Use of machine learning to analyse routinely collected intensive care unit data: a systematic review

Area of Science:

• Intensive Care Medicine
• Health Informatics
• Machine Learning Applications

Background:

• Intensive care units (ICUs) face significant operational constraints.
• Electronic health records (EHRs) collect extensive patient data.
• Machine learning (ML) offers potential for decision support in ICUs.

Purpose of the Study:

• To systematically review ML applications using routinely collected ICU data.
• To identify common study aims, ML methods, dataset sizes, and validation strategies.
• To assess the predictive accuracy of ML models in ICU settings.

Main Methods:

• Systematic review of Web of Science and MEDLINE databases.
• Exclusion of studies focused on image processing.
• Extraction of study aim, ML type, dataset size, validation, and accuracy measures.

Main Results:

• 258 studies met eligibility criteria; common aims included predicting complications and mortality.
• Median sample size was 488, with 41 studies using >10,000 patients.
• High predictive accuracy (median AUC 0.83-0.94) was reported, but independent validation was rare (6.2%). Neural networks and support vector machines were common ML methods.

Conclusions:

• Publication rates for ML in ICUs are rapidly increasing.
• Many studies utilize insufficient sample sizes for optimal ML performance.
• Methodological and reporting guidelines are crucial for clinical translation and reliable findings.