A multimodal deep learning model for predicting early neurological deterioration in patients with acute ischemic stroke

TL;DR

A deep learning model combining clinical data and imaging reports improves prediction of neurological decline in stroke patients.

Contribution

A novel multimodal deep learning model that integrates structured clinical data and radiology text for END prediction in acute ischemic stroke.

Findings

The multimodal model achieved AUCs of 0.877 (training) and 0.771 (test) for predicting END.

Key predictors included D-dimer, diastolic blood pressure, and textual descriptors like 'unsteadiness' and 'vision'.

The model effectively stratified patients into high- and low-risk groups with significantly different END outcomes.

Abstract

Timely identification of patients at high risk for early neurological deterioration (END) is critical for effective intervention after acute ischemic stroke; however, existing prediction models largely rely on structured clinical data and underutilize semantic information from imaging findings.

In this retrospective single-center study at Xiangyang No.1 People’s Hospital (January 2018–December 2023), 426 patients with acute ischemic stroke and imaging-confirmed middle cerebral artery occlusion who received non-endovascular treatment were included. Patients with other arterial occlusions, endovascular therapy, or incomplete data were excluded. END was defined as a ≥2-point increase in total National Institutes of Health Stroke Scale (NIHSS) score or a ≥1-point increase in the motor subscore within 7 days after admission. Structured clinical variables and radiology report text collected at admission were integrated into a multimodal deep learning model. Model performance was evaluated using AUC, accuracy, recall, precision, F1-score, calibration, and decision curve analysis, with interpretability assessed using SHAP and Integrated Gradients.

Among all patients, 38.0% exhibited END (30.3% early; 7.7% late). The multimodal Concat-Fusion model achieved AUCs of 0.877 (training) and 0.771 (test), surpassing single-modality models, with strong predictive capabilities for early (AUC = 0.842) and late END (AUC = 0.855). Subgroup analyses confirmed consistent performance across NIHSS scores, D-dimer levels, hypertension, and atrial fibrillation, with significantly higher AUC in diabetic patients (p = 0.002). SHAP analysis identified D-dimer, diastolic blood pressure, and heart rate as key contributors. Integrated Gradients revealed symptom descriptors including “unsteadiness,” “walking,” and “vision” as significant textual predictors. Risk stratification effectively distinguished high- and low-risk groups with significantly different cumulative END incidences within 48 h and 7 days (log-rank p < 0.0001). High-risk patients exhibited poorer NIHSS trajectories during 14-day follow-up.

This multimodal prediction model may improve early identification of END and support individualized clinical management. Larger prospective studies are required to validate its clinical utility.

Figures