Abstracts

Rationale: Post-stroke epilepsy (PSE) contributes to a substantial proportion of epilepsy burden, especially in older adults. Moving beyond the investigation of stroke features associated with PSE development, recent years have seen attempts to devise prognostic models. The two most commonly used prognostic models for individualized PSE prediction are the CAVE score for hemorrhagic stroke and the SeLECT score for ischemic stroke. In contrast to the CAVE score, the SeLECT model did not consider the role of volume of acute injury in PSE. Instead, it includes binary imaging predictors of cortical and MCA territory involvement. Our goal is to determine if the size of stroke, as measured by Diffusion-weighted imaging (DWI) volume, is a significant predictor of PSE and if its addition improves the concordance index (C-index) of the SeLECT Score.

Methods: In this single center retrospective study, we included all stroke survivors with outpatient follow-up who had DWI data. For each patient, we manually extracted data on the SeLECT score variables from electronic medical records. The DWI volume analysis was performed by a previously validated convolutional neural network with 2D U-Net architecture. We computed the SeLECT score for each patient and re-fit the SeLECT score using a multivariable Cox proportional hazard model where time to post-stroke seizure ( > 7 days post-stroke) was the dependent variable. We then fit various models by adding DWI volume, and then removed one or both of cortical involvement and MCA. This created 6 models for which we evaluated the statistical significance of each variable, compared discrimination of the model using the C-index, and examined model calibration graphically.

Results: Of the 375 available patients, 221 had data for DWI volume. Of the 221 patients included, 35 (15.8%) had PSE. On unadjusted analysis, patients who had PSE had a higher SeLECT score, were more likely to have an early seizure, and had larger DWI volumes. DWI volume was found to be a statistically significant variable in all multivariable models (Table 1). The original SeLECT score had a C-index of 0.666 in our sample. Our refit SeLECT model had a C-index of 0.662. Adding DWI volume increased the C-index to 0.667. Keeping DWI and removing cortical involvement resulted in a C-index of 0.670. Keeping DWI and removing both cortical involvement and MCA resulted in a C-index of 0.660. Calibration was generally good for all models.

Conclusions: Stroke volume, as measured by our automated pipeline, is a significant predictor of PSE, independent of the SeLECT variables. The predictive performance of the SeLECT model remains comparable when stroke volume replaces cortical and MCA territory involvement. Stroke volume is a more precise measure of the extent of acute brain injury, and its extraction using automated pipelines can improve the efficiency and objectivity of future PSE prognostic models that go beyond the binary cortical and MCA territory involvement.

Funding: N/A