We previously developed a personalized AI model to predict surgical transfusion risk. Here we simulated using it for preop T&S decisions at 45 US hospitals. Our model (S-PATH) did better than the standard of care approach (MSBOS). We also measured performance using a clinically meaningful benchmark, the number of T&S ordered. S-PATH needed ~ 1/3 fewer while maintaining 96% sensitivity for finding patients who need blood. Importantly, we used S-PATH out of the box. No retraining or fine tuning on individual hospitals. Nonetheless, it still performed well. This kind of robustness is rare among AI models, and suggests S-PATH could be immediately useful for many hospital systems.
Many studies have suggested that higher cognitive burden is associated with increased burnout and risk for errors. However, it has been challenging to measure the cognitive load associated with clinician work within the electronic health record. Here we developed a novel scalable method to measure cognitive load using audit log data and demonstrate its validity.
20 million patients have surgery in the US every year, and ~ 1 million of those patients require life-saving blood transfusion. Presurgical preparation for transfusion is important to allow for safe and timely transfusion during surgery, but excessive preparation is unfortunately common, costly, and contributes to blood waste. In this paper, we develop a personalized surgical transfusion risk prediction model using a database of 3 million surgical patients, and show that using such a model to guide presurgical type and screen orders can potentially improve patient safety while reducing the number of unnecessary orders. Reproducible code is provided to make predictions for new patients.