Development and External Validation of a Deep Learning Model to Predict Mortality in Aneurysmal Subarachnoid...

Area of Science:

• Neurosurgery
• Radiology
• Artificial Intelligence

Background:

• Traditional prognostication for aneurysmal subarachnoid hemorrhage (aSAH) relies on subjective clinical and radiological scores with high inter-rater variability.
• Existing methods require multiple variables, limiting their clinical applicability and reproducibility.

Purpose of the Study:

• To develop and validate a fully automated deep learning (DL) model for predicting 90-day mortality in aSAH patients.
• The model exclusively uses admission noncontrast computed tomography (NCCT) scans, aiming for objective, reproducible, image-only risk stratification.

Main Methods:

• A 3D DenseNet-121 DL model was trained using transfer learning on admission NCCT scans from 9 hospitals.
• Performance was evaluated against logistic regression models (Core, Imaging, Full Clinical) using discrimination, calibration, and decision-curve analysis.
• External validation was performed on data from 2 independent centers.

Main Results:

• The DL model achieved an area under the curve (AUC) of 0.855 (internal) and 0.806 (external validation).
• Performance was comparable to conventional models, with no statistically significant differences in discrimination or calibration.
• Decision-curve analysis indicated similar net benefit across clinically relevant thresholds.

Conclusions:

• A fully automated DL model using only NCCT can predict 90-day mortality in aSAH patients with performance comparable to traditional clinical models.
• This DL approach serves as a complementary decision-support tool, enhancing objective and reproducible risk stratification at the point of care.
• It requires no additional data collection beyond routine imaging, facilitating clinical adoption.