Abstract
Purpose :
We trained and assessed the performance of machine learning (ML) algorithms predicting diabetic retinopathy (DR) improvement in nonproliferative DR (NPDR) patients with DR Severity Scale (DRSS) 20-35 (mild NPDR) based on systemic and/or retinal imaging features.
Methods :
Baseline systemic data and retinal imaging features (optical coherence tomography [OCT], color fundus photographs [CFPs]) of patients with mild NPDR (DRSS 20-35) were pooled from the EUROCONDOR (NCT01726075) trial for neuroprotective eye drop assessment and the C-Tracer (NCT01607190) observational study, and were used to train random forest ML models predicting DRSS improvement over 2 years. We performed multifold cross-validation and compared model performances of ML algorithms when trained with systemic features only, retinal imaging features only, or combining all available features. The quantitative comparison is based on area under the receiver operator characteristics curve (AUROC).
Results :
Data from 309 patients/eyes with mild NPDR at baseline were used to train the ML models and test in a 10-fold cross-validation setting. At baseline, 130 and 179 patients/eyes had a DRSS level of 20 and 35, respectively. After 2 years, DRSS level was improved in 133 (43%) of these patients, did not change in 159 (51.5%) patients, and worsened in 17 (5.5%) patients. DR severity improvement was predicted with an AUROC = 0.62 (95% CI, 0.56, 0.67), based on systemic features only; AUROC = 0.71 (95% CI, 0.65, 0.77), based on retinal imaging features only; and AUROC = 0.76 (95% CI, 0.70, 0.82), based on both systemic and imaging features.
Conclusions :
Our results indicate that structural measurements from retinal images (OCT and CFP) have a higher predictive value than systemic features alone in predicting future DR improvement in patients with mild NPDR, while the combination of both feature families provided the best predictive outcome. Predictive ML models in NPDR patients could be used to inform personalized monitoring and follow-up.
This is a 2020 ARVO Annual Meeting abstract.