A logistic regression model using six clinical variables achieved an AUROC of 0.832 for predicting delayed cerebral ischemia after aneurysmal subarachnoid hemorrhage, but incomplete reporting and absence of external validation currently prevent clinical implementation.
Key Findings
Methods
DCI occurred in 175 of 544 patients in the training cohort and 42 of 136 patients in the internal-validation cohort.
The study included 680 adult patients with aSAH treated between July 2022 and December 2024 at a single center.
Patients were divided using an outcome-stratified 8:2 split into training (544 patients) and hold-out internal-validation (136 patients) cohorts.
DCI incidence was approximately 32.2% in the training cohort (175/544) and 30.9% in the validation cohort (42/136).
Methods
LASSO regression with 10-fold cross-validation selected six predictor variables for the DCI prediction model.
Predictor selection was performed in the training cohort using least absolute shrinkage and selection operator (LASSO) regression.
The six retained variables were: age, cerebral edema, hypoalbuminemia, modified Fisher grade, Hunt-Hess grade, and World Federation of Neurological Surgeons grade.
10-fold cross-validation was used during the LASSO selection process.
Results
Logistic regression outperformed or was selected among five candidate machine learning models for DCI prediction.
Five models were compared: logistic regression, extreme gradient boosting, light gradient boosting machine, support vector machine, and k-nearest neighbor.
Logistic regression was identified as the best-performing model based on validation cohort results.
The other models (XGBoost, LightGBM, SVM, KNN) were evaluated but did not surpass logistic regression performance in the internal-validation cohort.
Results
Logistic regression achieved an AUROC of 0.832 in the internal-validation cohort.
The area under the receiver operating characteristic curve was 0.832 (95% confidence interval, 0.758–0.906) in the internal-validation cohort of 136 patients.
The calibration slope was 0.98 and the calibration intercept was 0.02, suggesting good calibration.
The Brier score was 0.168, indicating moderate overall predictive accuracy.
Confidence intervals for the calibration slope, intercept, and Brier score were not available.
Discussion
Several methodological limitations prevent patient-level probability calculation and clinical implementation of the model.
Incomplete reproducibility records and the absence of the model intercept prevent patient-level probability calculation.
No resampling-based optimism correction (e.g., bootstrap) was applied to account for overfitting.
External validation in an independent dataset was not performed.
The authors note that 'the findings represent preliminary performance in a single hold-out internal-validation cohort.'
Confidence intervals for calibration estimates were unavailable.
What This Means
This research suggests that a statistical model using six clinical characteristics — patient age, presence of brain swelling (cerebral edema), low blood protein levels (hypoalbuminemia), and three standard neurological severity scores — can predict which patients with a ruptured brain aneurysm will go on to develop a serious complication called delayed cerebral ischemia (DCI). DCI is a condition where parts of the brain suffer reduced blood flow in the days following a hemorrhage, and it is a major cause of disability and death in these patients. The model, built from records of 680 patients at a single hospital, correctly distinguished patients who developed DCI from those who did not about 83% of the time, and its probability estimates appeared reasonably well-calibrated to actual outcomes.
The study compared five different modeling approaches and found that a traditional logistic regression model performed best, matching or exceeding more complex machine learning techniques. This is a meaningful finding because simpler models are generally easier for clinicians to understand and use. However, the study has important limitations: it was conducted at only one hospital, the model was only tested on a portion of that same hospital's data (not an entirely separate dataset), and key technical details needed to actually apply the model to individual patients were not fully reported.
This research suggests that clinical prediction tools for DCI are feasible and could eventually help clinicians identify high-risk patients earlier. However, further work — including external validation at other hospitals, correction for potential overoptimism in the results, and complete reporting of the model — would be needed before such a tool could be responsibly used in clinical practice.
Fei L, Lin Z. (2026). Internal Validation of a Clinical Prediction Model for Delayed Cerebral Ischemia After Aneurysmal Subarachnoid Hemorrhage.. Journal of visualized experiments : JoVE. https://doi.org/10.3791/72237