Leakage-Aware Multi-Representation EEG Fusion with Confidence-Gated Learnable MLP for Robust Epileptic Seizure Detection

Automated epileptic seizure detection from Electroencephalogram (EEG) signals remains challenging due to the non-stationary nature of EEG activity, strong inter-patient variability, and the tendency of window-based evaluation protocols to produce overly optimistic results when temporal or subject-level data leakage is not properly controlled. This study proposes a leakage-aware multimodal deep learning framework for robust EEG seizure detection on the CHB-MIT scalp EEG dataset. Each EEG window is represented using three complementary views: time-domain waveform images, Short-Time Fourier Transform (STFT) spectrograms, and Continuous Wavelet Transform (CWT) scalograms. Eight pretrained convolutional neural network (CNN) backbones, including DenseNet201, EfficientNetB0, GoogleNet, MobileNetV2, ResNet18, ResNet50, VGG16, and VGG19, are independently trained as modality-specific experts. Instead of fusing high-dimensional feature maps, the proposed framework performs compact probability-level fusion by learning from the posterior outputs of frozen experts through a lightweight multilayer perceptron (MLP) enhanced with a confidence-gated decision mechanism. The main contribution of this work is a unified seizure detection framework that combines complementary EEG representations, low-complexity learnable probability-level fusion, confidence-adaptive decision refinement, and rigorous leakage-safe validation. Three evaluation scenarios were implemented: patient-specific evaluation with Group ID-based splitting and fallback handling, patient-wise 5-fold cross-validation, and Leave-One-Patient-Out (LOPO) validation. These protocols explicitly reduce temporal, recording-level, and patient-level leakage, thereby providing a more reliable assessment of both individualized and cross-patient generalization. Experimental results demonstrate that the proposed confidence-gated fusion consistently outperforms conventional MLP fusion across the evaluated backbones. Among the evaluated backbone models, ResNet50 provided the best overall clinical performance, achieving a mean balanced accuracy of 99.23%, a mean F1-score of 99.12%, the highest mean sensitivity of 99.50%, a mean specificity of 98.96%, and a mean precision of 98.85%. ResNet50 was selected as the best-performing model due to its superior seizure detection capability, as reflected by its highest sensitivity while maintaining comparable performance across all other evaluation metrics. These results demonstrate that adaptive probability-level fusion of waveform, STFT, and CWT representations can provide highly accurate and stable seizure detection while maintaining a leakage-aware evaluation framework. Overall, the proposed framework offers a robust and clinically relevant strategy for automated EEG seizure detection under realistic validation conditions.