Applying Machine Learning to Predict Loss to Follow-Up Among People Living With HIV in Haiti Using a National Electronic Medical Record Cohort.

Objectives: Loss to follow-up (LTFU) among people living with HIV (PLHIV) remains a major barrier to epidemic control. This study developed machine learning (ML) models to forecast individual risk of LTFU using routine electronic medical record (EMR) data. Methods: We analyzed data from Haiti's national EMR database, with 115,822 PLHIV receiving antiretroviral therapy (ART) across 167 health facilities from 2018 to 2024. We trained four ML models, including demographic, clinical, and institutional predictors for LTFU. Model performance was assessed across four quarters using F2-score, recall, precision, ROC-AUC, PR-AUC, and calibration. SHAP values were used to interpret key predictors of LTFU risk. Results: The CatBoost model trained with weight adjustment performed best across all quarters. The highest F2-score and recall were observed in the first quarter, with modest declines over time. Predictive features with the strongest influence included prior visit and viral load test frequency, ART dispensation patterns, and facility location. Conclusion: ML models using national EMR data can effectively forecast LTFU risk among PLHIV in Haiti. Incorporating these models into routine care systems can support proactive, tiered interventions to improve retention outcomes.