Prediction for Perioperative Stroke Using Intraoperative Parameters.

Area of Science:

• Medical Informatics
• Neurosurgery
• Cardiology

Background:

• Perioperative stroke is a severe surgical complication.
• Existing prediction models primarily use preoperative patient characteristics.
• There is a need for more accurate stroke risk assessment tools.

Purpose of the Study:

• To develop a machine learning model for predicting perioperative stroke.
• To evaluate the impact of incorporating intraoperative variables into stroke prediction.
• To compare the performance of an integrated model with a preoperative-only model.

Main Methods:

• A machine learning model was developed using data from 15,752 patients (Seoul National University Hospital) and validated externally with 449 patients (Boramae Medical Center).
• The study defined perioperative stroke as new ischemic lesions on diffusion-weighted imaging within 30 days of noncardiac surgery.
• Two models were compared: one using pre- and intraoperative factors (integrated model) and another using preoperative factors alone.

Main Results:

• The integrated model showed superior predictive performance in both internal (AUC 0.824) and external validation (AUC 0.716).
• The preoperative-only model had significantly lower predictive accuracy (internal AUC 0.584, external AUC 0.505).
• Perioperative stroke occurred in 0.69% of patients in the development cohort and 2.45% in the validation cohort.

Conclusions:

• Incorporating intraoperative factors into prediction models significantly enhances accuracy for perioperative stroke.
• The developed integrated model offers improved risk stratification for patients undergoing surgery.
• This approach can aid in identifying at-risk patients and potentially mitigating stroke complications.