O-50 Predicting virological failure in pediatric and adolescent human immunodeficiency virus (HIV) patients in Haiti: a crisis-adjusted machine learning approach
Author(s):
J Wu, B Shaw, B Oni, D Dorestan, M Bien-Aime, D Compere-Louis, N Rachel Labbe
Year of Presentation:
2026
Objective: Haiti’s ongoing security crisis has widened the
disparity in HIV viral suppression between adults (–80%)
and pediatric patients (–59%). Since standard adherence
monitoring fails to capture the disruptions caused by violence and displacement, we aimed to develop and validate a
machine learning model to predict virologic failure among
children and adolescents (ages 0–19), identifying high-risk
patients for proactive triage before virological rebound
occurs.
Methods: We conducted a retrospective study using Electronic Medical Records from the iSanté system. To account for irregular clinic attendance, we utilized a patient-centric sliding window approach, anchoring behavioral features to the most recent viral load test. We engineered crisis-specific features, including status change frequency, and proportion of time on treatment. Three tree-based models (XGBoost, LightGBM, CatBoost) were trained using 5-fold crossvalidation and hyperparameter tuning, optimized for Recall (F2-Score) to minimize false negatives in a high-risk setting.
Results: The cohort included 4,674 pediatric and adolescent patients with a virologic failure prevalence of 20.6%. The CatBoost algorithm achieved the best performance (AUC: 0.74; Recall: 75.1% at a 0.45 threshold), identifying nearly four times more failing patients than random selection. SHAP analysis revealed that historical failure and prior viral load results were the strongest predictors. Crucially, the average length of treatment interruptions outperformed demographics, confirming that the intensity of care gaps is a better predictor than simple missed visits. Clinic type and marital status also demonstrated strong predictive associations. Age-stratified analysis confirmed optimal performance among adolescents (15–19 yrs), with predictive drivers varying across age groups.
Conclusion: Conventional adherence metrics are insufficient in humanitarian crisis settings. A crisis-adjusted machine learning model successfully identified 75% of pediatric viral failures. By integrating metrics of instability into routine surveillance, programs can effectively allocate limited resources to the most vulnerable children, potentially closing the gap between pediatric and adult outcomes despite prevailing instability.