6228 AiSEPTRON study: designing a paediatric sepsis prediction tool using machine learning

Objectives Background: Sepsis is a life-threatening and preventable condition that kills children worldwide.1 Predictive models developed from electronic health records (EHR), using artificial intelligence (AI) demonstrated high precision to detect sepsis2 and predict clinical outcomes3 in children attending the emergency department (ED). Objectives We report the performance of ML models to predict the outcomes of serious infection and sepsis in children attending ED. Methods Strategy: Develop a multi-disciplinary AI team, infrastructure and platforms for the extraction, storage and processing of a wide range of EHR variables.4 Study Design: The study was designed as a retrospective, observational study. Participants & Settings: Consecutive children under the age of 16 years that attended ED at the St Thomas’s Hospital, between 1 January 2018 to 31 Dec 2019 were included. Children presenting with trauma/minor injuries were excluded. PREDICTOR VARIABLES: The study used de-identified, routinely collected EHRs. PRIMARY OUTCOME: Patients admitted for intravenous antibiotics > = 48 hours. Figure 1 shows the predictor and outcome variables in detail. Model Building: A supervised ML approach was utilised to create the predictive models. The input data, together with their corresponding outcome labels, were used to train classification models using XGBoost, neural networks and random forests. We incorporated free-text notes from triage using Natural Language Processing using a method called transfer learning. Results A total of 35,414 patient attendances were included in the study. Triage Model: AUC of 0.80 (95% CI 0.77 to 0.83) for the test dataset to predict the need for intravenous antibiotics at triage. Blood Test Model 1: AUC of 0.78 (95% CI 0.74 to 0.82) to predict the need for intravenous antibiotics of any duration. Blood Test Model 2: AUC of 0.73 (95% CI 0.68 to 0.78) to predict the need for admission with intravenous antibiotics > = 48 hours. Figure 2 shows the performance of the 3 main models. Conclusion We pioneered the development of prediction models for sepsis outcomes, applying machine learning and natural language processing to EHRs, for the first time in the UK. We established the infrastructure and platforms for the extraction, storage and processing of multi-centre data, including a collaboration of a strong team of clinical experts, research nurses, data analysts, data scientists and statisticians. We demonstrated high precision in test cohorts to predict the need for intravenous antibiotics from triage, and moderate precision for intravenous antibiotics after blood tests. Further refinement and prospective evaluation is being planned. References Fleischmann-Struzek C, et al, Lancet Respir. Med, 2018. Le S, et al, Front. Pediatr., 2019. Goto T et al, JAMA Netw Open, 2019. AiSEPTRON Study: Designing a Paediatric Sepsis Prediction Tool using Machine Learning, 2023. Available at: http://aiseptron.co.uk (Accessed 01/11/2023).

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.