Machine learning models in anaesthesiology: bridging the gap from model training to implementation.

Abstract

Very few anaesthesiology-related machine learning models have successfully made the transition from retrospective validation to prospective implementation. A new publication in the British Journal of Anaesthesia describes the single-centre implementation of a mortality prediction model to be used by a float anaesthesiologist to trigger enhanced preoperative evaluation before add-on cases. Key adaptations and decisions during model implementation included selecting an appropriate decision threshold to trigger float anaesthesiologist review, retrieving fresh data for model input features only every 6 h, and reducing the number of input features used by the model. This implementation provides an excellent case study illustrating how a machine learning model must be tailored during deployment to meet the needs of the specific use case.