Abstracts

Rationale:
Ischemic stroke survivors risk developing post-ischemic stroke epilepsy (PISE). Current PISE models often neglect the risk of post-stroke death when modeling PISE potential. Some risk factors for both death and PISE after stroke overlap, so it is possible that these PISE models identify patients at high risk for both outcomes, which may mislead anti-epileptogenic trial enrollment and patient communication. We aim to improve a well-validated PISE risk model – SeLECT (Severity of stroke, Large-artery atherosclerosis etiology, Early seizure, Cortical involvement, Territory of middle cerebral artery) by simultaneously estimating both PISE and death risks.

Methods:
We reviewed clinical notes of ischemic stroke patients at Yale-New Haven Hospital (2014-2023). Inclusion criteria: aged ≥ 18 years, had neurological follow-up >7 days, and had neuroimaging and EEG assessment ≤ 7 days post-stroke. Exclusion criteria: histories of seizures or diseases known to be epileptogenic. Time from stroke to PISE/death was identified per patient. We first evaluated SeLECT performance in our cohort. We then applied the Cox model (cause-specific hazard) to identify PISE and death predictors. Using these predictors, we fit a random survival forest with a competing risk framework to construct a predictive model that estimates the risk for both PISE and death per patient.

Results:
We included 230 patients in the training and 50 in the testing cohort. Although SeLECT was predictive of PISE (1^st-year AUC=65%), the model was also prone to identifying patients with high probabilities for death until SeLECT ≥ 6. (Fig 1A-B). Fig 1C demonstrated that admission National Institutes of Health Stroke Scale (NIHSS) was associated with both PISE and death. To construct a model that predicts who are likely to survive and develop PISE, we identified novel quantitative PISE biomarkers (e.g., NIHSS at 72h post-stroke, infarct volume, peak epileptiform abnormality burden, power/rhythmicity asymmetries) and death covariates (Fig 1C). Finally, using SeLECT score, novel PISE predictors, and competing risk (i.e., death) covariates, we constructed a SeLECT-Quantitative-Competing Risk (“SeQuantCR”) model that estimated a patient’s risks for both PISE and death events (cross-validated 1^st-year PISE AUC=73%, Death AUC=76%; test PISE AUC=72%, Death AUC=77%; Fig 2A). The benefits of SeQuantCR as compared to SeLECT were the most prominent amongst high-risk patients (i.e., SeLECT ≥ 4, Fig 2B). Notably, the SeQuantCR model allowed patient stratification into four targeted groups: 1) Event-Free: low PISE and death risks; 2) PISE: high PISE and low death risks; 3) Death: high death and low PISE risks; and 4) Some-Bad-Event: high PISE and death risks (Fig 2C).

Conclusions:
Competing risk analysis facilitates a simultaneous estimation of both PISE and death risks. This allows clinicians and researchers to better focus trial interventions and patient communications based on the predicted probabilities of the outcome of greatest interest.

Funding: JAK received funding from the NINDS (R25N065743, K23NS112596-01A1, R01NS117904), the American Academy of Neurology Clinical Research Training Scholarship, the American Heart Association, and the Bee Foundation.