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

Area of Science:

• Ophthalmology
• Machine Learning
• Infectious Diseases

Background:

• Ocular tuberculosis (OTB) presents diagnostic and therapeutic challenges.
• Predicting treatment failure is crucial for timely intervention and better patient outcomes.

Purpose of the Study:

• To apply machine learning (ML) for predicting OTB treatment failure.
• To dynamically update predictions using patient history and new observations.

Main Methods:

• Utilized data from the Collaborative Ocular Tuberculosis Study (COTS), a multinational retrospective study of 836 patients.
• Evaluated nine ML models, including XGBoost and Random Forest (RF), to predict treatment failure at 6, 12, and 24 months.
• Employed explainability tools like weight of evidence and information value for feature importance.

Main Results:

• XGBoost and RF models demonstrated superior performance in predicting OTB treatment failure across all timepoints.
• At 6 months, XGBoost achieved an AUC of 0.915 and accuracy of 0.879.
• At 12 and 24 months, RF showed high performance with AUCs of 0.921 and 0.888, and accuracies of 0.944 and 0.960, respectively.

Conclusions:

• ML models, particularly XGBoost and RF, show significant promise for early and accurate prediction of OTB treatment failure.
• Explainability tools enhance the clinical interpretability of these ML models.
• This research bridges ML and clinical care, supporting data-informed treatment decisions for OTB management.