Compact machine learning model for perioperative stroke prediction prior to surgery: A retrospective cohort study

Area of Science:

• Neurosurgery
• Cardiology
• Artificial Intelligence

Background:

• Perioperative stroke is a significant complication affecting patient outcomes.
• Current methods for predicting perioperative stroke risk are neither accurate nor practical for clinical use.

Purpose of the Study:

• To develop a machine learning (ML) model for enhanced accuracy and usability in predicting perioperative strokes.
• To compare the performance of ML models against established cardiovascular risk scores.

Main Methods:

• Utilized data from 36,502 patients for internal validation and 404 patients for external validation.
• Developed and compared various ML models, including a compact model with the top 10 features.
• Defined perioperative stroke as ischemic brain infarction within 30 days post-surgery.

Main Results:

• The CatBoost-based ML model demonstrated superior discriminatory power for high-risk patients.
• The ML model significantly outperformed cardiovascular scores (revised cardiac index and CHA2DS2VASc) in external validation (AUC 0.867 vs. 0.528 and 0.706).
• A compact ML model also showed improved performance (AUC 0.875).

Conclusions:

• The developed ML-based model offers improved accuracy and clinical usability for perioperative stroke prediction.
• This novel approach has the potential to enhance patient safety and optimize surgical decision-making.