Assessing the Utility of a Machine-Learning Model to Assist With the Assignment of the American Society of Anesthesiology Physical Status Classification in Pediatric Patients.

Background: The American Society of Anesthesiologists Physical Status Classification System (ASA-PS) is used to classify patients' health before delivering an anesthetic. Assigning an ASA-PS Classification score to pediatric patients can be challenging due to the vast array of chronic conditions present in the pediatric population. The specific aims of this study were to (1) suggest an ASA-PS score for pediatric patients undergoing elective surgical procedures using machine-learning (ML) methods; and (2) assess the impact of presenting the suggested ASA-PS score to clinicians when making their final ASA-PS assignment. The intent was not to create a new ASA-PS score but to use ML methods to generate a suggested score, along with information on how the score was generated (ie, historical information on patient comorbidities) to assist clinicians when assigning their final ASA-PS score.

Methods: A retrospective analysis of 146,784 pediatric surgical encounters from January 1, 2016, to December 31, 2019, using eXtreme Gradient Boosting (XGBoost) methods to predict ASA-PS scores using patients' age, weight, and chronic conditions. SHapley Additive exPlanations (SHAP) were used to assess patient characteristics that contributed most to the predicted ASA-PS scores. The predicted ASA-PS model was presented to a prospective cohort study of 28,677 surgical encounters from December 1, 2021, to October 31, 2022. The predicted ASA-PS score was presented to the anesthesiology provider for review before entering the final ASA-PS score. The study focused on summarizing the available information for the anesthesiologist by using ML methods. The goal was to explore the potential for ML to provide assistance to anesthesiologists by highlighting potential areas of discordance between the variables that generated a given ML prediction and the physician's mental model of the patient's medical comorbidities.

Results: For the retrospective analysis, the distribution of predicted ASA-PS scores was 22.7% ASA-PS I, 48.5% II, 23.6% III, 5.1% IV, and 0.04% V. The distribution of clinician-assigned ASA-PS scores was 24.3% for ASA-PS I, 44.5% for ASA-PS II, 24.9% for ASA III, 6.1% for ASA-PS IV, and 0.2% for ASA-V. In the prospective analysis, the final ASA-PS score matched the initial ASA-PS 90.7% of the time and 9.3% were revised after viewing the predicted ASA-PS score. When the initial ASA-PS score and the ML ASA-PS score were discrepant, 19.5% of the cases have a final ASA-PS score which is different from the initial clinician ASA-PS score. The prevalence of multiple chronic conditions increased with ASA-PS score: 34.9% ASA-PS I, 73.2% II, 92.3% III, and 94.4% IV.

Conclusions: ML derivation of predicted pediatric ASA-PS scores was successful, with a strong agreement between predicted and clinician-entered ASA-PS scores. Presentation of predicted ASA-PS scores was associated with revision in final scoring for 1-in-10 pediatric patients.