Evaluation of machine learning models as decision aids for anesthesiologists

Area of Science:

• Anesthesiology
• Medical Informatics
• Artificial Intelligence

Background:

• Machine Learning (ML) models are increasingly used to predict clinical parameters.
• Accurate prediction of perioperative glucose and opioid requirements is crucial for patient management.

Purpose of the Study:

• To evaluate if ML models can improve anesthesiologists' predictions of peak intraoperative glucose and postoperative opioid requirements.
• To assess the impact of ML-driven decision support on clinical judgment.

Main Methods:

• 10 anesthesiologists predicted glucose and opioid needs for 100 patients, with and without ML model estimations.
• A web-based tool presented patient data and ML predictions.
• Anesthesiologists' prediction accuracy was compared, and feedback was collected via questionnaire.

Main Results:

• Anesthesiologists' accuracy in predicting peak glucose improved from 79.0% to 84.7% with ML assistance (p < 0.001).
• Accuracy in predicting opioid requirements rose from 18% to 42% with ML assistance (p < 0.001).
• ML predictions improved estimates for 8/10 anesthesiologists for glucose and 7/10 for opioids.

Conclusions:

• ML models can significantly enhance anesthesiologists' ability to predict key perioperative clinical parameters.
• ML predictions function effectively as reference tools, aiding clinicians in modifying their judgment and improving patient care outcomes.