Improving ART Retention Through Machine Learning-Guided Targeting of Interventions: A Monte Carlo Simulation Study in Lilongwe, Malawi

Retention in antiretroviral therapy care remains a major challenge in high-burden settings such as Malawi, where substantial loss to follow up undermines treatment outcomes and long-term epidemic control. Although machine learning models can accurately identify patients at high risk of disengagement, there is limited evidence on how these predictions can be translated into improved retention outcomes in practice. This study addresses this gap by linking machine learning-based risk stratification to the targeted allocation of retention interventions, providing a framework for evaluating their expected impact on ART retention outcomes. We developed a patient-level Monte Carlo simulation model that integrates individual predicted probabilities of loss to follow up from a validated Extreme Gradient Boosting model with intervention effect sizes derived from a meta-analysis of ART retention interventions conducted in sub-Saharan Africa. The study population included 1,705 ART patients receiving care at Lighthouse Trust clinics in Lilongwe, Malawi. Patients were stratified by predicted risk, and the highest-risk group (n = 512) was targeted for intervention. Six interventions were evaluated, including Expert Client support, psychosocial support, two-way text messaging, adherence clubs, community ART groups, and teen clubs, followed by subgroup-specific and combined approaches allocated based on predicted risk. The primary outcome was twelve-month ART retention, estimated over 5,000 simulation iterations. Subgroup and post-simulation analyses were conducted to assess heterogeneity in intervention response. Among patients classified as high risk (n = 512), baseline retention was 44.1%. Individual interventions improved retention to 52.7% with two-way texting (RR = 1.19; p < 0.001) and 55.0% with Expert Client support (RR = 1.25; p < 0.001). A combined intervention package produced larger gains, increasing retention to 64.0% (RR = 1.45; p < 0.001), corresponding to an absolute improvement of 19.9 percentage points. Intervention effects varied across subgroups, with significant improvements observed among newly initiated patients (43.0% to 58.9%; RR = 1.37; p < 0.001) and clinically unstable patients (28.3% to 39.1%; RR = 1.38; p = 0.01), while effects among adolescents were more modest (34.3% to 45.6%; RR = 1.33; p = 0.03). Despite these improvements, 46% of high-risk patients remained hard to retain after receiving multiple interventions. In this subgroup, expected retention increased only marginally from approximately 0.15 at baseline to 0.20 after intervention, with poor outcomes observed among patients who were virally unsuppressed, had depressive symptoms, or were younger. Machine learning-guided targeting of ART retention interventions can substantially improve retention outcomes, particularly when interventions are combined. However, a substantial subgroup of patients remains hard to reach and vulnerable to disengagement, indicating that existing strategies may be insufficient for individuals with complex clinical and psychosocial needs. This study contributes to knowledge by introducing an integrated framework that combines machine learning risk prediction, meta-analytic intervention effects, and patient-level Monte Carlo microsimulation to quantify twelve-month ART retention outcomes under risk-based targeting with subgroup-specific intervention allocation before real-world implementation. These findings highlight the potential of using individual risk to guide the delivery of retention interventions within routine ART programs to enable more efficient, proactive, and patient-centered allocation of retention resources.