Machine and Deep Learning Models for Hypoxemia Severity Triage in CBRNE Emergencies.

Area of Science:

• Medical informatics
• Computational biology
• Artificial intelligence in healthcare

Background:

• Hypoxemia poses a significant risk during emergency triage, especially in Chemical, Biological, Radiological, Nuclear, and Explosive (CBRNE) events.
• Accurate and timely prediction of hypoxemia severity is crucial for effective patient management and resource allocation.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting hypoxemia severity during emergency triage.
• To compare the performance of tree-based models (TBMs) and sequential models (LSTMs, GRUs) for real-time hypoxemia prediction.
• To identify key physiological variables influencing hypoxemia severity prediction.

Main Methods:

• TBMs (XGBoost, LightGBM, CatBoost, RFs) and sequential models (LSTM, GRU) were trained on MIMIC-III and IV datasets.
• A preprocessing pipeline handled missing data, class imbalances, and synthetic data.
• Models were evaluated using a 5-minute prediction window with minute-level interpolations.

Main Results:

• TBMs demonstrated superior speed, interpretability, and reliability compared to sequential models for real-time applications.
• Feature importance analysis highlighted six key physiological variables and the significance of mask and score features.
• Voting Classifier ensembles offered minor metric improvements but did not surpass individually optimized TBMs.

Conclusions:

• TBMs are effective for real-time hypoxemia prediction in emergency triage.
• Sequential models, while capable of temporal analysis, are computationally intensive.
• ML holds significant potential for enhancing triage systems and mitigating alarm fatigue.