Machine learning-based prediction of mortality risk in AIDS patients with comorbid common AIDS-related diseases or symptoms

TL;DR

This study develops a machine learning model to predict mortality risk in AIDS patients with related diseases or symptoms, aiming to help with early clinical intervention.

Contribution

The novel contribution is an XGBoost-based model optimized for predicting mortality in AIDS patients with comorbid conditions.

Findings

The optimal model achieved an AUC of 0.832 in the training set and 0.873 in the external validation set.

Key predictors included hemoglobin, infection pathway, and Pneumocystis jirovecii pneumonia.

The model demonstrated high clinical utility through decision-curve analyses and calibration curves.

Abstract

Early assessment and intervention of Acquired Immune Deficiency Syndrome (AIDS) patients at high risk of mortality is critical. This study aims to develop an optimally performing mortality risk prediction model for AIDS patients with comorbid AIDS-related diseases or symptoms to facilitate early intervention.

The study included 478 first-time hospital-admitted AIDS patients with related diseases or symptoms. Eight predictors were screened using lasso regression, followed by building eight models and using SHAP values (Shapley’s additive explanatory values) to identify key features in the best models. The accuracy and discriminatory power of model predictions were assessed using variable importance plots, receiver operating characteristic curves, calibration curves, and confusion matrices. Clinical benefits were evaluated through decision-curve analyses, and validation was performed with an external set of 48 patients.

Lasso regression identified eight predictors, including hemoglobin, infection pathway, Sulfamethoxazole-Trimethoprim, expectoration, headache, persistent diarrhea, Pneumocystis jirovecii pneumonia, and bacterial pneumonia. The optimal model, XGBoost, yielded an Area Under Curve (AUC) of 0.832, a sensitivity of 0.703, and a specificity of 0.799 in the training set. In the test set, the AUC was 0.729, the sensitivity was 0.717, and the specificity was 0.636. In the external validation set, the AUC was 0.873, the sensitivity was 0.852, and the specificity was 0.762. Furthermore, the calibration curves showed a high degree of fit, and the DCA curves demonstrated the overall high clinical utility of the model.

In this study, an XGBoost-based mortality risk prediction model is proposed, which can effectively predict the mortality risk of patients with co-morbid AIDS-related diseases or symptomatic AIDS, providing a new reference for clinical decision-making.

Figures