Machine learning in anesthesiology: Detecting adverse events in clinical practice

The credibility of threshold-based alarms in anesthesia monitors is low and most of the warnings they produce are not informative. This study aims to show that Machine Learning techniques have a potential to generate meaningful alarms during general anesthesia without putting constraints on the type of procedure. Two distinct approaches were tested - Complication Detection and Anomaly Detection. The former is a generic supervised learning problem and for this a simple feed-forward Neural Network performed best. For the latter, we used an Encoder-Decoder Long Short-Term Memory architecture that does not require a large manually-labeled dataset. We show this approach to be more flexible and in the spirit of Explainable Artificial Intelligence, offering greater potential for future improvement.