# Automated machine learning for predicting perioperative ischemia stroke in endovascularly treated ruptured intracranial aneurysm patients

## TL;DR

This study created a machine learning tool to predict the risk of stroke after endovascular treatment for brain aneurysms, helping doctors make better decisions.

## Contribution

The novel contribution is an interpretable machine learning model with a web-based calculator for predicting postoperative ischemic stroke in ruptured aneurysm patients.

## Key findings

- The random forest model achieved 92.11% accuracy in predicting perioperative ischemic stroke.
- Thirteen clinically actionable predictors were identified, including vascular risk factors and neuroimaging indicators.
- A web-based calculator was developed for clinicians to predict stroke risk and tailor treatment plans.

## Abstract

This study aims to develop and validate an automated machine learning model to predict perioperative ischemic stroke (PIS) risk in endovascularly treated patients with ruptured intracranial aneurysms (RIAs), with the goal of establishing a clinical decision-support tool.

In this retrospective cohort study, we analyzed RIA patients undergoing endovascular treatment at our neurosurgical center (December 2013–February 2024). The least absolute shrinkage and selection operator (LASSO) method was used to screen essential features associated with PIS. Based on these features, nine machine learning models were constructed using a training set (75% of participants) and assessed on a test set (25% of participants). Through comparative analysis, using metrics such as area under the receiver operating characteristic curve (ROCAUC) and Brier score, we identified the optimal model—random forest (RF)—for predicting PIS. To interpret the RF models, we utilized the Shapley Additive exPlanations (SHAP).

The final cohort comprised 647 consecutive RIA patients who underwent endovascular intervention. LASSO regression identified 13 clinically actionable predictors of PIS from the initial variables. These predictors encompassed: vascular risk factors (hyperlipidemia, arteriosclerosis); neuroimaging indicators of severity (modified Fisher scale, aneurysm location, and neck-to-diameter ratio); clinical status (Glasgow Coma Scale score, Hunt-Hess grade, age, sex); procedural complications (intraprocedural rupture, periprocedural re-rupture); and therapeutic determinants (therapy method and history of ischemic comorbidities). Nine machine learning algorithms were evaluated using stratified 10-fold cross-validation. Among them, the RF model demonstrated the best performance, with the ROCAUC of 92.11% (95%CI: 89.74–94.48%) on the test set and 87.08% (95%CI: 81.23–92.93%) on the training set. Finally, in a prospective validation cohort, the RF predictive model demonstrated an accuracy of 88.23% in forecasting the incidence of PIS. Additionally, based on this predictive model, this study developed a highly convenient web-based calculator. Clinicians only need to input the patient’s key factors into this calculator to predict the postoperative incidence of PIS and provide individualized treatment plans for the patient.

We successfully developed and validated an interpretable machine learning framework, integrated with a clinical decision-support system, for predicting postprocedural PIS in endovascularly treated RIAs patients. This tool effectively predicted the likelihood of PIS, enabling high-risk patients to promptly take specific preventive and therapeutic measures.