• Janno S Schouten, Melissa A C M Kalden, Eris van Twist, ReissIrwin K MIKM0000-0002-9474-6895Department of Neonatal and Pediatric Intensive Care, Division of Neonatology, Erasmus MC, University Medical Center-Sophia Children's Hospital, Rotterdam, The Netherlands., GommersDiederik A M P JDAMPJ0000-0001-6808-7702Datahub, Erasmus MC, University Medical Center, Rotterdam, The Netherlands.Department of Adult Intensive Care, Erasmus MC, University Medical Center, Rotterdam, The Netherlands., Michel E van Genderen, and H Rob Taal.
• Department of Neonatal and Pediatric Intensive Care, Division of Neonatology, Erasmus MC, University Medical Center-Sophia Children's Hospital, Rotterdam, The Netherlands.
• Intensive Care Med. 2024 Nov 1; 50 (11): 176717771767-1777.

PurposeDespite its promise to enhance patient outcomes and support clinical decision making, clinical use of artificial intelligence (AI) models at the bedside remains limited. Translation of advancements in AI research into tangible clinical benefits is necessary to improve neonatal and pediatric care for critically ill patients. This systematic review seeks to assess the maturity of AI models in neonatal and pediatric intensive care unit (NICU and PICU) treatment, and their risk of bias and objectives.MethodsWe conducted a systematic search in Medline ALL, Embase, Web of Science Core Collection, Cochrane Central Register of Controlled Trials, and Google Scholar. Studies using AI models during NICU or PICU stay were eligible for inclusion. Study design, objective, dataset size, level of validation, risk of bias, and technological readiness of the models were extracted.ResultsOut of the 1257 identified studies 262 were included. The majority of studies was conducted in the NICU (66%) and most had a high risk of bias (77%). An insufficient sample size was the main cause for this high risk of bias. No studies were identified that integrated an AI model in routine clinical practice and the majority of the studies remained in the prototyping and model development phase.ConclusionThe majority of AI models remain within the testing and prototyping phase and have a high risk of bias. Bridging the gap between designing and clinical implementation of AI models is needed to warrant safe and trustworthy AI models. Specific guidelines and approaches can help improve clinical outcome with usage of AI.© 2024. The Author(s).

14 keywords...

hide…