A pilot study of machine learning models for prediction of treatment response in patients with diabetic macular edema in a phase II clinical trial

Abstract

Purpose : To develop machine learning (ML) models using baseline clinical and imaging characteristics to predict treatment response to ranibizumab or faricimab in diabetic macular edema (DME) patients enrolled in BOULEVARD.

Methods : A total of 195 patients with best corrected visual acuity (BCVA) and central subfield thickness (CST) measurements at week 24 in the phase II trial BOULEVARD (NCT02699450) were included in the model development. For each patient, 20 features on demographic and clinical characteristics as well as 250 features from an in-house retinal optical coherence tomography (OCT) segmentation algorithm (ROSA) at baseline were collected from the study eye. Those two sets of features were divided into 2 levels each, which were then combined in 6 different ways and used as the model input (Figure 1). Two types of ML models, linear model and eXtreme Gradient Boosting tree (XGBoost), were trained to predict BCVA at week 24 and evaluated by nested 5-fold cross-validation (CV). The mean coefficient of determination (R² score) over outer folds was reported as a performance metric. Feature importance ranking was calculated based on the coefficients for the linear model and the variance reduction for XGBoost.

Results : R² scores of the developed models ranged from 0.33 to 0.46 (Table 1). XGBoost with only level-1 demographic and clinical features had the best performance. While the ranking of feature importance varied among different models, baseline BCVA and age consistently ranked at the top.

Conclusions : All developed ML models showed moderate predictive power for BCVA at week 24, with demographic and clinical features providing most of the information needed for model predictions. There was no significant improvement in model performance with ROSA features. This pilot study of DME treatment response prediction is potentially useful for clinical trial design and treatment planning in clinical practice. Future studies with large and diverse datasets are needed to validate these preliminary findings and explore the potential of ML models in this problem.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Figure 1. Demonstration of tiered features

 

Table 1. Mean nested cross-validation results