Adaptive Multi-Resolution Decomposition And Dual-Conditioned Generative Adversarial Networks For Patient-Independent Eeg Seizure Detection

Abstract

Existing VMD-cGAN approaches have demonstrated competitive performance in EEG-based seizure detection by combining Variational Mode Decomposition (VMD) with Conditional Generative Adversarial Networks (cGANs) for frequency-aware data augmentation. However, such methods are constrained by a fixed decomposition order K, a single class-conditioning signal, and a patient-specific evaluation protocol that limits cross-patient generalisation. This paper presents an Adaptive Multi-Resolution Decomposition and Dual-Conditioned GAN (AMRD-DCGAN) framework that overcomes these limitations through three principal advances: (1) an adaptive VMD order selection mechanism that automatically determines the optimal number of intrinsic mode functions (IMFs) per patient based on a spectral entropy criterion, replacing the fixed K=5 with a data-driven K ∈ [3, 8]; (2) a dual-conditioning architecture that simultaneously conditions the generator on both class label and patient identifier, enabling cross-patient generalisation of synthetic EEG synthesis; and (3) a multi-scale temporal discriminator that enforces structural fidelity at multiple temporal resolutions. Evaluated on the CHB-MIT Scalp EEG Database under a strict leave-one-patient-out (LOPO) cross-validation protocol, the proposed AMRD-DCGAN achieves 98.61% accuracy, 97.82% sensitivity, 99.13% specificity, and 94.27% F1-score, outperforming existing VMD-cGAN methods by 0.77%, 1.09%, 0.61%, and 1.34% respectively. The improvements are confirmed statistically significant at the 99% confidence level. The framework addresses the patient-independence limitation of current generative augmentation methods and establishes a new state of the art on the CHB-MIT benchmark.