Prediction of new HIV infection in men who have sex with men based on machine learning: secondary analysis of a prospective cohort study from Western China

This study aimed to construct a model based on machine learning to predict new HIV infections in HIV-negative men who have sex with men (MSM). This is a secondary analysis of a previous random clinical trial aiming to evaluate the preventive effects of PrEP on new HIV infection in MSM. During 2013–2015, 1455 HIV-negative MSM were enrolled. Participants were divided into treatment group and control group and regularly followed up until they seroconverted to HIV positive or until the 2-year endpoint reached. Five machine-learning approaches were applied to predict the risk of HIV infection. Model performance was evaluated using Harrel’s C-index and area under the receiver operator characteristic curve (AUC) and validated in an external validation cohort. To explain this model, shapley additive explanation (SHAP) values were calculated and visualized. During the observation period, 102 MSM developed HIV infection. Thirteen parameters are selected to construct the model. The random survival forest model showed the best performance in the validation cohort, with a C-index of 0.7013, and could significantly categorize MSM into three groups. Our model indicated that MSM with younger age, receptive anal intercourse, and multiple male sexual partners had an increased risk of HIV infection, and those with higher AIDS knowledge scores had a lower risk. We presented a machine learning-based model to predict their risk of developing HIV infection. This model could be applied to recognize MSM who are at a higher risk of developing HIV infection.

本研究旨在构建基于机器学习的模型，用于预测男男性行为者（men who have sex with men, MSM）中HIV阴性人群的新发HIV感染风险。本研究为一项既往随机临床试验的二次分析，该试验旨在评估暴露前预防（PrEP）对男男性行为者新发HIV感染的预防效果。2013年至2015年间，共纳入1455名HIV阴性男男性行为者。研究对象被随机分为治疗组与对照组，并接受定期随访，直至其HIV血清学转换为阳性或达到2年随访终点。本研究采用五种机器学习方法构建HIV感染风险预测模型，模型性能通过哈雷尔C指数（Harrell’s C-index）与受试者工作特征曲线下面积（area under the receiver operator characteristic curve, AUC）进行评估，并在外部验证队列中完成验证。为阐释模型的可解释性，本研究计算并可视化展示了夏普利可加性解释（SHAP）值。在观察期内，共有102名男男性行为者发生HIV感染。研究最终筛选出13项参数用于构建该模型。随机生存森林模型在验证队列中表现最优，其C指数达0.7013，且可将男男性行为者显著划分为三个风险组别。本研究模型显示，年龄较轻、存在接受式肛交行为以及拥有多个男性性伴侣的男男性行为者HIV感染风险升高，而艾滋病知识得分较高者感染风险则相对较低。本研究提出了一种基于机器学习的HIV感染风险预测模型，该模型可用于识别HIV感染风险较高的男男性行为者群体。