(1437) An Extracorporeal Life Support (ECLS) Predictive Mortality Model for PICU Patients

Introduction: Several models have been developed to predict ECLS patient mortality (Bailly 2019; Hilder 2017; Shah 2022; Geisser 2022). Despite high discriminative power in certain studies, data is limited by single-site, small cohort, or restricted registry covariates. Moreover, standard PICU models like PIM3 and PRISM III fall short in predictive performance in ECLS patients. This study develops a model to predict mortality in the ECLS cohort using generalized covariates and a large multi-centered PICU dataset

Methods: 11,046 PICU cases requiring ECLS support between 2009 and 2024 from 131 units were obtained from the Virtual Pediatric Systems registry. Cases that did not collect PRISM III or had incomplete data were excluded. Data was divided into independent training and testing sets with a 75/25 split. Four logistic regression models were developed. Two models used PIM3 and PRISM III variables plus clinically relevant fields: ECLS cannulation type (venoarterial, venovenous), time from admission to ECLS cannulation, extracorporeal CPR, and renal replacement therapy. Model 1 used all 26 variables; Model 2 used forward selection with Akaike information criterion (AIC), resulting in 19 variables. Two models used PRISM III laboratory fields plus the clinically relevant fields. Model 3 used all 16 variables; Model 4 used forward selection with AIC, resulting in 8 variables. Model performance was assessed on the testing set with area under the receiver operating characteristic curve (AUROC), Hosmer-Lemeshow test, and GiVITI calibration belt.

Results: Model 2 was selected for further analysis due to its well-rounded discrimination, calibration, and parsimonious properties. The training and testing sets had mortality rates of 40.3% (3,235 cases) and 41.1% (1,100 cases), respectively. Model 2 had the second highest AUROC of 71.7%. The standardized mortality ratio was closest to one at 1.01. The Hosmer-Lemeshow test yielded a p-value = 0.33. The GiViTI calibration belt yielded a p-value = 0.09, suggesting good calibration. Notably, Model 2 outperformed PIM3 (59.1%) and PRISM III (58.4%) AUROC.

Conclusions: Our model matches the performance of existing ECLS mortality models and outperforms PIM3 and PRISM III in the ECLS cohort. However, our model has widespread applicability as it uses more universal covariates.