Ocular Microbiology and Immunology Group
Back to OMIG Main Page

< Previous | 2026 Agenda and Abstracts

2026 OMIG Abstract

POSTER PRESENTATION


Field Evaluation of an AI-based Smartphone Application for Automated Detection of Trachomatous Inflammation-Follicular in School-Aged Children in Thailand

Hady Yazbeck¹, Ifrah Khurram², Monthira Jermjutitham¹, Allison Summers¹, Oliver Valdivia Camacho¹, Duncan Fuller¹, Ratipark Tamornpark³, and Travis K Redd²

¹Casey Eye Institute, Oregon Health & Science University, Portland, Oregon; ²University of Colorado Anschutz Medical Campus, Aurora, Colorado; ³Mae Fah Luang University, Mae Fah Luang District, Chiang Rai Province, Thailand

Purpose: Artificial intelligence (AI) models for trachomatous inflammation-follicular (TF) detection have mostly been evaluated retrospectively, with no evidence from real-world screening workflows. The purpose of this study was to evaluate Trackoma, an AI-based smartphone application for automated TF detection from conjunctival photographs, during school-based screening in Thailand.

Methods: Trackoma was deployed in four schools in Mae Fah Luang District, Chiang Rai Province using two devices, a Samsung S22 Ultra and an iPhone 13 Pro Max. The application supported participant registration, image acquisition, local model inference, clinical grading entry, image-quality assessment, and data storage. Predictions were compared with masked expert grader assessments. Outcomes included workflow timing, image quality, device reliability, diagnostic performance, and population-level TF prevalence.

Results: 1480 eyes from 751 participants were examined using the Samsung device; and 826 eyes from 445 participants using the iPhone. Grader-derived TF prevalence ranged from 0.67% to 2.00%. The model missed the few grader-positive cases but maintained high specificity, resulting in low prevalence estimates that were consistent with grader-derived prevalence estimates. AUROC ranged from 0.60 to 0.83. Median screening time was under 40 seconds per eye, and median grading registration time was 6-7 seconds per image. Low-quality ranged from 3.7% to 11.2%. No application failures occurred on the Samsung device, whereas iPhone application crashes led to incomplete screening and three corrupted images.

Conclusions: Trackoma could be integrated into a school-based screening workflow and produced prevalence estimates consistent with expert grading in this low-prevalence setting. Positive-class detection was poor, although this should be interpreted in the context of very low TF prevalence. Larger external validation, improved sensitivity, and cross-device software stabilization are needed before independent deployment.