Machine learning predicts mortality in septic patients using only routinely available ABG variables: a multi-centre...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Critical Care Medicine

Background:

• Intensive Care Unit (ICU) mortality prediction is crucial for patient management and resource allocation.
• Current methods may not fully capture the dynamic nature of patient status in the ICU.
• Arterial blood gas (ABG) values offer a readily available yet complex data source for physiological assessment.

Purpose of the Study:

• To evaluate Deep Neural Networks (DNN), specifically Long Short-Term Memory (LSTM) models, for predicting ICU mortality.
• To develop a model that predicts mortality within 96 hours, simulating a clinical
• re-triage
• scenario.
• To assess the feasibility of using only ABG values for predictive modeling to enhance real-world applicability.

Main Methods:

• Retrospective analysis of septic patients from the eICU (multi-centre) and MIMIC-III (single-centre) datasets.
• Inclusion of patients alive after 48 hours with available ABG data.
• Development and comparison of predictive models including SOFA score, logistic regression, and an LSTM-based DNN.

Main Results:

• The LSTM-based model achieved superior performance compared to SOFA score and logistic regression.
• The multi-centre study showed an Area Under the Curve (AUC) of 0.88 for the LSTM model.
• The single-centre study demonstrated an AUC of 0.85 for the LSTM model, confirming its robustness.

Conclusions:

• LSTM-based DNN models show significant potential for accurate ICU mortality prediction using ABG data.
• This approach can assist clinicians in patient
• re-triage
• and treatment limitation decisions for patients with poor prognoses.
• Restricting input to ABG values enhances the model's practical utility in clinical settings.