Evaluation of explainable artificial intelligence for subarachnoid hemorrhage detection and outcome modelling using neuroimaging data

Subarachnoid hemorrhage (SAH) is a life-threatening neurologic emergency. SAH Detection Artificial Intelligence (AI)-Explainable AI (SAHDAI-XAI) is a cloud-based machine learning (ML) model that classifies computed tomography (CT) images as positive or negative for SAH across 8 hemorrhage spaces. Phase 1 of our project used Vertex AI (Google Cloud) automated ML to train models on progressively larger noncontrast head CT datasets (20–300 images), evenly labeled as hemorrhage-positive or hemorrhage-negative.

Explainability was applied to images, and the final round evaluated differentiation between hemorrhages and other pathologies. Phase 2 focused on explainability by visualizing predictive features and localizing hemorrhages with the Brain Lesion Analysis and Segmentation Tool for CT (BLAST-CT) model, which segments and quantifies 4 hemorrhage regions. In Phase 1, the first 2 automated ML rounds showed 100.0% precision due to the small dataset size.

With larger datasets, third-round precision was 97.9% with 1 false negative, and fourth-round precision was 94.4% with 2 false positives. In Phase 2, after preprocessing using BLAST-CT, topographic hemorrhage maps were generated, with some accurately capturing a large portion of the hemorrhage, while others performed less reliably. SAHDAI-XAI improves transparency in SAH detection by using BLAST-CT visualizations to clarify predictive features and reduce black-box constraints.

The Scientific Publications staff at Mayo Clinic provided copyediting, proofreading, administrative, and clerical support. There was no direct funding for this work; however Dr. Freeman wishes to acknowledge that Mayo receives generous philanthropic support from Louis V.

Gerstner Philanthropies and he is Principal Investigator of the Mayo Gerstner AI Scholars Program for SAHVAI (Subarachnoid Hemorrhage Volumetric AI) which is focused on artificial intelligence creation and mathematics of SAH hemorrhage volumetrics to advance and translate back to patient care. Freeman received no financial compensation for this work. Department of Neurologic Surgery, Mayo Clinic, 4500 San Pablo Rd, FL, 32224, Jacksonville, USA Sarah Morgan, Saif Salman, Rabih Tawk & W.

David Freeman Department of Neurology, Mayo Clinic, Jacksonville, FL, USA Sarah Morgan, Saif Salman & W. David Freeman Department of Critical Care Medicine, Mayo Clinic, Jacksonville, FL, USA Department of Information Technology, Mayo Clinic, Rochester, MN, USA Department of Radiology, Mayo Clinic, Jacksonville, FL, USA The authors declare no competing interests. Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.