Deep learning methods for seizure detection in ElectroEncephalograms (EEG) must combine information of EEG electrodes into a single representation vector of reduced dimensionality that it is the input to classifiers. This work analyzes various fusion strategies, focusing on their ability to handle inter- and intra-subject variability. We propose modular architectures incorporating early (raw input) and mid-level (feature representation) channel fusion strategies, with different methods such as concatenation and trainable weighted averaging. The proposed models are validated on the CHB-MIT public database using three levels of generalization: window, seizure and subject. Standard metrics like recall, precision, and F1-score are reported, and results are statistically analyzed to assess model generalization and robustness. Mid-level fusion models, particularly those employing trainable weighted averages, consistently outperform early fusion approaches, achieving higher accuracy and lower variability across metrics. Comparisons with state-of-the-art methods show that the proposed mid-level fusion strategies provide significant improvements, particularly in subject level generalization. This paper highlights the importance of validation protocols that account for temporal dependencies and diverse patient populations.