JMIR Publications

ABSTRACT

Background:

Diabetic retinopathy (DR) is the leading cause of preventable blindness worldwide. Machine learning (ML) systems show potential to enhance DR in community-based screening. However, predictive power models assessing their usability and performance are scarce.

Objective:

This study used data from three university hospitals in Korea to provide a simple and accurate assessment of ML-based risk prediction for DR development, which can be universally applied to adults with type 2 diabetes mellitus (T2DM).

Methods:

This study predicted DR using data from independent electronic medical record-based cohorts; namely, a discovery cohort (one hospital, n=68,009) and a validation cohort (two hospitals, n=18,895). The primary outcome was the presence or absence of DR at three years. Different ML-based models were selected through hyperparameter tuning in the discovery cohort and analyzed the area under the receiver operating characteristic curve in the validation cohort.

Results:

Among 68,009 patients screened for inclusion, 14,694 (21.61%) were eligible for study analysis, and 348 (2.37%) patients were referred for DR. For DR, the XGBoost system had an accuracy of 73.10% (95% confidence interval [CI], 71.27–74.93), with a sensitivity of 72.71% (71.03–74.39) and a specificity of 73.11% (71.27–74.94) in the original dataset. Among the validation data set, XGBoost had an accuracy of 66.86%, a sensitivity of 67.15%, and a specificity of 66.84%. The most common feature in the XGBoost model was dyslipidemia, followed by cancer, hypertension, chronic kidney disease, neuropathy, and cardiovascular disease.

Conclusions:

Among 68,009 patients screened for inclusion, 14,694 (21.61%) were eligible for study analysis, and 348 (2.37%) patients were referred for DR. For DR, the XGBoost system had an accuracy of 73.10% (95% confidence interval [CI], 71.27–74.93), with a sensitivity of 72.71% (71.03–74.39) and a specificity of 73.11% (71.27–74.94) in the original dataset. Among the validation data set, XGBoost had an accuracy of 66.86%, a sensitivity of 67.15%, and a specificity of 66.84%. The most common feature in the XGBoost model was dyslipidemia, followed by cancer, hypertension, chronic kidney disease, neuropathy, and cardiovascular disease.