A representation-level wavelet–EMD fusion framework for spectral analysis and detection of epileptiform patterns in pediatric EEG: computational validation with a roadmap for external benchmarking

This is the authors' abstract. We don't add key points for this paper.

Background Pediatric EEG analysis for epilepsy detection is challenged by the non-stationary and developmentally heterogeneous nature of epileptiform activity. Conventional Fourier approaches provide limited time–frequency localization, whereas single-domain methods such as the wavelet transform and empirical mode decomposition (EMD) each exhibit complementary weaknesses. Methods We developed a representation-level wavelet–EMD fusion framework that integrates wavelet and EMD spectra before feature extraction. The pipeline included adaptive wavelet denoising, energy-criterion EMD, weighted time–frequency fusion, extraction of eight spectral–temporal features, and classification using a soft-voting ensemble of a radial basis function support vector machine, random forest, and neural network. Evaluation was performed on 2,000 synthetic pediatric EEG segments spanning canonical epileptiform morphologies and multiple signal-to-noise ratios. A feature-level fusion variant was implemented as an ablation comparator to isolate the effect of fusion strategy while preserving identical feature extraction, classifier architecture, and validation procedures. Results Across 10-fold stratified cross-validation with 100 Monte Carlo replications, the proposed framework achieved 94.6 ± 0.6% accuracy, 93.7 ± 0.8% sensitivity, 95.5 ± 0.5% specificity, F1 score of 0.943 ± 0.007, and AUROC of 0.971 (95% CI, 0.962–0.978). Performance exceeded standalone wavelet, EMD, and Fourier baselines (all p < 0.001). Representation-level fusion also outperformed feature-level fusion by 1.4 percentage points in accuracy, with the greatest gains observed for polyspike and slow-wave morphologies. On the synthetic corpus the framework was competitive with EEGNet and a 1D convolutional network while retaining explicit time–frequency interpretability. In an external feasibility validation on the CHB-MIT pediatric scalp EEG corpus (14 subjects, leave-one-subject-out), representation-level fusion attained the highest performance of all methods (AUROC 0.901, accuracy 86.7%), confirming that the fusion strategy transfers to real recordings, with performance decreasing as expected. Conclusion Representation-level wavelet–EMD fusion provides an interpretable and transparent computational framework for analyzing non-stationary pediatric epileptiform EEG whose advantage over classical time–frequency baselines persists, at feasibility scale, on real CHB-MIT recordings and remains competitive with deep-learning baselines. Because the primary evaluation is synthetic and the external validation is a feasibility-scale demonstration, these figures should be read as evidence of methodological viability and external transportability rather than definitive clinical performance; validation on larger, multi-center and prospective pediatric cohorts is the essential next step.