Validating and updating the OHTS-EGPS model predicting 5-year glaucoma risk among ocular hypertension patients using electronic records

Abstract

ObjectiveTo validate and update the OHTS-EGPS model predicting risk of conversion from OHT to glaucoma using electronic medical records (EMR). DesignEvaluation and update of a risk prediction algorithm using EMRs and linked visual field (VF) tests. ParticipantsNewly diagnosed OHT patients attending hospital glaucoma services in England. Inclusion criteria: IOP 22-32 mmHg (either eye); normal baseline VF test, defined as Glaucoma Hemifield Test (GHT) ‘within normal range’ in a reliable VF test; at least two VF tests in total; no significant ocular co-morbidities. MethodsRisk factors: age, ethnicity, sex, IOP, vertical cup-to-disc ratio, central corneal thickness, VF pattern standard deviation, family history of glaucoma, systemic hypertension, diabetes mellitus, glaucoma treatment. Glaucoma conversion was defined as two consecutive and reliable VF tests with GHT ‘outside normal limits’ and/or need for glaucoma surgery. For validation, the OHTS-EGPS model was applied to predict a patient’s risk of developing glaucoma in 5 years. In the updating stage, the OHTS-model was re-fitted by re-estimating the baseline hazard and regression coefficients. The updated model was cross-validated and several variants were explored. Main Outcome MeasuresMeasures of discriminative ability (c-index) and calibration (calibration slope) were calculated and pooled across hospitals using random effects meta-analysis. ResultsFrom a total of 138,461 patients from ten hospital glaucoma services in England 9030 patients with OHT fitted the inclusion criteria. A total of 1530 (16.9%) patients converted to glaucoma during this follow-up period. The OHTS-EGPS model provided a pooled c-index of 0.61 (95% confidence interval: 0.60, 0.63), ranging from 0.55 to 0.67 between hospitals. The pooled calibration slope was 0.45 (0.38, 0.51), ranging from 0.25 to 0.64 among hospitals. The overall re-fitted model performed better than the OHTS-EGPS model, with a pooled c-index of 0.67 (0.65, 0.69), ranging from 0.65 to 0.75 between hospitals. ConclusionsWe performed an external validation of the OHTS-EGPS model in a large English population. Re-fitting the model achieved modest improvements in performance. Given the poor performance of the OHTS-EGPS model in our population, one should use caution in its application to populations that differ from those in the OHTS and EGPS.