Efficient mortality prediction for acute respiratory failure: a resource-constrained machine learning approach using MIMIC databases

Accurate prediction of mortality in Acute Respiratory Failure (ARF) patients at intensive care unit (ICU) admission can improve patient outcomes and resource management. However, ICU environments often face challenges like missing test results and limited resources. This study presents a complete pipeline for predicting ARF mortality, focusing on effective feature extraction, data imputation, and class imbalance handling. Key preprocessing steps include iterative imputation for missing data and upsampling techniques like SMOTE and deep learning-based generators. Using the MIMIC-III and MIMIC-IV databases, logistic regression, random forest, extreme gradient boosting, and neural networks were tested. Findings demonstrate that neural networks, along with ensemble methods, achieved high sensitivity and \(\hbox {F}_\beta \) scores, which are essential for accurate mortality predictions. Notably, when class distribution was balanced, the models performed equally well on specificity and sensitivity. SMOTE proved particularly effective in addressing class imbalance, suggesting that advanced upsampling methods like GANs could further enhance prediction accuracy without reducing dataset size.

Graphical abstract

This graphical abstract of the work that illustrates that a patient is admitted to the hospital, admission to the ICU is determined, the test results of the first 24 hours are collected, missing parameters are imputed, the data are normalized and a machine learning model is applied to predict mortality outcomes.

Graphical abstract illustrating the process