An automated machine learning-based model predicts postoperative mortality using readily-extractable preoperative...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Surgical Risk Stratification

Background:

• Accurate preoperative risk identification is crucial for optimal resource allocation in healthcare.
• Existing risk assessment tools often lack patient-level specificity or require manual chart review.
• The American Society of Anesthesiologists (ASA) physical status classification necessitates clinician chart review.

Purpose of the Study:

• To develop a fully automated score for predicting postoperative in-hospital mortality.
• To utilize machine learning algorithms with structured electronic health record data.
• To compare the performance of the automated score against established risk assessment tools.

Main Methods:

• Employed random forest machine learning algorithms.
• Extracted 58 preoperative features from electronic health records of 53,097 surgical patients.
• Validated the automated score against POSPOM, Charlson comorbidity, and ASA physical status scores.

Main Results:

• The automated score achieved an area under the curve (AUC) of 0.932, outperforming POSPOM (AUC 0.660), Charlson (AUC 0.742), and ASA (AUC 0.866).
• Incorporating ASA physical status with automated features yielded an AUC of 0.936.
• The developed score demonstrated superior predictive accuracy for in-hospital mortality.

Conclusions:

• The automated machine learning score significantly outperforms traditional risk scores in predicting in-hospital mortality.
• This automated tool enhances the precision of preoperative risk assessment.
• Integration with postoperative scores highlights dynamic perioperative risk changes.