Dynamic Prediction of Treatment Failure in Ocular Tuberculosis Using Machine Learning and Explainable AI

Le résumé fourni par la source

Purpose: Ocular tuberculosis (OTB) poses significant challenges in treatment because of its complex diagnostic and therapeutic landscapes. Predicting treatment failure effectively is crucial for timely intervention and improving patient outcomes. We report the application of machine learning (ML) approaches to (i) allow predictions using baseline data and (ii) dynamically update predictions based on patient history and new observations. Methods: The Collaborative Ocular Tuberculosis Study (COTS) was a multinational retrospective study encompassing data from 836 patients with tubercular uveitis across 27 international eye care centers. This study evaluated the performance of nine ML models to predict treatment failure at six, 12, and 24 months using baseline and longitudinal data. Metrics such as area under the curve (AUC), precision, accuracy, F1-score, and model complexity were reported. Top features and their importance were identified using XGBoost, with weight of evidence and information value calculated to enhance interpretability. Results: Data were collected from 836, 769, and 418 patients at six, 12, and 24 months, respectively. XGBoost and Random Forest (RF) models consistently showed superior performance across all timepoints. At 6 months, XGBoost achieved an AUC of 0.915 ± 0.019 and accuracy of 0.879 ± 0.027. At 12 months, RF outperformed with an AUC of 0.921 ± 0.011 and accuracy of 0.944 ± 0.022. At 24 months, RF maintained high accuracy (0.960 ± 0.029) despite a slight drop in AUC (0.888 ± 0.099). Deep Neural Networks and TT-net models were underfitted. Conclusions: ML models like XGBoost and RF demonstrate promise for early and accurate prediction of treatment failure in OTB, with explainability tools enhancing clinical interpretability. Translational Relevance: This study bridges basic ML research and clinical care by offering explainable, performance-driven models that support real-time, data-informed treatment decisions in managing OTB, potentially improving long-term outcomes.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.