Early neurological deterioration (END) remains a critical determinant of outcomes following acute ischemic stroke (AIS). This retrospective, single-center study, published in the Frontiers in Neurology conducted at Xiangyang No.1 People’s Hospital (January 2018–December 2023) assessed whether integrating structured clinical variables with radiology report text using a multimodal deep learning model could improve END prediction. 

A total of 426 patients with AIS and imaging-confirmed middle cerebral artery (MCA) occlusion who received non-endovascular treatment were included. Patients with other arterial occlusions, those undergoing endovascular therapy, or with incomplete datasets were excluded. END was defined as a ≥2-point increase in total National Institutes of Health Stroke Scale (NIHSS) score or ≥1-point increase in motor subscore within 7 days.

Overall, 38.0% of patients developed END (30.3% early; 7.7% late). The Concat-Fusion multimodal model demonstrated superior performance compared with single-modality approaches, achieving area under the curve (AUC) values of 0.877 (training) and 0.771 (test). It showed strong predictive ability for early (AUC 0.842) and late END (AUC 0.855). 

Performance remained consistent across subgroups defined by NIHSS, D-dimer, hypertension, and atrial fibrillation, with higher AUC observed in diabetes (p=0.002). Key predictors included D-dimer, diastolic blood pressure, and heart rate, while textual features such as “unsteadiness,” “walking,” and “vision” contributed significantly. Risk stratification differentiated high- and low-risk groups with significant differences in END incidence and NIHSS trajectories.

This model may support early END identification and individualized management. Further prospective validation is required to confirm clinical applicability.