An explainable machine learning framework for computable physiologic risk representation in preanesthetic assessment:...

Area of Science:

• Anesthesiology and Perioperative Medicine
• Artificial Intelligence in Healthcare
• Clinical Risk Assessment

Background:

• Preanesthetic evaluation synthesizes complex clinical data, but current risk communication often uses subjective, categorical methods like the American Society of Anesthesiologists Physical Status (ASA-PS) classification.
• The ASA-PS classification exhibits inter-rater and institutional variability, limiting its precision in capturing true physiologic risk.
• There is a need for more objective and data-driven approaches to represent physiologic severity in preanesthetic assessments.

Purpose of the Study:

• To develop and externally validate an explainable machine learning (ML) framework for representing physiologic severity in preanesthetic risk assessment.
• To create a computable tool that complements existing subjective risk classification systems.
• To enhance the consistency and accuracy of perioperative risk communication.

Main Methods:

• A retrospective study utilized data from two Taiwanese institutions for model development (n=1,200) and external validation (n=113).
• Supervised learning employed a physiologic severity label derived from postoperative Acute Physiology and Chronic Health Evaluation II (APACHE II) scores (≥12).
• Least Absolute Shrinkage and Selection Operator (LASSO) regression was used for feature selection, and EasyEnsemble-Light Gradient Boosting Machine (LightGBM) addressed class imbalance. Model performance was assessed using AUROC, calibration, and DCA.

Main Results:

• The ML model incorporated six key variables: surgical site, age, white blood cell count, heart rate, mean arterial pressure, and smoking status.
• Internal validation showed an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.94, significantly outperforming ASA-PS (0.71).
• External validation achieved an AUROC of 0.88, also superior to ASA-PS (0.76), with Decision Curve Analysis (DCA) favoring the ML approach for clinical utility.

Conclusions:

• The developed ML framework provides an externally validated, computable representation of physiologic severity.
• This explainable AI tool complements the traditional ASA-PS classification, offering more nuanced risk stratification.
• The framework has the potential to support more consistent perioperative risk communication and improve preanesthetic decision-making.