ABSTRACT NUMBER: ESOC2026A1638 DEEP LEARNING-BASED ONSET TIME ESTIMATION IN ACUTE ISCHEMIC STROKE

Abstract Background and aims In acute ischemic stroke, the time from onset to presentation guides treatment decisions, yet this stroke age is often unknown. Cross-sectional brain imaging may be used for stroke-age estimation, yet MRI is costly and NCCT shows only very subtle early ischemic changes, making visual assessment difficult. We aim to develop an automatic and clinically useful stroke age estimation tool that can detect subtle imaging signs while remaining robust across scanners and centers. Methods We used 1,633 non-contrast CT scans with annotated stroke ages from the two national multi-center cohorts: MR CLEAN Registry and MR CLEAN Late. Stroke ages were categorized into three classes using clinically relevant cut-offs at 4.5 h and 6 h. We developed ISANet, combining energy-guided contrastive learning and disentanglement learning to classify stroke age, enhancing clinical detection of subtle ischemic changes despite heterogeneous scanner and imaging parameters. Results ISANet demonstrated robustness to scanner-domain variations and label imbalance, achieving a macro F1-score of 0.57 and a macro AUC of 0.69 (Table I). Its performance on clinically critical <4.5 h and >6 h groups improved notably compared with traditional methods. As shown in Fig. 1, ISANet also captures subtle gray–white matter blurring and additional hypodense regions, revealing imaging biomarkers relevant to stroke age. Conclusions In this work, we have introduced an effective deep learning-based model ISANet for stroke age estimation, which is robust to scanner-type bias and class imbalance. The experimental results demonstrate its potential to support stroke-age window prediction for clinical decision-making. Conflict of interest the authors have nothing to disclose Table 1 - belongs to Results Figure 1 - belongs to Results