Machine learning quantifies immuno-virological interactions: a TRIPOD+AI compliant prediction model for HIV-1 salvage therapy outcomes
Background The management of multidrug-resistant HIV-1 in patients experiencing virologic failure remains a critical clinical challenge. Traditional linear scoring systems often fail to adequately capture the complex evolutionary dynamics between the virus, host immunity, and antiretroviral regimens. This study aims to construct and validate a prediction model in line with the TRIPOD+AI statement to quantify the multidimensional, nonlinear interactions among “virus-host-drug” and to guide individualized clinical salvage therapy. Methods This retrospective cohort study integrated de-identified data from 18 clinical trials in the Stanford HIV Drug Resistance Database (n = 6,844). Seven machine learning algorithms were compared with traditional models. Model evaluation metrics included area under the receiver operating characteristic curve (AUC), Brier score, and calibration curves. TreeSHAP quantified feature contributions and interactions. Ablation studies (DeLong’s test) evaluated the incremental predictive value of integrating virological, immunological, and treatment history domains. The algorithmic fairness of the model across populations with different immune statuses and viral loads was evaluated through subgroup analysis, and the Effective Sample Size (ESS) was introduced to assess individual prediction uncertainty. Results The XGBoost model best predicted 24-week virologic suppression (AUC: 0.887), significantly outperforming the baseline model (AUC: 0.816) with excellent calibration (Brier score: 0.096). Ablation studies confirmed that the integrated model significantly outperformed partial models restricted to single feature domains (all P < 0.05). SHAP interaction analysis revealed a significant modification effect of baseline CD4+ T cell count on the predictive weight of viral load; meanwhile, a temporal decay in drug resistance test results was observed, significantly diminishing the negative predictive weight of a heavy treatment history. Reclassification analysis showed that the XGBoost model corrected 61.90% of actual failures misclassified by the baseline model, demonstrating a significant net clinical benefit (Net Reclassification Improvement: 0.490). Clinical fairness checks confirmed that the model performed stably in subgroups with severe immune compromise and high viral loads, without showing systematic bias. Conclusion The developed XGBoost model overcomes the limitations of traditional linear scoring and achieves precise prediction of HIV salvage therapy outcomes by quantifying immune modulatory effects and therapeutic exhaustion markers. This model acts as a clinical safeguard to identify ineffective treatments while maintaining algorithmic fairness across patient severities. The developed web-based calculator and risk stratification system help clinicians optimize resource allocation and advance novel drug use in complex resistance scenarios, promoting evidence-based HIV precision medicine practices.