O1017: Identifying Pediatric Patients at Risk of Developing Severe Complications from sars-cov-2 Infection – Harnessing the Power of Artificial Intelligence-driven Predictive Models

Background: At initial evaluation in the Emergency Department (ED), it is not known which children infected with SARS-CoV-2 might be at risk of severe outcomes. Our objective was to derive and validate models for predicting the development of severe complications of SARS-CoV-2 in children in the ED using electronic health record data and artificial intelligence-based machine learning methods.

Methods: This was an observational study of one academic and two community hospitals. ED encounters from patients aged 0-21 years were included if patients underwent SARS-CoV-2 testing between March 2020 and December 2022. Primary outcomes were: need for general inpatient care (cardiovascular or respiratory dysfunction, or readmission within 72 hours) or critical care (intensive care admission, need for vasopressors, need for ventilatory support, or death). Predictor variables were chosen based on existing literature and univariable analyses. Models were derived using predictors available at various time points during the ED visit: at triage (A), once labs were obtained (B), once any respiratory support was given (C), and at ED disposition (D). Models were trained and internally validated on electronic health record data from the academic hospital and externally validated on data from each community hospital. Data were analyzed using descriptive statistics and random forest supervised machine learning (ML) analyses.

Results: For the 42490 ED visits included, median age was 4.0 years (interquartile range [IQR] 1.0 – 11.0), females accounted for 48.6% of visits, 45.5% of patients were Black, 23.0% White, and 18.9% were Hispanic. The most common chief complaints were fever (26.2%) and respiratory symptoms (26.9%). A need for inpatient care occurred in 5181 visits (12.2%) and for critical care in 1758 visits (4.1%). The area under the receiver operating characteristic curve (AUC) for time points A, B, C and D were 0.84 (95% confidence interval [CI] 0.83 – 0.85), 0.85 (95%CI 0.84 – 0.85), 0.85 (95%CI 0.84 – 0.86), and 0.85 (95%CI 0.84 – 0.86), respectively for predicting need for inpatient care, and 0.87 (95%CI 0.86 – 0.88), 0.88 (95%CI 0.87 – 0.89), 0.88 (95%CI 0.87 – 0.89), and 0.88 (95%CI 0.87 – 0.89) respectively for predicting need for critical care in the internal validation dataset. Figure 1 depicts corresponding AUCs for the external validations at each community hospital.

Conclusion: Machine learning models using data obtained during an ED visit accurately predicted need for inpatient and critical care in children with SARS-CoV-2. Models using variables available at triage (demographics, past medical history, chief complaint, and triage vital signs) performed as well as those adding data collected during the rest of the ED encounter. Similar methods could be applied to other disease processes to accurately predict risk of decompensation at presentation to the ED.