The Role of AI-Based Software BrainScan in the Interpretation of Non-Contrast Head CT in Acute Ischemic Stroke: An External Validation Study.
Authors
Affiliations (6)
Affiliations (6)
- Department of Neurology, Medical University of Plovdiv, 15A Vasil Aprilov Blvd., 4002 Plovdiv, Bulgaria.
- Clinic of Neurology, UMHAT "St George", 66 Peshtersko Shose Blvd., 4001 Plovdiv, Bulgaria.
- Department of Radiology, UMHAT "St George", 66 Peshtersko Shose Blvd., 4001 Plovdiv, Bulgaria.
- Department of Social Medicine and Public Health, Medical University of Plovdiv, 15A Vasil Aprilov Blvd., 4002 Plovdiv, Bulgaria.
- Environmental Health Division, Research Institute, Medical University of Plovdiv, 15A Vasil Aprilov Blvd., 4002 Plovdiv, Bulgaria.
- Department of Forensic Medicine and Deontology, Medical University of Plovdiv, 15A Vasil Aprilov Blvd., 4002 Plovdiv, Bulgaria.
Abstract
Artificial intelligence (AI) tools are increasingly integrated into acute stroke imaging workflows, but real-world performance for ischemia detection on non-contrast CT (NCCT) remains incompletely validated by investigators independent of the developer. This study externally validated the BrainScan AI system in an unselected, consecutively enrolled emergency cohort. Consecutive adult patients undergoing NCCT under the routine acute stroke protocol at a single tertiary centre between January and December 2025 were prospectively enrolled. The reference standard was the post-consensus radiological diagnosis, supplemented where available by follow-up imaging and clinical course. Primary outcomes were diagnostic accuracy for ischemia and intracranial haemorrhage detection, assessed by sensitivity, specificity, predictive values, likelihood ratios, and area under the ROC curve (AUC; DeLong). Pre-specified secondary analyses included regional sensitivity, confidence-score behaviour, artefact robustness, threshold sensitivity, a cluster-robust bootstrap for within-patient correlation, and a quantitative bias analysis under non-differential reference-standard misclassification. Sample size adequacy was assessed using a precision-based framework. A total of 1419 NCCT examinations from 1260 patients were analysed. Ischemia sensitivity was 59.2% (95% CI 52.1-66.1) and specificity was 99.8% (99.4-100), with an AUC of 0.930 (0.906-0.954). The Youden-optimal threshold (0.055) recovered sensitivity to 86.1% with negligible specificity loss, reflecting a markedly bimodal score distribution. Regional sensitivity was lower in infratentorial structures. Bias-corrected estimates were stable across all reference-standard parameters consistent with the data. Haemorrhage detection performed substantially better (sensitivity 96.7%; AUC 0.983). The system shows excellent specificity and strong discrimination but moderate sensitivity for ischemia, supporting its role as a rule-in adjunct rather than a stand-alone tool, pending multicentre validation and site-specific threshold recalibration.