Machine learning-driven in-hospital mortality prediction in HIV/AIDS patients with Cytomegalovirus infection: a single-centred retrospective study

<b>Introduction.</b> <i>Cytomegalovirus</i> (<i>CMV</i>) is a widely disseminated betaherpesvirus that typically induces latant infections. In immunocompromised populations, especially transplant and HIV-infected patients, <i>CMV</i> infection increases in-hospital mortality.<b>Gap statement.</b> Although machine learning models have been widely used in clinical diagnosis and prognosis prediction, reports on machine learning model predictions for the in-hospital mortality of HIV/AIDS patients with <i>CMV</i> infection have not been reported.<b>Aim.</b> Analyze the general gemographic and clinical characteristics of HIV/AIDS patients with <i>CMV</i> infection and identify the factors affecting the prognosis of this population, which will help to reduce their in-hospital mortality.<b>Methods.</b> Hospitalized HIV/AIDS patients with <i>CMV</i> infection were recruited from the Fourth People's Hospital of Nanning, Guangxi, from 2012 to 2019. After dividing them into survival and death groups based on their in-hospital survival status, their general and clinical profiles were described. Following 1 : 3 propensity score matching to equalize baseline characteristics, three machine-learning models (Random Forest, Support Vector Machine and eXtreme Gradient Boosting) were deployed to forecast factors influencing prognosis. The SHapley Additive exPlanations tool explained the models.<b>Results.</b> A total of 1102 HIV/AIDS patients with <i>CMV</i> infection were analysed. There was no statistical difference in the general condition of the study subjects (<i>P</i>＞0.05). Prevalent complications/coinfections included pneumonia (63.6%), <i>tuberculosis</i> (47.2%) and oral fungal infections (44.6%). There were significant differences between the groups in pneumonia, <i>cryptococcosis</i> and hypoproteinaemia (<i>P</i><0.05). The differences in laboratory indicators between patients were also statistically significant (<i>P</i><0.05). The three machine learning models demonstrated good performance, identifying primary predictors of mortality. Pneumonia, urea, indirect bilirubin and platelet distribution width exhibited positive associations with death, with higher levels correlating with an increased mortality risk. Conversely, CD4 T-cell count, CD8 T-cell count and platelet displayed negative correlations with mortality.<b>Conclusions.</b> HIV/AIDS patients with CMV infection exhibit distinctive clinical features impacting survival outcomes. Machine learning models accurately identify key influencing factors and predict mortality risk in this population, which appears to be essential to reducing in-hospital mortality.