Machine Learning-Based Identification of Risk Factors for ICU Mortality in 8902 Critically Ill Patients with Pandemic...

Area of Science:

• Critical Care Medicine
• Infectious Diseases
• Data Science in Healthcare

Background:

• The COVID-19 and influenza A (H1N1)pdm09 pandemics caused significant ICU admissions and mortality.
• Identifying early risk factors for ICU mortality is crucial for optimizing clinical management.
• Different analytical models may yield varying risk factor profiles.

Purpose of the Study:

• To compare the risk factors and predictive performance of a linear model (generalized linear model, GLM) versus a non-linear model (random forest, RF).
• To analyze a large national cohort of critically ill patients with SARS-CoV-2 or influenza A (H1N1)pdm09.

Main Methods:

• Retrospective analysis of 8902 critically ill patients from 184 Spanish ICUs.
• Data included demographics, clinical, laboratory, and microbiological information from the first 24 hours of admission.
• Prediction models (GLM and RF) were developed and their performance evaluated using AUC, precision, sensitivity, specificity, OOB error, and accuracy.

Main Results:

• Overall ICU mortality was 25.8%.
• Both GLM (AUC 76%) and RF (AUC 75.6%) demonstrated similar predictive performance.
• GLM identified 17 risk factors, while RF identified 19 (accuracy reduction) to 23 (Gini index reduction).
• Key differences included laboratory markers (procalcitonin, WBC, lactate, D-dimer) significant in RF but not GLM, and acute kidney injury and *Acinetobacter* spp. significant in GLM but not RF.

Conclusions:

• Linear and non-linear models showed comparable performance in predicting ICU mortality.
• However, the identified risk factors differed significantly between the GLM and RF models.
• Clinicians should be aware of the limitations and benefits of studies relying on a single modeling approach.