Predicting immune reconstitution after antiretroviral therapy in HIV/AIDS using ensemble machine learning: a real-world study.

BACKGROUND

Accurate prediction of long-term CD4+ T-cell recovery trajectories in people living with HIV on antiretroviral therapy (ART) is a crucial unmet need for personalized monitoring and treatment optimization. Traditional statistical models have limited ability to capture the complex, non-linear relationships inherent in longitudinal clinical data.

METHODS

We developed a heterogeneous stacking ensemble framework to predict longitudinal CD4+ count, CD8+ count, and CD4/CD8 ratio. The model integrates four tree-based algorithms-XGBoost, LightGBM, Random Forest, and Gradient Boosting-with a Ridge regression meta-learner. It was trained and tested on a retrospective cohort of 5,436 patients who initiated ART between 2016 and 2025, using only demographic and clinical features while explicitly excluding baseline CD4+/CD8+ counts to prevent data leakage.

RESULTS

On an independent test set (n=1,088), the ensemble achieved an R2 of 0.768 (MAE: 74.8 cells/μL) for CD4+ count, 0.636 (MAE: 300.5 cells/μL) for CD8+ count, and 0.131 (MAE: 0.137) for the CD4/CD8 ratio. This represents a relative improvement in R2 of 66.4% for CD4+ and 128.6% for CD8+ predictions compared to a baseline Robust Transformer model. The model accurately replicated the statistical distributions of observed outcomes and demonstrated stable learning dynamics without overfitting.

CONCLUSION

Our ensemble learning framework provides a robust and clinically applicable tool for forecasting multi-dimensional immune reconstitution in HIV care. By synthesizing diverse algorithmic perspectives without relying on baseline immunology, it offers a foundation for data-driven clinical decision support to personalize long-term treatment monitoring.