Predicting sepsis treatment decisions in the paediatric emergency department using machine learning: the AiSEPTRON study.

BACKGROUND

Early identification of children at risk of sepsis in emergency departments (EDs) is crucial for timely treatment and improved outcomes. Existing risk scores and criteria for paediatric sepsis are not well-suited for early diagnosis in ED.

OBJECTIVE

To develop and evaluate machine learning models to predict clinical interventions and patient outcomes in children with suspected sepsis.

DESIGN

Retrospective observational study.

SETTING

ED of a tertiary care hospital, UK.

PATIENTS: Electronic health records of children <16 years of age attending between 1 January 2018 and 31 December 2019. Patients presenting with minor injuries were excluded.

METHODS

Prediction models were developed and validated, using 15 key predictors from triage and post-blood test data. XGBoost, the best-performing machine learning model, integrated these predictors with triage note information extracted via Natural Language Processing.

MAIN OUTCOMES: (1) Administration of antibiotics; (2) critical care: antibiotics with fluid resuscitation above 20 mL/kg or non-elective mechanical ventilation; (3) serious infection: hospital admission for antibiotics >48 hours.Model performance was evaluated using area under the receiver operating characteristic curve (AUC), likelihood ratios and positive and negative predictive values.

RESULTS

Triage model: predicted antibiotics at triage (n=35 795; 3.2% with outcome) with an AUC of 0.80 (95% CI 0.76 to 0.84).Antibiotic model: predicted antibiotics post-blood tests (n=4700; 24.2%) with an AUC of 0.78 (95% CI 0.73 to 0.81).Critical care model: predicted critical care (n=4700; 3.3%) with an AUC of 0.78 (95% CI 0.72 to 084).Serious infection model: predicted serious infection (n=4700; 9.4%) with an AUC of 0.76 (95% CI 0.71 to 0.81).Key predictors included triage category, temperature, capillary refill time and C reactive protein.

CONCLUSION

Machine learning models demonstrated good accuracy in predicting antibiotic use following triage and moderate accuracy for critical care and serious infection. Further development and external validation are ongoing.