In-Hospital Mortality Prediction of Patients Requiring Extracorporeal Membrane Oxygenation Using Composite Lactate...

Area of Science:

• Critical Care Medicine
• Biomedical Engineering
• Data Science in Healthcare

Background:

• Accurate mortality prognostication for adult patients on extracorporeal membrane oxygenation (ECMO) remains a significant clinical challenge.
• Current methods lack established accuracy, necessitating the development of improved predictive tools.

Purpose of the Study:

• To evaluate the efficacy of composite lactate-based metrics and machine learning (ML) models in enhancing in-hospital mortality prognostication for ECMO patients.
• To compare the predictive performance of these novel approaches against lactate alone.

Main Methods:

• A retrospective study included 104 adult ECMO patients (2022-2024).
• Composite metrics were derived from lactate and acid-base markers; their significance was assessed against lactate alone using Benjamini-Hochberg correction.
• Multilayer perceptron (MLP) models and a weight-of-evidence (WoE) model, guided by LIME feature attribution, were developed using initial laboratory values.

Main Results:

• A composite metric (average arterial lactate × average arterial bicarbonate) was significantly elevated in non-survivors, outperforming lactate alone (FDR-adjusted p = 4.49 x 10-4).
• An MLP model (100x50 architecture) achieved 85% accuracy, 87% precision, and an AUC of 0.879.
• The WoE model, incorporating key lab values, demonstrated 80% accuracy and an AUC of 0.886.

Conclusions:

• Composite lactate metrics and the WoE model significantly improve in-hospital mortality prediction in ECMO patients.
• These findings suggest a promising advancement in critical care prognostication.
• Further prospective studies and external validation are recommended to confirm these results.