AI-driven prediction of HIV-1 salvage therapy outcomes: a TRIPOD+AI compliant cohort study

Background

Achieving “Undetectable=Untransmittable” in multidrug-resistant HIV management is challenged by linear Genotypic Susceptibility Scores (GSS) that overlook viral epistasis and host interactions, leading to futile treatment and resource waste. We developed a TRIPOD+AI-compliant XGBoost model to predict salvage therapy outcomes, ensuring transparency and fairness.

Methods

Using Stanford HIV Drug Resistance Database data (n = 6,844), we built an XGBoost model integrating viral genetics, immune markers, and treatment history to predict viral suppression (<50 copies/mL). Feature contributions were quantified via exact TreeSHAP using marginal background distributions. Interpretability robustness was validated by assessing feature ranking stability across 5-fold cross-validation. Fairness was evaluated in severe immune-suppressed (CD4<200 cells/µL) and high-viral-load subgroups.

Results

XGBoost achieved superior discrimination (AUC: 0.887 vs. 0.816) and calibration (Brier: 0.096) compared with the GSS-based model. As a “circuit breaker,” it correctly reclassified 61.9% (671/1,084) of failures misidentified by GSS (NRI: 0.490). Stability analysis confirmed consistent feature rankings (mean Spearman’s rho = 0.9452), validating signal robustness. SHAP revealed patterns consistent with an “immune buffer” effect, in which robust immunity (CD4>233 cells/µL) attenuated the risk of high viral load, highlighting a protective mechanism absent in severe cases. Crucially, the model maintained high accuracy (AUC: 0.871) in severely immunosuppressed patients, with no systematic bias.

Conclusions

This model surpasses traditional scores by quantifying nonlinear interactions. Its “circuit breaker” function prevents futile regimens and the accumulation of resistance. Validated fairness ensures equitable care. Implementing this tool enables optimized antiviral allocation, supporting global HIV elimination targets via precision medicine.

Key messages

• An interpretable XGBoost model acts as a “clinical circuit breaker,” correctly reclassifying 61.9% of futile treatments to prevent resistance accumulation and optimize resource allocation.

• The model demonstrates algorithmic fairness in severe cases, supporting “U=U” goals by guiding aggressive interventions for immunocompromised patients without systematic bias.