Background: Retention and viral suppression remain uneven among adults living with HIV (PLHIV) in Tanzania. To proactively identify individuals at high risk of disengagement from care, a machine-learning (ML) algorithm was previously developed using routine electronic health record data to predict disengagement risk. Cash transfer-enhanced adherence counseling (EAC) may improve viral suppression among people living with HIV (PLHIV) at high risk of disengagement, but benefits may be heterogeneous. Identifying which subgroups benefit most is important for equitable and cost-effective program targeting in resource-limited settings.

Methods: In the RKPK Objective 2 trial (four facilities in Geita, Tanzania), 692 machine-learning–flagged high-risk PLHIV were randomized 1:1 to EAC alone versus EAC plus cash. The primary outcome was viral suppression (<50 copies/mL) within 9-16 months. We used honest causal forests with cross-fitting to estimate conditional average treatment effects (CATEs) across baseline demographic, clinical, and ML-risk covariates. Because global heterogeneity diagnostics suggested limited precision for individual-level ranking, primary inference emphasized pre-specified subgroup effects with multiplicity control and facility-clustered bootstrap confidence intervals.

Results: Overall, 12-month viral suppression was 86.5% in the incentive arm versus 81.3% in standard EAC (risk difference +5.3%; 95% CI -0.3, +10.9). Global heterogeneity tests did not support reliable individual-level ranking. Variable-importance measures identified ML risk score, ART duration, sex, and enrollment period as the strongest candidate effect modifiers. Subgroup analyses identified larger effects among men with 3-7 years on ART (+27.9%; 95% CI +6.9, +49.0) and among participants in ML risk quintile 4 (60th-80th percentile; +18.5%; 95% CI +5.5, +31.5).

Conclusion: In this ML-targeted trial population, cash incentives layered onto EAC yielded modest overall benefit, with greater effects among men with intermediate ART duration and individuals at intermediate ML-predicted risk. These findings provide actionable guidance for equity-focused, precision implementation of EAC programs in sub-Saharan Africa.