Applying machine learning for perioperative adverse event prediction: a narrative review toward better clinical efficacy and usability

Abstract

Abstract Early prediction of the major perioperative adverse events is of great significance for reducing mortality, morbidity and medical costs. Machine learning (ML) leverages the capacity for predicting the probability of perioperative adverse events, revealing the promise to facilitate risk stratification, tailored prevention, and individualized perioperative management. However, significant heterogeneity has been demonstrated in the model’s performance of discrimination, calibration, interpretability, and transparency among studies, which raises concerns over their clinical efficacy and usability. A lack of guidance for non-expert medical professionals and stakeholders hinders rigorously conducting research with standard procedure, appropriate methodology, consistent measures, and complete reports. We established a multidisciplinary team consisting of clinicians, data scientists, computer scientists. Multiple libraries including Medline, PubMed, Web of Science, Embase, and CINAHL were searched. We comprehensively summarized critical issues within the entire workflow of ML-based model study, including scenarios and problems, task definition, data collecting and processing, feature representation, model development and validation, clinical implementation and evaluation, aiming to provide guidance and insights for this topic. This review provides a practical checklist of the ML workflow tailored for perioperative teams, bridging technical innovations with clinical translation.

Ähnliche Arbeiten