Borderline trachoma cases can be identified with a non-dichotomous latent class analysis (LCA)

Abstract

Purpose : The WHO and other stakeholders aim to eliminate blinding trachoma by 2030, but this requires accurate grading of clinical cases. Disagreement amongst experts is common in trachoma grading. This study tested the hypothesis that non-dichotomous latent class analysis (LCA) can be used to identify discrepant cases.

Methods : De-identified photos of upper right tarsal conjunctivae of 200 children (0-5 y) from hyperendemic communities in Niger and Ethiopia were randomly selected from the databases of the Partnership for Rapid Elimination of Trachoma (PRET) and Trachoma Elimination Follow-up (TEF) trials (2003-2014). Ten trained graders independently evaluated 200 photos for the presence or absence of each grade (Trachomatous Inflammation–Follicular, TF, and Trachomatous Inflammation–Intense, TI). Given 3 potential result clusters (positive, negative, borderline), we conducted a 3-class LCA. Cohen’s κ-statistic for agreement between graders was calculated before and after removing the third class, hypothesized to represent a discrepant-case class (i.e., borderline).

Results : TF Prevalence by majority was found to be 0.42 (95% CI = 0.36-0.48). When the borderline cases were removed, the κ-statistic increased by 0.10 (95% CI = 0.72-0.85, P < 0.001) for TF and 0.13 (95% CI = 0.81-0.90, P < 0.001) for TI. Fig. 1 shows that the photos identified as borderline by the 3-class LCA were highly correlated with greater levels of disagreement between graders.

Conclusions : LCA with more than 2 classes can be used to identify borderline trachoma cases. Inclusion of the borderline cases in trachoma test sets reduces agreement considerably, making certification of graders more difficult although perhaps more realistic. Clinical diagnosis of trachoma that is not certain but probabilistic also has ramifications for training convolutional neural networks to perform grading. Further analytic research will be needed to determine the practical uses of LCA in large trachoma grading databases.

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

 

Figure 1. (A) The number of photos for x number of graders that said TF. (B) The probability of the photos being either TF or Non-TF for x number of graders that said TF, as predicted by the 2-class model. (C) The probability of the photos being Borderline, Non-TF, or TF for x number of graders that said TF, as predicted by the 3-class model. (D) (E) and (F) show the same data as (A) (B) and (C), respectively, for TI instead of TF.