Development, External Validation, and Biomolecular Corroboration of Interoperable Models for Identifying Critically Ill Children at Risk of Neurologic Morbidity

ImportanceDeclining mortality in the field of pediatric critical care medicine has shifted practicing clinicians attention to preserving patients neurodevelopmental potential as a main objective. Earlier identification of critically ill children at risk for incurring neurologic morbidity would facilitate heightened surveillance that could lead to timelier clinical detection, earlier interventions, and preserved neurodevelopmental trajectory. ObjectiveDevelop machine-learning models for identifying acquired neurologic morbidity while hospitalized with critical illness and assess correlation with contemporary serum-based, brain injury-derived biomarkers. DesignRetrospective cohort study. SettingTwo large, quaternary childrens hospitals. ExposuresCritical illness. Main Outcomes and MeasuresThe outcome was neurologic morbidity, defined according to a computable, composite definition at the development site or an order for neurocritical care consultation at the validation site. Models were developed using varying time windows for temporal feature engineering and varying censored time horizons prior to identified neurologic morbidity. Optimal models were selected based on F1 scores, cohort sizes, calibration, and data availability for eventual deployment. A generalizable created at the development site was assessed at an external validation site and optimized with spline recalibration. Correlation was assessed between development site model predictions and measurements of brain biomarkers from a convenience cohort. ResultsAfter exclusions there were 14,222-25,171 encounters from 2010-2022 in the development site cohorts and 6,280-6,373 from 2018-2021 in the validation site cohort. At the development site, an extreme gradient boosted model (XGBoost) with a 12-hour time horizon and 48-hour feature engineering window had an F1-score of 0.54, area under the receiver operating characteristics curve (AUROC) of 0.82, and a number needed to alert (NNA) of 2. A generalizable XGBoost model with a 24-hour time horizon and 48-hour feature engineering window demonstrated an F1-score of 0.37, AUROC of 0.81, AUPRC of 0.51, and NNA of 4 at the validation site. After recalibration at the validation site, the Brier score was 0.04. Serum levels of the brain injury biomarker glial fibrillary acidic protein measurements significantly correlated with model output (rs=0.34; P=0.007). Conclusions and RelevanceWe demonstrate a well-performing ensemble of models for predicting neurologic morbidity in children with biomolecular corroboration. Prospective assessment and refinement of biomarker-coupled risk models in pediatric critical illness is warranted. Key PointsQuestion Can interoperable models for predicting neurological deterioration in critically ill children be developed, correlated with serum-based brain-derived biomarkers, and validated at an external site? Findings A development site model demonstrated an area under the receiver operating characteristics curve (AUROC) of 0.82 and a number needed to alert (NNA) of 2. Predictions correlated with levels of glial fibrillary acidic protein in a subset of children. A generalizable model demonstrated an AUROC of 0.81 and NNA of 4 at the validation site. Meaning Well performing prediction models coupled with brain biomarkers may help to identify critically ill children at risk for acquired neurological morbidity.