Identifying Risk Factors of Perioperative Stroke: An Analysis of Feature Selection Techniques Using the American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP) Database
Objective:
To identify a parsimonious but highly predictive set of risk factors for perioperative cerebrovascular accident (CVA) by systematically evaluating four machine learning feature selection algorithms on a large national surgical database.
Background:
Since perioperative CVA carries high morbidity, it is used in risk models such as the Society of Thoracic Surgeons (STS) cardiac surgery risk model. Methods to identify significant risk factors from large databases is a necessary and an often underappreciated step when creating machine learning models. We compared different algorithms for optimal feature selection in identifying patients who may be at risk of perioperative CVA.
Design/Methods:
We conducted a retrospective analysis by using a random set of 60% (645,864 patients) of the 2019 ACS-NSQIP database patients for feature selection and the other 40% (430,576) for CatBoost model generation and validation. Four distinct feature selection methods - Boruta, Recursive Feature Elimination (RFE), Sequential Forward Selection (SFS), and Sequential Backward Selection (SBS) - were employed to reduce an initial set of 60 perioperative variables. The predictive value of each resulting feature subset was assessed by training a CatBoost gradient boosting model and evaluating its calibration and accuracy using the Brier score. Analysis was done using Python 3.12.
Results:
The Boruta algorithm identified a subset of 7 features while achieving the lowest Brier score of 0.0017738. SFS selected 34 features (Brier score of 0.001777), SBS selected 60 variables (Brier score of 0.00178), and RFE selected one feature (Brier score of 0.00178). The seven features identified using Boruta were: age, ASA class, inpatient/outpatient status, primary CPT code, surgical specialty, preoperative sodium, and whether the surgery was elective.
Conclusions:
The Boruta algorithm identified an ideal minimal feature set to predict perioperative CVA. This is ideal for developing a practical, user-friendly preoperative risk stratification tool or for integration into electronic health records to automatically flag high-risk patients for targeted intervention.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.