BACKGROUND: 

Mortality prognostication in adult patients requiring extracorporeal membrane oxygenation (ECMO) is not accurate or established. We hypothesized that composite lactate-based metrics and machine-learning models would improve in-hospital mortality prognostication compared with lactate alone in this population.

STUDY DESIGN: 

We conducted a retrospective study of adult patients supported with ECMO at a Level I trauma center from 2022 to 2024 (N = 104). Composite metrics were generated by multiplying lactate by acid–base markers and comparing performance to lactate after Benjamini-Hochberg correction to control the false discovery rate. For machine-learning models, 3 multilayer perceptron architectures were developed using laboratory markers from a patient’s hospital admission. Feature attribution through local interpretable model-agnostic explanations guided the development of a weight-of-evidence (WoE) model using only initial laboratory values.

RESULTS: 

The composite metric of average arterial lactate × average arterial bicarbonate was the most significantly elevated marker in nonsurvivors and more significant than lactate alone (false discovery rate-adjusted p = 4.49 × 10^–4 vs 2.14 × 10^–3). The 100 × 50 multilayer perceptron architecture achieved 85% accuracy (95% CI 80 to 89%), 87% precision (95% CI 82% to 92%), F1 score of 0.81 (95% CI 0.74 to 0.86), and area under the curve of 0.879 (95% CI 0.84 to 0.92). Local interpretable model-agnostic explanations identified kidney replacement therapy, surrogate respiratory quotient, lactate gradient, and arterial bicarbonate as metrics incorporated in the WoE model achieving 80% accuracy (95% CI 76% to 84%), 70% precision (95% CI 63% to 77%), F1 score of 0.72 (95% CI 0.66 to 0.78), and area under the curve of 0.886 (95% CI 0.853 to 0.920).

CONCLUSIONS: 

Composite lactate metrics and the WoE model improved in-hospital mortality prediction in patients requiring ECMO. Prospective studies and external validation are warranted to confirm these findings.