Abstract

The natural history of seizure risk after brain tumor resection is not well understood. Identifying seizure-naive patients at highest risk for postoperative seizure events remains a clinical need. In this study, the authors sought to develop a predictive modeling strategy for anticipating postcraniotomy seizures after brain tumor resection.The IBM Watson Health MarketScan Claims Database was canvassed for antiepileptic drug (AED)- and seizure-naive patients who underwent brain tumor resection (2007-2016). The primary event of interest was short-term seizure risk (within 90 days postdischarge). The secondary event of interest was long-term seizure risk during the follow-up period. To model early-onset and long-term postdischarge seizure risk, a penalized logistic regression classifier and multivariable Cox regression model, respectively, were built, which integrated patient-, tumor-, and hospitalization-specific features. To compare empirical seizure rates, equally sized cohort tertiles were created and labeled as low risk, medium risk, and high risk.Of 5470 patients, 983 (18.0%) had a postdischarge-coded seizure event. The integrated binary classification approach for predicting early-onset seizures outperformed models using feature subsets (area under the curve [AUC] = 0.751, hospitalization features only AUC = 0.667, patient features only AUC = 0.603, and tumor features only AUC = 0.694). Held-out validation patient cases that were predicted by the integrated model to have elevated short-term risk more frequently developed seizures within 90 days of discharge (24.1% high risk vs 3.8% low risk, p < 0.001). Compared with those in the low-risk tertile by the long-term seizure risk model, patients in the medium-risk and high-risk tertiles had 2.13 (95% CI 1.45-3.11) and 6.24 (95% CI 4.40-8.84) times higher long-term risk for postdischarge seizures. Only patients predicted as high risk developed status epilepticus within 90 days of discharge (1.7% high risk vs 0% low risk, p = 0.003).The authors have presented a risk-stratified model that accurately predicted short- and long-term seizure risk in patients who underwent brain tumor resection, which may be used to stratify future study of postoperative AED prophylaxis in highest-risk patient subpopulations.

View details for DOI 10.3171/2022.1.FOCUS21751

View details for PubMedID 35364580