Abstract
Purpose :
To predict visual acuity (VA) in patients with microbial keratitis (MK), at 90-day after diagnosis and at presentation, from data at the initial clinical ophthalmic encounter.
Methods :
Patients with MK were identified in the University of Michigan electronic health record between August 2012 and February 2021. VA was extracted for the affected eye in MK patients with a unilateral infection, or for the better seeing eye in patients with bilateral infections, at the time of diagnosis and at day 90 (final). Random forest (RF) models were used to predict initial and final VA (iVA and fVA) that were <20/40. Predictors included age, gender, iVA (for the 90-day prediction model), and information documented in the clinical notes at presentation but excluding the assessment and plan. Model diagnostics are reported with 95% confidence intervals (CI) for area under the curve (AUC), misclassification, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).
Results :
1791 MK patients were identified. Patients averaged 48.0 years old (standard deviation, SD=20.3) with 1734 (96.8%) unilateral infections and 57 (3.2%) bilateral infections. LogMAR iVA was on average 0.85 (Snellen equivalent=20/142; SD=1.25) in the affected or better eye and was <20/40 in 43.0% of patients. fVA was <20/40 in 26.6% of patients. The RF model for predicting fVA of <20/40 had an AUC of 0.95 (CI, 0.94- 0.97) and a misclassification rate of 11% (8-13%). The sensitivity, specificity, PPV, and NPV were 91% (86-95%), 89% (86- 91%), 73% (65-79%), and 97% (95-98%), respectively. Older age, worse presenting VA, and more mentions of “hypopyon,” “PKP,” and “OS” in the clinical note were found to be associated with 90-day VA <20/40. The RF model for predicting iVA of <20/40 had an AUC of 0.88 (CI, 0.85-0.91) and a misclassification rate of 17% (14- 20%). The sensitivity, specificity, PPV, and NPV were 77% (71- 82%), 88% (84-91%), 83% (78- 88%), and 83% (79-87%), respectively. Older age, less mentions of the term “quiet” in the clinical note, and more mentions of “hypopyon,” “OD,” and “BID” were associated with iVA of <20/40.
Conclusions :
RF models showed strong performance in predicting fVA and iVA. The model identified morphologic features, past ocular surgery, and indirect measure of medication use as key risk factors associated with poor VA outcomes. Predicting fVA can inform clinicians when risk stratifying patients with MK.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.