Machine Learning-Based Prognostic Modeling and BAR Trajectory Subphenotyping in Critically Ill Patients with Viral...

Viral infection is a major cause of severe community-acquired pneumonia, yet existing risk stratification tools may lack timeliness and precision. This study aimed to develop machine learning models to predict 28-day mortality and new initiation of invasive mechanical ventilation in critically ill patients with viral pneumonia.

This retrospective study used the MIMIC-IV (v3.1) database, with external validation of the mortality model from Huashan Hospital. Predictors were selected using least absolute shrinkage and selection operator regression. Logistic regression, ridge regression, support vector machine, random forest, and XGBoost models were developed and compared. Blood urea nitrogen and albumin were further analyzed to construct the BUN-to-albumin ratio (BAR). Latent class mixed modeling was applied to identify BAR trajectories and assess their prognostic value.

A total of 1,285 patients were included, with 28-day mortality of 23.3%. XGBoost demonstrated the best discriminative performance. The 9-predictor mortality model achieved AUCs of 0.870 (95% CI: 0.844-0.896) in the training set, 0.750 (0.695-0.805) in the internal validation set, and 0.762 (0.701-0.823) in the external validation cohort. The 7-variable ventilation model yielded AUCs of 0.861 (0.837-0.886) and 0.745 (0.693-0.797) in the training and internal validation sets, respectively. An interactive web-based tool was developed to support individualized risk estimation. Three BAR trajectory classes were identified, and an increasing BAR trajectory was independently associated with higher 28-day mortality.

Routinely available variables allowed individualized prediction of mortality and new invasive ventilation risk. Dynamic BAR trajectories provided additional prognostic information and may help refine risk stratification. Further external validation and recalibration are warranted before clinical implementation.