Related Experiment Video
Updated: Jan 7, 2026
Prediction of HIV-1 Coreceptor Usage Tropism by Sequence Analysis using a Genotypic Approach
Published on: December 1, 2011
Early prediction of immunological non-responders in people living with HIV using machine learning: Model development
Rui Sun1, Binyu Gao2, Hanxi Zhang1
1Beijing Ditan Hospital, Capital Medical University, Beijing 100015, China.
Background:
Despite antiretroviral therapy (ART), 10-40 % of people living with HIV (PLWH) fail to normalize CD4+ T cells, known as immune non-responders (INR), associated with poor clinical outcomes. Due to the complex pathogenesis and the absence of effective treatments, early prediction and intervention of INR are critical. With the rapid advancements in artificial intelligence, particularly in machine learning (ML), developing an interpretable ML model to identify individuals at high risk of INR can facilitate personalized treatment strategies and improve clinical management.
Methods:
This retrospective study was conducted from a long-term multicenter cohort involving 30938 PLWH attending three hospitals from January 2003 to December 2023. Seven ML algorithms were employed to construct prediction models. The area under the receiver operating characteristic curve (AUC), precision-recall curves, calibration plots, clinical impact curves, and decision curve analysis were used to evaluate and identify the optimal model. We evaluated the final model using internal cross-validation and validated it in an external cohort. The Python-based Streamlit framework was applied to develop a web platform.
Results:
11287 PLWH, including 1325 (11.74 %) INR, were analyzed in the derivation cohort. Twenty baseline clinical indicators of PLWH were used to construct ML models. After feature reduction and comparative evaluation, the 9-feature RF model demonstrated strong discrimination (AUC = 0.864), superior calibration, and favorable clinical utility, outperforming the traditional model (AUC = 0.856) and other models. The model performance was also confirmed in internal validation (mean AUC = 0.884 ± 0.003) and the external validation (AUC = 0.855).
Conclusions:
In this study, an interpretable ML model with nine baseline clinical indicators was constructed for early INR prediction in ART-naïve PLWH and was incorporated into a convenient web platform to facilitate individualized treatment and management.
