# Quantitative color fundus photography parameters as potential biomarkers of axial length progression: evidence from a machine learning cohort study

## TL;DR

This study shows that machine learning models using retinal images can accurately predict myopia progression in children.

## Contribution

A novel machine learning model combining clinical data and retinal imaging metrics to predict axial length progression in children.

## Key findings

- The Random Forest model achieved an AUC of 0.961 in predicting axial length progression.
- Retinal venous density and vascular fractal dimension were key predictors of myopia progression.
- The model demonstrated favorable net benefit across clinically relevant thresholds.

## Abstract

Early identification of children at risk for accelerated axial elongation is essential for implementing timely myopia control strategies. Quantitative parameters derived from color fundus photography (CFP) may capture subtle structural and microvascular features relevant to axial length (AL) progression, yet their predictive value remains insufficiently characterized. To develop and validate a machine learning–based model integrating CFP-derived quantitative biomarkers and clinical characteristics to predict 1-year AL progression in school-aged children.

This cohort study included 693 children aged 6–10 years from Tianjin, China. AL progression >0.2 mm over 1 year was defined as significant elongation. Baseline clinical variables and 144 quantitative CFP metrics were evaluated. Feature selection was performed using Least Absolute Shrinkage and Selection Operator (LASSO) regression, logistic regression screening, and expert ophthalmologic assessment. Seven machine learning algorithms were developed using fivefold cross-validation, with hyperparameters optimized by grid search. Model performance was evaluated on an independent validation set using the area under the receiver operating characteristic (ROC) curve (AUC), F1 score, and other metrics. The best-performing model was interpreted using Shapley Additive Explanations (SHAP) and Local Interpretable Model-Agnostic Explanations (lime).

Of the 693 included children, 457 (65.9%) exhibited AL progression >0.2 mm. LASSO regression selected 39 candidate variables, and 12 predictors were ultimately incorporated into the model construction. Among all algorithms, the Random Forest (RF) model achieved the best discrimination, with an AUC of 0.961 (95% CI: 0.933–0.984) and the highest F1 score. Decision curve analysis (DCA) demonstrated a favorable net benefit across clinically relevant thresholds. SHAP analysis indicated that retinal venous density, venous fractal dimension, presence of leopard-spot lesions, vascular fractal dimension, and inferior-region vascular density were among the most influential predictors of AL progression.

The RF model, which combines clinical characteristics with CFP-derived quantitative biomarkers, accurately predicts short-term AL progression in children. Retinal microvascular and fundus structural parameters significantly contributed to model performance, underscoring their potential as early indicators of myopic AL elongation.