Wspomagana diagnoza epilepsji na podstawie sygnałów EEG z wykorzystaniem rekonstrukcji z podwójnym progiem ICEEMDAN i DW-ACFS

Abstract Objective. Electroencephalography (EEG) is the primary neurological method for epilepsy detection, but it is time-consuming. EEG signals are non-stationary, nonlinear, and noisy, making epileptic EEG recognition challenging. To address mode mixing and reconstruction distortion in traditional mode decomposition methods, as well as insufficient redundancy suppression in feature selection, this paper proposes an assisted diagnosis method for epileptic EEG signals based on dual-threshold reconstruction using improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and dynamic weighted adversarial collaborative feature selection (DW-ACFS). Approach. First, ICEEMDAN decomposes the EEG data, and a dual-threshold criterion based on correlation coefficient and energy ratio selects effective intrinsic mode function (IMF) components for signal reconstruction. Then, 18-dimensional features comprising time-domain, frequency-domain, and nonlinear dynamic characteristics are extracted from the reconstructed signals. This paper proposes the DW-ACFS method: feature importance is evaluated via dual pathways using maximum relevance minimum redundancy (mRMR) and F-statistic, fusion weights are adaptively determined based on validation set performance, and an adversarial redundancy penalty mechanism suppresses highly collinear features. Finally, the feature importance ranking obtained by DW-ACFS is used to perform three-class classification (healthy, interictal, ictal) on the resulting feature dataset using classifiers. Main Results. Experiments on the Bonn epilepsy EEG dataset using support vector machine (SVM), random forest, and extreme gradient boosting (XGBoost) classifiers show that the feature subset selected by DW-ACFS achieves higher classification accuracy than those selected by either mRMR or F-statistic alone. With 13 features selected, XGBoost achieves accuracy, recall, and specificity of 98.12%, 98.13%, and 99.06%, respectively, outperforming SVM and random forest. Significance. Experimental results demonstrate that the combination of ICEEMDAN dual-threshold reconstruction and DW-ACFS enables high-precision, low-complexity recognition of epileptic EEG signals, providing an effective assisted diagnosis method and efficient solution for clinical auxiliary diagnosis.