Predicting loss and delay to follow-up of amblyopia patients

Abstract

Purpose : Amblyopia can lead to irreversible loss of vision without timely monitoring and treatment. Previous work found associations between social determinants of health (SDoH) and amblyopia follow-up, and predicted general pediatric no-show risk. We extended the work to predict amblyopia patients at risk of delay or loss to follow-up.

Methods : Data consisted of newly diagnosed amblyopia patients seen at Casey Eye Institute in 2021 and 2022 with a recommended follow-up time of <6 months after the initial visit. The amblyopia status and follow-up recommendation were automatically identified using natural language processing methods on visit notes. We implemented a logistic regression model to predict whether a patient completed their first follow-up within 2 months of suggested date (predictors in Table 1b). We did a 90/10 train/test split (770/85 patients). We calculated odds-ratios and p-values (with Bonferroni correction for age) based on the Wald test for each predictor. We calculated accuracy, sensitivity, specificity, and AUC (area under the curve) of the Receiver Operating Characteristic (ROC) curve for the model on the test data.

Results : There were 855 patients in the cohort (446/409 yes/no follow up in 2 months); demographics are in Table 1a and predictors’ odds ratios and p values are in 1b. Commercial insurance was associated with higher odds of completing the follow-up without delay (p<0.001). Having at least one cancellation and English as preferred language were associated with lower odds of completing the follow-up without delay (p<0.001, p=0.03). Age <1 year had higher odds of completing the follow-up without delay as compared to age >5 years (p=0.01), but a Wald test of the effect of age overall was non-significant (p=0.06). Other predictors were non-significant. On test data, we achieved 70.6% accuracy, with 76.7% sensitivity and 64.3% specificity. Our AUC was 73.1% (CI: 61.9%-84.2%). The ROC curve is in Figure 1.

Conclusions : Our results are consistent with previous accuracy of no-show risk. Our significant predictors—English preferred, commercial insurance, having a cancellation—are all related to patient level SDoH. The community level SDoH predictor (ADI) was not, indicating a need to explore additional patient-level predictors. This model is a first attempt at the important aim of identifying amblyopia patients at risk of delayed care or loss to follow-up, in order to develop interventions and prevent vision loss.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.