Prediction of outcomes in perioperative medicine is key to decision-making and various prediction models have been created to help quantify and communicate those risks to both patients and clinicians. Increasingly, machine learning (ML) is being favoured over more traditional techniques to improve prediction of outcomes, however, the studies are of varying quality. It is also not known whether any increase in predictive performance using ML algorithms transpires into a clinically meaningful benefit. This coupled with the difficulty in interrogating ML algorithms is a potential cause of concern within the medical community. In this review, we provide a concise appraisal of studies which develop perioperative predictive ML models and compare predictive performance to traditional statistical models.

The search strategy, title and abstract screening, and full-text reviews produced 37 studies for data extraction. Initially designed as a systematic review but due to the heterogeneity of the population and outcomes, was written in the narrative.

Perioperative ML and traditional predictive models continue to be developed and published across a range of populations. This review highlights several studies which show that ML can enhance perioperative prediction models, although this is not universal, and performance for both methods remain context dependent. By focusing on relevant patient-centred outcomes, model interpretability, external validation, and maintaining high standards of reporting and methodological transparency, researchers can develop ML models alongside traditional methods to enhance clinical decision-making and improve patient care.