Reliable Identification of Paediatric Sepsis Onset from EHR Data for AI Model Development

This poster, presented at ESPID 2026 (The European Annual Meeting of the European Society for Paediatric Infectious Diseases) showcases research carried our within the PHEMS project on the reliable identification of paediatric sepsis onset using routinely collected Electronic Health Record (EHR) data. The study evaluates different operational definitions of sepsis to identify the most robust approach for generating accurately labelled datasets for artificial intelligence (AI) model development. Using retrospective data from PICU, the analysis compares established sepsis frameworks, including Goldstein, Phoenix, and IPSO criteria, considering timing of detection, data completeness, and clinical performance. The findings support the combination of an antibiotics-plus-cultures definition for suspected infection with the Goldstein sepsis criteria as the most reliable operational framework for identifying sepsis onset. This approach provides a strong foundation for developing AI-based prediction models that can enable earlier recognition of paediatric sepsis and improve clinical decision-making. The poster contributes to the PHEMS mission of advancing federated health data research and trustworthy AI for paediatric healthcare across Europe.