Machine learning algorithms to predict the risk of admission to intensive care units in HIV-infected individuals: a single-centre study

Abstract

Antiretroviral therapy (ART) has transformed HIV from a rapidly progressive and fatal disease to a chronic disease with limited impact on life expectancy. However, people living with HIV(PLWHs) faced high critical illness risk due to the increased prevalence of various comorbidities and are admitted to the Intensive Care Unit(ICU). This study aimed to use machine learning to predict ICU admission risk in PLWHs. 1530 HIV patients (199 admitted to ICU) from Beijing Ditan Hospital, Capital Medical University were enrolled in the study. Classification models were built based on logistic regression(LOG), random forest (RF), k-nearest neighbor (KNN), support vector machine (SVM), artificial neural network(ANN), and extreme gradient boosting(XGB). The risk of ICU admission was predicted using the Brier score, area under the receiver operating characteristic curve (ROC-AUC), and area under the precision-recall curve(PR-ROC) for internal validation and ranked by Shapley plot. The ANN model performed best in internal validation (Brier score = 0.034, ROC-AUC = 0.961, PR-AUC = 0.895) to predict the risk of ICU admission for PLWHs. 11 important features were identified to predict predict ICU admission risk by the Shapley plot: respiratory failure, multiple opportunistic infections in the respiratory system, AIDS defining cancers, baseline viral load, PCP, baseline CD4 cell count, and unexplained infections. An intelligent healthcare prediction system could be developed based on the medical records of PLWHs, and the ANN model performed best in effectively predicting the risk of ICU admission, which helped physicians make timely clinical interventions, alleviate patients suffering, and reduce healthcare cost.

Keywords: Admission; Artificial neural network (ANN); HIV; HIV-related comorbidities; ICU; Machine learning; Predictive modeling; Risk factors; SHAP analysis.

Fig. 1

Study flow diagram. Note: ICU: intensive care unit, LASSO: least absolute shrinkage and selection operator, ROC: receiver operating characteristic curve, PR: precision-recall

Fig. 2

LASSO feature selection. A. Optimal λ determination via 5-fold cross-validation. B. Coefficient profiles of 11 features showing regularization effects. LASSO: Least absolute shrinkage and selection operator

Fig. 3

Model performance metrics. A. AUROC (Area under ROC curve) B. AUPRC (Area under precision-recall curve) RF: Random forest; KNN: k-nearest neighbors; SVM: Support vector machine; ANN: Artificial neural network; XGB: XGBoost; LOG: Logistic regression

Fig. 4

SHAP summary plot for the ANN model A. Feature importance ranking (top: high impact) with value-direction coding: Color scale: High (red)/Low (blue) feature values; Horizontal position: Risk decrease (left) / increase (right). B.Characteristic contribution of a single sample. ANN: Artificial neural network; SHAP: Shapley additive explanations