Abstract 034: Deep Learning to Predict Intracranial Atherosclerosis‐Related Large Vessel Occlusion Using Neuroimaging and Clinical Data

Predicting successful recanalization in acute ischemic stroke (AIS) patients undergoing mechanical thrombectomy (MT) is crucial for optimizing treatment strategies but remains challenging. Intracranial atherosclerosis‐related large vessel occlusion (ICAS‐LVO) may require early bailout strategies after a failed MT attempt with standard devices. Existing predictive models often struggle to effectively capture the complex interactions between neuroimaging and clinical data. This study aims to develop predictive models for ICAS‐LVO following an unsuccessful initial MT pass by integrating both neuroimaging and clinical data. A retrospective, single‐center study was conducted on patients with anterior LVO stroke from 2017 to 2023. ICAS‐LVO was defined by angiographic stenosis >50% or failed MT without another identifiable cause. Neuroimaging data included non‐contrast CT and CTA volumes, automated vascular segmentation, and CTP output values. Clinical variables, including patient demographics, stroke severity, and comorbidities, were also incorporated. A Dynamic Affine Feature Map Transform (DAFT) mechanism integrated imaging data, encoded by a convolutional neural network, with clinical data. Multiple deep learning models were compared to traditional machine learning models trained on clinical data alone. The data were split 80% for training with four‐fold cross‐validation and 20% for testing. The study included 262 patients (median age 81, IQR 69‐88; 61% women; 11.83% with ICAS‐LVO) with anterior LVO and an initial failed MT attempt (eTICI 0‐2a). The training set comprised 212 patients (12.73% with ICAS‐LVO), while 50 patients (8% with ICAS‐LVO) were reserved for testing. The best model achieved an AUC of 0.73 ± 0.10 and an F1 score of 0.44 ± 0.15 in validation, with an AUC of 0.64 ± 0.12 and an F1 score of 0.21 ± 0.09 on testing. While predicting ICAS‐LVO is challenging, models combining neuroimaging and clinical data outperformed single‐source models, highlighting the importance of multi‐modal integration to enhance predictive accuracy and optimize treatment strategies.