Exécution de notre outil de synthèse sur un nouvel article
Vérifiez l’article ci-dessous, puis lancez sa synthèse.
Predicting ICU Seizures Using 1 hour of cEEG
Souvik Phadikar (Georgia Institute of Technology, University of Wisconsin–Madison), Santiago Philibert‐Rosas (University of Wisconsin–Madison, Washington University in St. Louis), Steven E. Haworth (University of Wisconsin–Madison), Safoora Fatima (University of Wisconsin–Madison, SIU Medicine), Maximilian Grobbelaar (University of Wisconsin–Madison), Chenxi Sun (Harvard University, SIU Medicine), Jin Jing (Beth Israel Deaconess Medical Center, Harvard University), Sahar Zafar (Beth Israel Deaconess Medical Center), Jennifer A. Kim (Yale University), Brandon Westover (Beth Israel Deaconess Medical Center, Harvard University), Aaron F. Struck (University of Wisconsin–Madison, Washington University in St. Louis)
Résumé : PURPOSE: Continuous electroencephalography (cEEG) is the gold standard for detecting electrographic seizures in the intensive care unit (ICU), but limited resources limit prolonged monitoring for all patients. We developed an automated, interpretable pipeline that uses a single hour of cEEG to estimate near-term seizure risk across multiple prediction windows. METHODS: We retrospectively studied 50 adults who underwent ICU cEEG monitoring at a single tertiary center. From a 1-hour segment, we computed automated interictal-ictal continuum (IIC) burden probabilities for lateralized periodic discharges (LPDs), generalized periodic discharges (GPDs), and lateralized rhythmic delta activity (LRDA), along with quantitative background EEG features reflecting signal organization and hemispheric asymmetry. Segments were labeled by seizure occurrence within the subsequent 1, 2, 4, 8, 12, or 24 hours. Support vector machine (SVM) and bagged-tree classifiers were trained using clinical-only, background-only, and combined feature sets with a 70/30 patient-level train-test split and cross-validation within the training set. RESULTS: For the 4-hour horizon, the combined-feature SVM achieved 76% accuracy and the lowest Brier score (0.25), outperforming clinical-only (72%) and background-only (70%) models. Across horizons, SVM accuracy declined from 94% (1 hour) to 79% (24 hours) with increasing Brier scores, indicating reduced calibration at longer intervals. AUC-receiver operating characteristic increased with wider prediction windows, consistent with higher seizure prevalence at longer lead times. CONCLUSIONS: Integrating automated IIC probabilities with quantitative background EEG features enables moderate, interpretable near-term seizure prediction from a single hour of ICU cEEG. Initial external validation on an independent multicenter ICU cohort supports generalizability, although broader prospective and multicenter evaluation are needed before clinical deployment.