Abstract
INTRODUCTION: The use of machine learning (ML) in intensive care units (ICUs) has led to a large yet fragmented body of literature. It is imperative to conduct a systematic analysis and synthesis of this research to identify methodological trends, clinical applications, and knowledge deficits. METHODS: A systematic literature review was conducted in accordance with the PRISMA guidelines, encompassing 2,507 ICU-focused ML publications from 2019 to 2024. Latent Dirichlet Allocation (LDA), an unsupervised topic modeling approach, was used with n-gram and no-n-gram tokenization strategies. Bayesian optimization approaches were used to increase model coherence and diversity. RESULTS: The analysis demonstrated a substantial degree of methodological variability, emphasizing the predominance of studies on infection surveillance and complication prediction. N-gram tokenization efficiently identified clinically specific topics, but no-n-gram techniques produced larger interpretative groups. Underexplored fields include emerging research areas like drug response prediction, pediatric-specific modeling, and surgical risk classification. CONCLUSION: In conclusion, the study highlights the significance of methodological transparency and tokenization strategies while offering a thorough topic overview and identifying methodological trends in the literature on ICU - ML. Future research should prioritize neglected areas such as pediatric care modeling and therapy response, utilizing advanced ML techniques and multimodal data integration to enhance the outcomes of ICU patients.