ABSTRACT
Background and Purpose
Stroke is a leading cause of epilepsy, especially in older adults. The SeLECT score remains the standard among post‐stroke epilepsy (PSE) prediction tools. However, its broader validation is limited by the need to manually extract neuroimaging predictors (cortical and middle cerebral artery [MCA] involvement). Unlike the CAVE score, SeLECT did not evaluate acute stroke volume, which can now be quantified automatically. We aimed to determine whether stroke volume independently predicts PSE and compare its predictive contribution to SeLECT's neuroimaging variables.
Methods
SeLECT variables were manually extracted. Diffusion‐weighted imaging volume was quantified using a validated convolutional neural network. Cox proportional hazards models for time to PSE were built by adding stroke volume (per 10 mL) and then removing cortical and/or MCA involvement. For each model, we analyzed variable significance, discrimination, and calibration.
Results
Among 221 patients, 35 (15.8%) developed PSE. In our cohort, the original SeLECT score and the refit model had a C‐index of 0.669 and 0.642, respectively. Adding stroke volume resulted in a C‐index of 0.656. Retaining volume while removing cortical and MCA involvement resulted in C‐indices of 0.664 and 0.668, respectively. Keeping stroke volume and removing both variables increased the C‐index to 0.679. Calibration was good for all models. Stroke volume in crease by 10 mL was an independent predictor of 12% increased PSE risk across all models.
Conclusions
Acute stroke volume is an independent PSE predictor. Stroke volume offered comparable discrimination to the neuroimaging components of the SeLECT score, supporting its use as a scalable and automated alternative.
Keywords: acute stroke volume, AI tools, diffusion‐weighted imaging, post‐stroke epilepsy, prognostic models
1. Introduction
Post‐stroke epilepsy (PSE) contributes to a substantial proportion of epilepsy burden, especially in older adults, where it accounts for half of all the new epilepsy cases [1]. Our population is rapidly aging. Substantial effort has, therefore, been made in the last decade to create prognostic PSE models. The two most used predictive models for individualized PSE prognostication are the CAVE score for intracerebral hemorrhage (ICH) and the SeLECT score for ischemic stroke [2, 3]. The SeLECT score includes five predictors: severity of stroke (NIHSS), large‐artery atherosclerosis etiology, early seizure (</ = 7 days), cortical involvement, and territory of the middle cerebral artery (MCA) [2]. It was developed and validated in a European cohort. While several iterations of the SeLECT score have been published, we lack large‐scale studies to validate and improve the SeLECT score in geographically diverse populations, especially in the United States [4]. A major barrier to applying the SeLECT score in large datasets is the need to manually extract MCA and cortical involvement from neuroimaging or clinical notes. These features are not routinely or consistently coded and require expert review of imaging or detailed electronic health records (EHRs) chart review.
Recent neuroimaging research on lesion mapping found that the volume of post‐stroke chronic encephalomalacia is strongly associated with epilepsy development [5]. Acute neuroimaging data for the CAVE score found ICH volume to be a significant predictor of epilepsy development when it is more than 10 mL [3]. However, the SeLECT model did not consider the role of volume of acute ischemic stroke in PSE and only included two binary neuroimaging predictors (cortical and MCA territory involvement) [2, 3]. Advances in machine learning and artificial intelligence have allowed automated tools for calculating ischemic stroke volumes using diffusion‐weighted imaging (DWI) [6, 7]. The goal of our study was to determine if the size of acute ischemic stroke is a significant predictor of PSE and if its use can help overcome the limitation of manual extraction of neuroimaging data used in the SeLECT score.
2. Methods
After institutional review board approval, we conducted a retrospective, single‐center study at the Cleveland Clinic in Cleveland, OH, USA. The study included adults with acute stroke admissions from January 2013 to October 2018, and DWI data from acute Magnetic Resonance Imaging (MRI). Cortical and MCA territory involvement was independently extracted by two board‐certified neurologists (VD and VP). Disagreements were mutually resolved. We obtained data on each variable of the SeLECT score through manual extraction from the EHRs. Patients without an MRI during admission or insufficient data to calculate the SeLECT score were excluded. Using these variables, we computed the SeLECT score for each patient in our sample. Clinical seizures during hospital admission and follow‐up, determined and documented in the EHR by a licensed healthcare provider, were extracted. A clinical seizure within ≤7 days of stroke was classified as an acute symptomatic seizure (ASyS), and a seizure > 7 days after stroke was classified as a PSE [8].
2.1. MRI Acquisition
Clinically acquired MRIs were obtained retrospectively through the Picture Archiving and Communication System. Axial diffusion‐weighted echo planar images, including b = 0 and trace (b‐value = 1000 s/mm2), were used to calculate the ischemic stroke volume. The sequence median (interquartile range [IQR]) parameters were echo time = 114 (98, 114) ms, repetition time = 3802 (3500, 4200) ms, in‐plane pixel resolution = 1.8 (1.7, 1.8) mm, and slice thickness = 7.0 (7.0, 7.0) mm.
2.2. DWI Analysis
The training datasets came from a separate cohort of 175 patients with their first acute ischemic stroke, incomplete outcomes, or missing baseline MRI [9]. The b0 and trace images were used to segment regions of hyperintense areas on trace images. For the training and validation, we manually segmented hyperintense DWI regions with consultation with a neuroradiologist (Appendix Figure 1). The median DWI volume (IQR) was 3.0 (0.8−10.3) mL. We developed a convolutional neural network with a 2D U‐Net architecture [10]. Briefly, the architecture had six levels with skip connections, and each unit with convolutions and batch normalization layers. The data were split into training, validation, and testing sets as a 5:1:1 ratio.
The mean Dice coefficient on the validation dataset was 0.87, and the testing dataset was 0.81, which was within a typical range [6]. The Dice similarity coefficient score is a statistical measure used to quantify the spatial overlap between two binary segmentations, typically, the automated segmentation output and the ground truth manual segmentation, of stroke lesions in neuroimaging studies [11]. As a post‐processing, we used the b0 and trace images to calculate the maps of apparent diffusion coefficients (ADCs), and the hypointense ADC areas were retained if the lesion‐wise 25th percentile ADC value was smaller than 600×10−6 mm2/s and if the lesion size is 4 voxels or greater with an 18‐connectivity. The results were visually checked by a rater and modified if necessary.
For external validation, we applied the model to the Ischemic Stroke Lesion Segmentation 2022 (ISLES22) training dataset, consisting of 250 studies [12]. Briefly, ISLES22 included subjects 18 years or older who had received MR imaging of the brain for previously diagnosed or suspected stroke. The MRI protocols included FLAIR and DWI trace images at a b‐value of up to 1000 s/mm2, as well as an ADC map at a 2 mm isotropic resolution. The zero‐shot validation showed a median Dice (IQR) of 0.65 (0.38, 0.76). However, the median (IQR) DWI lesion volume in ISLES 22 was 6.66 (1.58, 21.2) mL, which was larger than the internal training dataset (t‐test, p<0.001). After the model was fine‐tuned to ISLES22, the Dice improved to 0.82 (0.65, 0.88).
2.3. Statistical Analysis
Descriptive statistics were computed for the entire cohort and stratified by whether the patient developed PSE. Mean with standard deviation or median with IQR was used for continuous variables, and frequency with percentage was used for categorical variables. Comparisons were made using two‐sample t‐tests or Wilcoxon rank sum tests for continuous variables and Fisher's exact test for categorical variables.
We first applied the published point‐based SeLECT score to our cohort [2]. Then, we refit the SeLECT model in our cohort by re‐estimating the coefficients of its original predictors using a multivariable Cox proportional hazards model with time to PSE as the outcome. This allowed us to assess how the published SeLECT model performed in our data and to compare it with both the refit model and the additional models we developed. Patients who did not develop PSE were censored at their last follow‐up date. The proportional hazards assumption was evaluated using Schoenfeld residuals, and no major violations were detected.
We then fit various models by adding stroke volume and then removing one or both of cortical involvement and MCA. Stroke volume was modeled as a continuous variable (per 10 mL). Nonlinearity was examined visually, and no major deviations from linearity were observed. Therefore, the total number of models that were fit includes the following:
Refit SeLECT model (NIHSS + LAA + early seizure + cortical involvement + MCA)
SeLECT + stroke volume
SeLECT + stroke volume—cortical involvement
SeLECT + stroke volume—MCA
SeLECT + stroke volume—cortical involvement—MCA
For each of these models, we compared discrimination using the concordance index (C‐index). Model calibration was examined graphically. We internally validated each of these models using bootstrapping with 500 iterations. To determine whether the optimism‐corrected C‐index was different between models, we used bootstrapping with 1000 bootstrap samples to compute 95% confidence intervals for the difference between each pair of optimism‐corrected C‐index values.
Furthermore, for clinical use purposes, we analyzed whether there was a stroke volume threshold beyond which the size of the stroke is significantly more predictive of PSE. Because time to PSE is a time‐dependent outcome, we performed time‐dependent receiver operating characteristic (ROC) analysis, where time to PSE was the dependent variable, and stroke volume was the predictor. Specifically, we used the cumulative sensitivity and dynamic specificity method [13]. We created ROC curves for each of the 6‐month, 1‐year, and 2‐year periods, and the area under the ROC curve (AUC), along with a 95% confidence interval, was computed. The optimal cutoff point for stroke volume was determined at each time point using Youden's index (the maximum vertical distance from the ROC curve to the diagonal line) [14]. The sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) at each time point, based on the corresponding stroke volume cutoff, were estimated. All computations were performed in R, version 4.3.1 [15]. All tests were two‐sided, and statistical significance was defined as p < 0.05.
3. Results
Of the 375 patients qualifying for the study criteria, 221 had both DWI data and data for all SeLECT variables and were included in the final analysis. Compared with the study cohort (N = 221), excluded patients had significantly more frequent cortical involvement and higher NIHSS scores, resulting in higher SeLECT scores (Appendix Table 1). Representative examples of automated DWI lesion segmentation are shown in Figure 1. Cortical involvement on initial Computed Tomography (CT) is strongly associated with higher NIHSS [16, 17]. The confirmation of stroke on the initial CT scan explains the lower likelihood of undergoing a subsequent MRI, leading to their exclusion from the current analysis.
FIGURE 1.
Examples of DWI trace images (left) and automatically segmented lesion masks shown by red (right) in two patients with (a, b) and without (c, d) post‐stroke epilepsy development. Top row (a, b) patient with stroke volume of 43.0 mL. Bottom row (c, d) with stroke volume of 22.8 mL. Both patients have cortical involvement (1 score point in SeLECT). The bottom row patient, with lower stroke volume, did not develop PSE despite MCA territory infarct (1 score point in SeLECT). In contrast, the top row patient, with a larger stroke volume, developed PSE, despite a non‐MCA territory stroke.
The mean age of the study cohort was 63.5 (±13.8) years and included 51% females. MRIs were acquired on Siemens systems in 96% of patients, with 93% from 1.5T scanners and the remainder from 3T scanners. During a median follow‐up of 59 months (IQR, 25–79), 35 (15.8%) patients developed PSE. Table 1 shows descriptive statistics for the study cohort, stratified by their PSE status. On unadjusted analyses (Table 1), patients who developed PSE had significantly higher SeLECT score, were more likely to have ASyS, and had larger stroke volume.
TABLE 1.
Patient and clinical characteristics, stratified by whether the patient had post‐stroke epilepsy.
| All patients | Post‐stroke epilepsy | No post‐stroke epilepsy | p‐value | |
|---|---|---|---|---|
| N | 221 | 35 | 186 | |
| Age (years), mean (SD) | 63.5 (13.8) | 65.1 (14.6) | 63.2 (13.7) | 0.479 |
| Female | 112 (50.7%) | 16 (45.7%) | 96 (51.6%) | 0.583 |
| Follow‐up duration (months), median (IQR) | 59 (25, 79) | 65 (34.5, 84) | 58 (22.5, 79) | 0.220 |
| NIHSS at admission, median (IQR) | ||||
| Median (IQR) | 7 (2, 16) | 11 (2, 17.5) | 7 (2, 15) | 0.803 |
| 0–3 | 69 (31.2%) | 11 (31.4%) | 58 (31.2%) | 0.223 |
| 4–10 | 65 (29.4%) | 6 (17.1%) | 59 (31.7%) | |
| 11+ | 84 (38.0%) | 18 (51.4%) | 66 (35.5%) | |
| Missing | 3 (1.4%) | 0 (0.0%) | 3 (1.6%) | |
| LAA | 47 (21.3%) | 7 (20.0%) | 40 (21.5%) | 1.000 |
| Early seizure | 25 (11.3%) | 11 (31.4%) | 14 (7.5%) | < 0.001 |
| Cortical involvement | 173 (78.3%) | 29 (82.9%) | 144 (77.4%) | 0.655 |
| MCA | 152 (68.8%) | 25 (71.4%) | 127 (68.3%) | 0.843 |
| SeLECT score, median (IQR) | 4 (3, 5) | 5 (4, 6) | 4 (3, 5) | < 0.001 |
| Stroke volume (mL), median (IQR) | 5.9 (1.5, 22.6) | 11.0 (5.0, 69.8) | 5.0 (1.4, 20.2) | 0.022 |
Abbreviations: IQR, Interquartile Range; LAA, Large Artery Atherosclerosis; MCA, Middle Cerebral Artery; NIHSS, National Institute of Health Stroke Scale; %, Percentage; SeLECT, Severity of stroke, Large artery atherosclerosis, Early seizure, Cortical Involvement, middle cerebral artery Territory; SD, Standard Deviation
Table 2 shows the results of the models we examined. The original SeLECT score had a C‐index of 0.669 in our sample. Our refit SeLECT model had an optimism‐corrected C‐index of 0.642. Stoke volume remained an independently significant predictor of PSE in all models. Adding stroke volume to the SeLECT model resulted in an optimism‐corrected C‐index of 0.656. Keeping stroke volume while removing only cortical involvement or MCA territory resulted in an optimism‐corrected C‐index of 0.664 and 0.668, respectively. Keeping stroke volume but removing both cortical involvement and MCA resulted in an optimism‐corrected C‐index of 0.679. Table 3 presents the differences in optimism‐corrected C‐index values for each pair of models, along with 95% confidence intervals. Given that all 95% confidence interval differences contained the null value of 0, we conclude that none of the models had significantly higher or lower optimism‐corrected C‐index values. Calibration was generally good for all models, as shown in Figure 2.
TABLE 2.
Results of multivariable Cox proportional hazard models.
|
SeLECT score C‐index = 0.669 |
SeLECT refit optimism‐corrected C‐index = 0.642 |
SeLECT + stroke volume optimism‐corrected C‐index = 0.656 |
SeLECT + stroke volume– cortical optimism‐corrected C‐index = 0.664 |
SeLECT + stroke volume– MCA optimism‐corrected C‐index = 0.668 |
SeLECT + stroke volume– cortical—MCA optimism‐corrected C‐index = 0.679 |
|||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| HR (95% CI) | p‐value | HR (95% CI) | p‐value | HR (95% CI) | p‐value | HR (95% CI) | p‐value | HR (95% CI) | p‐value | HR (95% CI) | p‐value | |
| 0.669 | 0.642 | 0.656 | 0.664 | 0.668 | 0.679 | |||||||
| NIHSS severity (vs. 0–3) | ||||||||||||
| 4–10 | 1.7 (1.0–3.1) | 0.06 | 0.41 (0.15–1.14) | 0.089 | 0.38 (0.14–1.04) | 0.061 | 0.38 (0.14–1.04) | 0.061 | 0.38 (0.14–1.05) | 0.062 | 0.38 (0.14–1.05) | 0.062 |
| 11+ | 2.7 (1.5–4.9) | < 0.001 | 1.05 (0.48–2.30) | 0.904 | 0.64 (0.26–1.60) | 0.338 | 0.64 (0.26–1.58) | 0.334 | 0.64 (0.25–1.60) | 0.337 | 0.64 (0.26–1.59) | 0.338 |
| LAA | 1.5 (0.9–2.5) | 0.15 | 0.95 (0.41–2.22) | 0.907 | 0.96 (0.41–2.25) | 0.933 | 0.96 (0.41–2.25) | 0.934 | 0.97 (0.42–2.26) | 0.941 | 0.97 (0.42–2.26) | 0.950 |
| Early seizure | 4.8 (2.5–9.3) | < 0.001 | 4.69 (2.22–9.94) | < 0.001 | 5.56 (2.57–12.00) | < 0.001 | 5.56 (2.58–11.99) | < 0.001 | 5.52 (2.57–11.87) | < 0.001 | 5.52 (2.57–11.85) | < 0.001 |
| Cortical involvement | 4.2 (1.9–9.0) | < 0.001 | 1.07 (0.39–2.90) | 0.896 | 1.01 (0.35–2.90) | 0.987 | 1.05 (0.41–2.66) | 0.923 | ||||
| MCA | 1.8 (0.8–4.1) | 0.12 | 1.38 (0.60–3.19) | 0.450 | 1.07 (0.44–2.63) | 0.882 | 1.07 (0.48–2.38) | 0.860 | ||||
| Stroke volume (per 10 mL) | 1.12 (1.04–1.20) | 0.002 | 1.12 (1.04–1.20) | 0.002 | 1.12 (1.04–1.20) | 0.002 | 1.12 (1.04–1.20) | 0.001 | ||||
Abbreviations: CI, Confidence Interval; c‐index, Concordance index; HR, Hazard Ratio, IQR, Interquartile Range; LAA, Large Artery Atherosclerosis; MCA, Middle Cerebral Artery; NIHSS, National Institute of Health Stroke Scale; %, Percentage; SeLECT, Severity of stroke, Large artery atherosclerosis, Early seizure, Cortical Involvement, middle cerebral artery Territory.
TABLE 3.
Difference in optimism‐corrected C‐index of all predictive models.
| Model 1 | Model 2 | Difference in optimism‐corrected C‐statistic (95% CI) |
|---|---|---|
| SeLECT refit | SeLECT + stroke volume | −0.013 (–0.207, 0.042) |
| SeLECT refit | SeLECT + stroke volume—cortical | −0.022 (–0.212, 0.056) |
| SeLECT refit | SeLECT + stroke volume—MCA | −0.026 (–0.209, 0.048) |
| SeLECT refit | SeLECT + stroke volume—cortical—MCA | −0.037 (–0.214, 0.058) |
| SeLECT + stroke volume | SeLECT + stroke volume—cortical | −0.008 (–0.027, 0.061) |
| SeLECT + stroke volume | SeLECT + stroke volume—MCA | −0.012 (–0.032, 0.054) |
| SeLECT + stroke volume | SeLECT + stroke volume—cortical—MCA | −0.023 (–0.047, 0.067) |
| SeLECT + stroke volume—cortical | SeLECT + stroke volume—cortical | −0.004 (–0.064, 0.063) |
| SeLECT + stroke volume—cortical | SeLECT + stroke volume—MCA | −0.015 (–0.042, 0.065) |
| SeLECT + stroke volume—MCA | SeLECT + stroke volume—cortical—MCA | −0.011 (–0.034, 0.062) |
Note: 95% confidence intervals computed through bootstrapping.
Abbreviations: 95% confidence intervals computed through bootstrapping; c‐index: Concordance index; CI, Confidence Interval; MCA, Middle Cerebral Artery; SeLECT, Severity of stroke, Large artery atherosclerosis, Early seizure, Cortical Involvement, middle cerebral artery Territory.
FIGURE 2.
Calibration plots for Cox proportional hazard models predicting seizure‐free status at 1‐year post‐stroke. The gray line represents the ideal, where model‐predicted probabilities are equal to observed probabilities. The black line represents the observations made in our patient sample. The blue line represents the bias‐corrected calibration line, based on bootstrapping.
Given that stroke volume is a statistically significant predictor in all models, we post hoc evaluated a threshold at which it becomes more predictive of PSE. Figure 3 shows ROC curves for stroke volume at 6‐month, 1‐year, and 2‐year post‐stroke. Table 4 shows the AUC, optimal stroke volume cutoff point, sensitivity, specificity, PPV, and NPV for each time point. AUC was higher for earlier time points. Stroke volume cutoff points varied by time point and ranged from 11.6 to 16.7 mL.
FIGURE 3.
Area under the curve for stroke volume predicting post‐stroke seizure recurrence.
TABLE 4.
Area under the curve (AUC), stroke volume cutoff point, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) at 6‐month, 1‐year, and 2‐year.
| Time | AUC (95% CI) | Stroke volume cutoff point (mL) | Sensitivity | Specificity | PPV | NPV |
|---|---|---|---|---|---|---|
| 6‐month | 0.69 (0.54, 0.84) | 16.7 | 0.50 | 0.73 | 0.13 | 0.95 |
| 1‐year | 0.67 (0.54, 0.80) | 11.6 | 0.54 | 0.67 | 0.17 | 0.92 |
| 2‐year | 0.61 (0.49, 0.73) | 14.7 | 0.43 | 0.69 | 0.20 | 0.87 |
4. Discussion
We found that acute stroke volume is a significant and independent predictor of PSE. There was a 12% increased risk of PSE for every 10 mL increase in stroke volume, which remained a significant predictor across all models. The predictive performance (discrimination and calibration) of the SeLECT model, when the stroke volume substitutes for both cortical and MCA involvement, was comparable to that of the original SeLECT score. This lack of significant difference in model performance, and stroke volume remaining a strong predictor across various models, suggests that it can reliably serve as a practical, automated proxy for the current neuroimaging components of the SeLECT score.
Automated DWI‐based lesion segmentation has now become feasible and accurate, achieving a Dice coefficient in the range of 0.81−0.85 [6, 7]. These methods are robust to variations in scanner type and image quality. They have been validated in multicenter cohorts [6, 18]. Our DWI‐based volume analysis algorithm achieved a Dice coefficient on our internal dataset (0.87) and the external dataset (0.82), comparable to those of the Artifical Intelligence (AI) tools in the literature. Additionally, segmentation error is unlikely to bias our results meaningfully, as measurement error in stroke volume would tend to attenuate effect estimates rather than create spurious associations. Given the strong and consistent association across all models, the predictive value of stroke volume appears robust. The availability of open‐source, state‐of‐the‐art, and reliable tools for ischemic stroke volume calculation enables the calculation to be performed at scale for large datasets [7]. In addition, researchers have developed validated, automated methods to extract NIHSS scores from EHRs using natural language processing [19]. With both NIHSS and stroke volume now extractable at scale, most of the core predictors for PSE risk can be assembled without manual review, enabling efficient, reproducible, large‐scale research and clinical implementation.
A very recent systematic review found that the SeLECT score's application in various geographical locations, including China, Turkey, and Japan, has limitations in its discrimination of PSE and is yet to be tested in an exclusively US population [4]. Our study, to our knowledge, is the first application of the SeLECT score in a US cohort, demonstrating a more than 10% loss in its discriminatory power for predicting PSE (c‐index of 0.77 vs. 0.64) [2]. This difference may reflect racial differences in PSE risk as shown in several extensive studies. The Northern Manhattan Study found that PSE risk was two ‐to three‐fold higher in Hispanics and non‐Hispanic Blacks compared to non‐Hispanic Whites [20]. Similarly, the Cardiovascular Health Study and the Atherosclerosis Risk in Communities (ARIC) study have found a higher risk of PSE among Black participants, with risk factors like hypertension and diabetes exerting a more substantial effect in these minority populations [21, 22].
When placed in the context of chronic lesional studies, our finding suggests that the volume of brain injury plays a critical role in epileptogenesis from the acute to the chronic pathological stages of stroke [5]. It is noteworthy that, like the CAVE score, it seems that the minimum ischemic stroke volume that significantly increases PSE is more than 10 mL [3].
Limitations of this study include the retrospective, single‐center approach. Additionally, more severe stroke patients and those with cortical involvement on CT scan did not receive MRI and hence, were excluded from the current analysis. To overcome this selection bias and increase the generalizability of the stroke volume as a robust PSE predictor, we plan to evaluate it using automated tools that use CT scans [23]. Additionally, future multicenter studies with larger sample sizes are needed to validate our findings and create new prognostic models that utilize stroke volume to predict PSE [24].
Overall, stroke volume provides a more precise and quantitative measure of acute brain injury than binary regional designations. Our findings indicate that automated DWI‐based volume estimation can serve as a practical alternative to manually abstracted cortical and MCA involvement. Automated segmentation is feasible and sufficiently accurate, and it substantially reduces the burden of manual data extraction. Incorporating stroke volume into prognostic modeling may enhance the efficiency and objectivity of PSE risk prediction and, with further validation, could become a preferred imaging‐based predictor.
Funding
The authors have nothing to report.
Acknowledgments and Disclosure
The authors have nothing to report.
Conflicts of Interest
The authors declare no conflicts of interest.
1.
TABLE A1.
Patient and clinical characteristics for the study cohort (included) and excluded patients.
| All Patients | Included | Excluded | p‐value | |
|---|---|---|---|---|
| N | 375 | 221 | 154 | |
| Age (years), mean (SD) | 64.6 (14.2) | 63.5 (13.8) | 66.2 (14.7) | 0.068 |
| Female | 176 (79.6%) | 112 (50.7%) | 64 (41.6%) | 0.093 |
| Follow‐up Duration (months), median (IQR) | 45 (9.5, 76.5) | 59 (25, 79) | 21.5 (2, 52) | <0.001 |
| NIHSS at Admission, median (IQR) | ||||
| Median (IQR) | 9 (3, 17.5) | 7 (2, 16) | 11 (4, 20) | 0.002 |
| 0–3 | 101 (45.7%) | 69 (31.2%) | 32 (20.8%) | 0.032 |
| 4–10 | 105 (47.5%) | 65 (29.4%) | 40 (26.0%) | |
| 11+ | 161 (72.9%) | 84 (38.0%) | 77 (50.0%) | |
| Missing | 8 (3.6%) | 3 (1.4%) | 5 (3.2%) | |
| LAA | 78 (35.3%) | 47 (21.3%) | 31 (20.1%) | 0.897 |
| Early Seizure | 53 (24.0%) | 25 (11.3%) | 28 (18.2%) | 0.071 |
| Cortical Involvement | 267 (120.8%) | 140 (63.3%) | 127 (82.5%) | <0.001 |
| MCA | 257 (116.3%) | 144 (65.2%) | 113 (73.4%) | 0.113 |
| SeLECT Score, median (IQR) | 4 (3, 5) | 3 (2, 5) | 5 (3, 5) | <0.001 |
FIGURE A1.
Examples of DWI manual segmentation and automatically generated lesion masks. Lesion masks shown by green (true positive agreement), red (false negative), and blue (false positive) by the algorithm. The top row (a,b) shows the high 3D Dice of 0.93 while the bottom row (c,d) shows low 3D Dice of 0.61 with a greater extent of false negative when the abnormality is less pronounced.
References
- 1. Sen A., Jette N., Husain M., and Sander J. W., “Epilepsy in Older People,” Lancet 395, no. 10225 (2020): 735–748, 10.1016/S0140-6736(19)33064-8. [DOI] [PubMed] [Google Scholar]
- 2. Galovic M., Döhler N., Erdélyi‐Canavese B., et al., “Prediction of Late Seizures After Ischaemic Stroke With a Novel Prognostic Model (The SeLECT Score): A Multivariable Prediction Model Development and Validation Study,” Lancet Neurology 17, no. 2 (2018): 143–152, 10.1016/S1474-4422(17)30404-0. [DOI] [PubMed] [Google Scholar]
- 3. Haapaniemi E., Strbian D., Rossi C., et al., “The CAVE Score for Predicting Late Seizures After Intracerebral Hemorrhage,” Stroke 45, no. 7 (2014): 1971–1976, 10.1161/STROKEAHA.114.004686. [DOI] [PubMed] [Google Scholar]
- 4. Ayaz A., Hashmi S. A., Ibrahim M., Sharma S., Abbas M. A., and Zawar I., “To Select or Not to Select: Evaluating the Effectiveness of SeLECT Score in Predicting the Risk of Post‐Stroke Seizures: A Systematic Review,” Epilepsy & Behavior 173 (2025): 110794, 10.1016/j.yebeh.2025.110794. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Schaper F., Nordberg J., Cohen A. L., et al., “Mapping Lesion‐Related Epilepsy to a Human Brain Network,” JAMA Neurology 80, no. 9 (2023): 891, 10.1001/jamaneurol.2023.1988. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Wong K. K., Cummock J. S., Li G., et al., “Automatic Segmentation in Acute Ischemic Stroke: Prognostic Significance of Topological Stroke Volumes on Stroke Outcome,” Stroke 53, no. 9 (2022): 2896–2905, 10.1161/STROKEAHA.121.037982. [DOI] [PubMed] [Google Scholar]
- 7. de la Rosa E., Reyes M., Liew S. L., et al., “DeepISLES: A Clinically Validated Ischemic Stroke Segmentation Model From the ISLES'22 Challenge,” Nature Communications 16, no. 1 (2025): 7357, 10.1038/s41467-025-62373-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Fisher R. S., Cross J. H., French J. A., et al., “Operational Classification of Seizure Types by the International League Against Epilepsy: Position Paper of the ILAE Commission for Classification and Terminology,” Epilepsia 58, no. 4 (2017): 522–530, 10.1111/epi.13670. [DOI] [PubMed] [Google Scholar]
- 9. Sokola M., Thoomukuntla B., Lapin B., et al., “Abstract WMP50: Lesion‐Symptom Mapping for Post‐Stroke Depression,” Stroke 54, no. Suppl_1 (2023): AWMP50, 10.1161/str.54.suppl_1.WMP50. [DOI] [Google Scholar]
- 10. Ronneberger O., Fischer P., and Brox T., “U‐Net: Convolutional Networks for Biomedical Image Segmentation,” in Medical Image Computing and Computer‐Assisted Intervention—MICCAI 2015, ed. Navab N, Hornegger J, Wells WM, and Frangi AF (Springer International Publishing, 2015), 234–241, 10.1007/978-3-319-24574-4_28. [DOI] [Google Scholar]
- 11. Yu Y., Xie Y., Thamm T., et al., “Use of Deep Learning to Predict Final Ischemic Stroke Lesions From Initial Magnetic Resonance Imaging,” JAMA Network Open 3, no. 3 (2020): e200772, 10.1001/jamanetworkopen.2020.0772. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Hernandez Petzsche M. R., de la Rosa E., Hanning U., et al., “ISLES 2022: A Multi‐Center Magnetic Resonance Imaging Stroke Lesion Segmentation Dataset,” Scientific Data 9, no. 1 (2022): 762, 10.1038/s41597-022-01875-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Kamarudin A. N., Cox T., and Kolamunnage‐Dona R., “Time‐Dependent ROC Curve Analysis in Medical Research: Current Methods and Applications,” BMC Medical Research Methodology [Electronic Resource] 17, no. 1 (2017): 53, 10.1186/s12874-017-0332-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Youden W. J., “Index for Rating Diagnostic Tests,” Cancer 3, no. 1 (1950): 32–35, 10.1002/1097-0142(1950)3:1<32::aid-cncr2820030106>3.0.co;2-3. [DOI] [PubMed] [Google Scholar]
- 15. R Development Core Team , R: A Language and Environment for Statistical Computing Published online 2020, http://www.r‐project.org/.
- 16. Payabvash S., Taleb S., Benson J. C., and McKinney A. M., “Acute Ischemic Stroke Infarct Topology: Association With Lesion Volume and Severity of Symptoms at Admission and Discharge,” American Journal of Neuroradiology 38, no. 1 (2017): 58–63, 10.3174/ajnr.A4970. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Furlanis G., Ajčević M., Stragapede L., et al., “Ischemic Volume and Neurological Deficit: Correlation of Computed Tomography Perfusion With the National Institutes of Health Stroke Scale Score in Acute Ischemic Stroke,” Journal of Stroke and Cerebrovascular Diseases 27, no. 8 (2018): 2200–2207, 10.1016/j.jstrokecerebrovasdis.2018.04.003. [DOI] [PubMed] [Google Scholar]
- 18. Wu O., Winzeck S., Giese A. K., et al., “Big Data Approaches to Phenotyping Acute Ischemic Stroke Using Automated Lesion Segmentation of Multi‐Center Magnetic Resonance Imaging Data,” Stroke 50, no. 7 (2019): 1734–1741, 10.1161/STROKEAHA.119.025373. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Fernandes M., Westover M. B., Singhal A. B., and Zafar S. F., “Automated Extraction of Stroke Severity From Unstructured Electronic Health Records Using Natural Language Processing,” Journal of the American Heart Association 13, no. 21 (2024): e036386, 10.1161/JAHA.124.036386. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Choi H., Thacker E. L., Liu M., et al., “Racial/Ethnic Differences in the Association of Incident Stroke With Late Onset Epilepsy: The Northern Manhattan Study,” Epilepsia 65, no. 12 (2024): 3561–3570, 10.1111/epi.18156. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21. Choi H., Pack A., Elkind M. S. V., Longstreth W. T., Ton T. G. N., and Onchiri F., “Predictors of Incident Epilepsy in Older Adults: The Cardiovascular Health Study,” Neurology 88, no. 9 (2017): 870–877, 10.1212/WNL.0000000000003662. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Johnson E. L., Krauss G. L., Lee A. K., et al., “Association Between Midlife Risk Factors and Late‐Onset Epilepsy: Results From the Atherosclerosis Risk in Communities Study,” JAMA Neurology 75, no. 11 (2018): 1375–1382, 10.1001/jamaneurol.2018.1935. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Qiu W., Kuang H., Teleg E., et al., “Machine Learning for Detecting Early Infarction in Acute Stroke With Non‐Contrast‐Enhanced CT,” Radiology 294, no. 3 (2020): 638–644, 10.1148/radiol.2020191193. [DOI] [PubMed] [Google Scholar]
- 24. Riley R. D., Ensor J., Snell K. I. E., et al., “Calculating the Sample Size Required for Developing a Clinical Prediction Model,” BMJ 368 (2020): m441, 10.1136/bmj.m441. [DOI] [PubMed] [Google Scholar]