A novel multivariate discriminant approach to predict sight threatening diabetic retinopathy (STDR) cases – data from the Liverpool Diabetic Eye Study

Abstract

Purpose : To identify patients developing STDR over a one year period using the patients’ demographic, clinical and retinopathy data collected over time.

Methods : Data from 13,775 people with diabetes registered in Liverpool between 2009 and 2016 were analysed using a novel multivariate discriminant approach. The dataset included: attendance to screening for DR, level of retinopathy assessed from digital photographs (41,350 screening episodes) as well as chronological primary care demographic and biochemical data.

Results : Retinopathy grade at first screening visit was distributed as follows: 63.8% No DR, 25.8% mild non-proliferative (NP) DR in one eye and 10.4% mild NPDR in both eyes. Level of retinopathy showed a gradual increase in risk of STDR across the 3 DR stages: 0.1% for no DR in either eye, 2.1% for mild NP in one eye and 18.1% for mild NP in both eyes. Duration of diabetes and HbA1C (median, IQR) were respectively (4.8, 2.3-9.4) years and (65, 53-87) mmol/mol in the STDR group, and (1.8, 0.3-4.6) years and (50, 44-60) mmol/mol in the no STDR group. Higher rates (difference in %) of appointments missed (20.3%), type I diabetes (15.8%) and males (6.6%) were observed in the STDR group.

Our predictive model achieved high levels of accuracy: 87.7% of all STDR cases (423 cases) and 83.2% of all no STDR cases were correctly predicted (sensitivity and specificity, respectively); the AUC of the ROC was 89.8% (95%CI: 89.5-90.1%). Cross-validation was conducted with 70% of the data for training and 30% for testing (process repeated 100 times).

Conclusions : Clinical measurements collected over time and baseline data can be jointly modelled to effectively identify patients who develop STDR within a year.

This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.