, and can assist clinicians in early screening of high-risk individuals for timely intervention.