Predicting the immunological nonresponse to antiretroviral therapy in people living with HIV: a machine learning-based multicenter large-scale study.

Background: Although highly active antiretroviral therapy (HAART) has greatly enhanced the prognosis for people living with HIV (PLWH), some individuals fail to achieve adequate immune reconstitution, known as immunological nonresponse (INR), which is linked to poor prognosis and higher mortality. However, the early prediction and intervention of INR remains challenging in South China. Methods: This study included 1,577 PLWH who underwent at least two years of HAART and clinical follow-up between 2017 and 2022 at two major tertiary hospitals in South China. We utilized logistic multivariate regression to identify independent predictors of INR and employed restricted cubic splines (RCS) for nonlinear analysis. We also developed several machine-learning models, assessing their performance using internal and external datasets to generate receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA). The best-performing model was further interpreted using Shapley additive explanations (SHAP) values. Results: Independent predictors of INR included baseline, 6-month and 12-month CD4+ T cell counts, baseline hemoglobin, and 6-month hemoglobin levels. RCS analysis highlighted significant nonlinear relationships between baseline CD4+ T cells, 12-month CD4+ T cells and baseline hemoglobin with INR. The Random Forest model demonstrated superior predictive accuracy, with ROC areas of 0.866, 0.943, and 0.897 across the datasets. Calibration was robust, with Brier scores of 0.136, 0.102, and 0.126. SHAP values indicated that early CD4+T cell counts and CD4/CD8 ratio were crucial in predicting INR. Conclusions: This study introduces the random forest model to predict incomplete immune reconstitution in PLWH, which can significantly assist clinicians in the early prediction and intervention of INR among PLWH.