Zmienność sygnału EEG jako biomarker do klasyfikacji stanów napadowych: Walidacja klasyfikatora Random Forest na poziomie pacjenta z użyciem bazy danych Uniwersytetu Bonn
EEG Signal Variance as a Biomarker for Ictal State Classification: Subject-Level Validation of a Random Forest Classifier using the University of Bonn Dataset
W skrócie
[Preprint - wstępne wyniki] Badacze przetestowali czy sztuczna inteligencja może automatycznie rozpoznawać napady epilepsji na podstawie zapisu EEG. Wykazali, że sygnały mózgu podczas napadu mają wyraźnie większą zmienność niż w stanach bez napadu, a specjalny algorytm komputerowy (Random Forest) potrafił prawidłowo odróżnić napady od normalnych zapisów mózgu. To jest ważne, bo w wielu krajach brakuje specjalistów do interpretacji zapisów EEG, a automatyczne rozpoznawanie mogłoby pomóc milionom pacjentów na całym świecie.
Oryginalny abstract (angielski)
Abstract Background Epilepsy affects an estimated 50 million people globally and imposes a disproportionate burden on low- and middle-income countries where specialist electroencephalographic (EEG) interpretation is scarce. Automated machine learning approaches to EEG analysis hold significant promise for scalable seizure detection. This study evaluated two pre-specified hypotheses: (1) that ictal EEG epochs exhibit significantly elevated signal variance compared to non-ictal states, as quantified by formal statistical testing; and (2) that a Random Forest (RF) classifier trained on raw EEG time-series features can achieve high internal validation accuracy in differentiating active epileptic seizures from recordings obtained in the epileptogenic zone during seizure-free intervals. Methods This retrospective diagnostic accuracy study adhered to the STARD 2015 and TRIPOD + AI reporting guidelines. The publicly available University of Bonn Epileptic Seizure Recognition dataset (N = 11,500 epochs from 500 subjects; five classes) was analysed. To prevent data leakage, all model development and evaluation employed subject-level data partitioning: an 80/20 GroupShuffleSplit (195 training subjects, 49 test subjects, zero overlap) and 10-fold subject-stratified GroupKFold cross-validation. Signal variance was compared across clinical states using the Kruskal–Wallis H test with pairwise Mann–Whitney U post-hoc comparisons (Bonferroni-corrected α = 0.005) and Cliff’s delta effect sizes. The RF classifier (100 trees; Gini criterion) was benchmarked against Decision Tree and Logistic Regression baselines. Performance was evaluated by accuracy, sensitivity, specificity, positive and negative predictive values, F1-score, AUC-ROC, Average Precision, Brier score, and bootstrap 95% confidence intervals (2,000 resamples). Results The Kruskal–Wallis test confirmed statistically significant variance differences across all five clinical states (H = 6,078.47, p