Developing a Predictive Model for Ischemic Stroke Onset Time Using Transfer Learning.

Abstract

Introduction: Identification of acute ischemic stroke (AIS) patients within the 4.5-h therapeutic window is critical for therapy. Diffusion-weighted imaging (DWI) and fluid-attenuated inversion recovery (FLAIR) sequences are an approach to determine whether the time since stroke (TSS) is within 4.5 h. However, interobserver variability and limited accuracy are observed in visual assessments. We aimed to develop a transfer learning model for predicting AIS onset within 4.5 h.

Materials and methods: A total of 266 AIS patients with known TSS who underwent imaging scans before treatment were retrospectively analyzed, divided into a training set (n = 211) and a validation set (n = 55). The model was built using DWI and FLAIR sequences. After image preprocessing and data augmentation, a 3D ResNet-18 pretrained on the Kinetics dataset was selected and adapted via transfer learning with DWI-FLAIR input. The model performance was compared with human visual assessment, which was based on the DWI-FLAIR mismatch principle. Partial mismatch was defined as hyperintense infarct on DWI with a smaller corresponding hyperintense area on FLAIR.

Results: Baseline characteristics did not differ between the training and validation sets. On the validation set, the model achieved sensitivity of 0.833 (0.703-0.941), specificity of 0.880 (0.737-1.000), and AUC of 0.929 (0.758-0.935), outperforming human visual assessment (sensitivity 0.767 [0.613-0.903]; specificity 0.360 [0.185-0.560]; AUC 0.563 [0.451-0.693]). For partial DWI-FLAIR mismatch cases, the model correctly classified all 15 cases, whereas humans classified 4.

Conclusion: The 3D ResNet-18 model shows promise in identifying AIS within 4.5 h, including partial DWI-FLAIR mismatch, but requires multicenter validation before use.