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

Comprehensive Summary

This article dives into using machine learning to predict post-operative mortality in patients, specifically add on cases that are new for the float anesthesiologist. The goal of having a machine learning model would be to have it alert for urgent cases, with the goal to minimize false positives. To do this, a higher decision threshold can be chosen, reducing the number of false positives but conversely potentially missing moderate risk patients. The decision threshold was set at 7%, which was determined to be the optimal threshold where intervention, a review of plans and a discussion of the patient, was indicated while avoiding cancellation of the surgery. Another benefit is using the information of the model. While a model accurately predicted incidence of mortality, it is determined to be as accurate as a physician who knows the patients and their associated risks and would assign an ASA category, indicating it may not be more accurate, but it may be useful in physicians who do not know their patient. In addition to ASA categorization, the type of surgery also plays a key role in determining the risk of the patient. To address this, the model used a colored system based on key measurements shown to lead to the risk associated with the surgery, rather than solely the surgery itself. To reduce burden of computational modeling, data was retrieved every 6 hours, an interval deemed long enough to gather pertinent information if a surgery is next day, but other information such as vitals would need to be gathered and imported into this model to give an accurate risk score. To incorporate this model into hospital systems at a low cost, some input features were reduced or deleted to account for the way the data was presented to the model from the charts, indicating that the model in the previous study that was tested would not be the exact model in use in a hospital system. This article was a continuation of a previous published study where a learning model was used to identify the risks of mortality in post-operative patients. It outlines the transition from a model using data from a retrospective study to live data in a hospital setting and outlines the strengths and weaknesses and how it can be used to potentially benefit physicians using it.

Outcomes and Implications

This article outlines the transition from using retrospective data for an output for machine learning to now incorporating the model into a hospital system with live fluctuating data. It outlines the strengths and weaknesses in the model transitioning process and can be used for future models to improve the designs initially, so when the model is transitioned into a hospital system, few design changes need to be made. This ultimately results in the models being studied from retrospective data and used in published articles to be closer to the real-life models that will be used in data and will allow for the studies to be more accurate regarding specific machine learning. This article also brings up important topics of conversation for the integration of machine learning, such as how often data is collected and the result this has on computing power for the hospital system. In this study, 6 hour intervals proved sufficient for next day surgeries but were lacking in rapid unforeseen surgeries, where there is less known about the patient to begin with, leading to a bigger unknown. Overall, this article can be used by other future machine learning model designers to format the machine to use data readily available within the hospital, allowing future studies to test the most accurate models that will be used inside a hospital system. It also gives future ideas and concepts that can began to be explored now to overcome future obstacles when it comes to integrating machine learning in hospitals.