Predicting the development of clinically significant macular edema (CSME) in patients with mild nonproliferative diabetic retinopathy (NPDR) using machine learning (ML)

Abstract

Purpose : To develop and evaluate ML models employing systemic and/or retinal imaging features for predicting whether patients with mild NPDR will develop CSME within 2 years from baseline.

Methods : Systemic and retinal imaging features (optical coherence tomography and color fundus photographs) from 348 patients (129 female; mean age, 60.9 years) with mild NPDR (Diabetic Retinopathy Severity Scale [DRSS] 20–35) and type 2 diabetes enrolled in a cohort study (NCT00763802) were pooled. Data were acquired 6 months (M6) and 2 years (M24) after baseline. Linear regression models were evaluated for predicting these events using systemic data only, imaging features only, and both combined. Area under the receiver operating characteristics (AUROC) curve was employed as a performance metric.

Results : Table 1 presents a summary of data acquired at baseline, M6, and M24. Twelve patients developed CSME in the study eye by M6, and 34 patients by M24. When using data obtained at baseline, CSME at M24 was predicted with an AUROC = 0.550 (95% CI, 0.521, 0.580) with systemic data; 0.727 (95% CI, 0.700, 0.753) with imaging features; and 0.734 (95% CI, 0.708, 0.760) with both combined. When using data from M6, CSME at M24 was predicted with an AUROC = 0.536 with systemic data, 0.734 with imaging data, and 0.713 with both combined. When using data combined from both M6 and baseline, CSME at M24 was predicted with an AUROC = 0.508 with systemic data, 0.749 with imaging features, and 0.738 with both combined. These results are summarized in Table 2.

Conclusions : Our results indicate that development of CSME in patients with mild NPDR can be more accurately predicted with retinal imaging features than with systemic data on their own. When predicting the development of CSME by M24 using data obtained at M6 only, the performance was not significantly different compared with using both baseline and M6. In the future, we plan to extend the evaluation to larger datasets, allowing more confident conclusions regarding the performance and predictive potential of the different types of data. Such predictive models of CSME in patients with NPDR could be employed to inform personalized monitoring and follow-up.

This is a 2021 ARVO Annual Meeting abstract.

 

Table 1. Mean values of the features

 

Table 2. AUROC performance when using data obtained from either M6 alone or from both baseline and M6