The paper introduces the KAN-EEG model, a novel architecture based on the Kolmogorov-Arnold Network (KAN) for efficient seizure detection using Electroencephalogram (EEG) signals. The KAN model is designed to replace traditional Multilayer Perceptron (MLP) models, offering improved accuracy, efficiency, and interpretability. The study evaluates the KAN-EEG model on three diverse datasets from the USA, Europe, and Oceania, demonstrating its ability to generalize across different geographical regions. The empirical findings show that while both KAN and MLP models perform well in seizure detection, the KAN model exhibits superior out-of-sample generalization and computational efficiency. The KAN architecture's resilience to model size reduction and shallow network configurations further highlights its versatility and potential for real-world applications. The study underscores the promising role of KANs in medical diagnostics, particularly in epilepsy detection, and suggests that KAN-EEG could be a viable alternative to MLP-based models.The paper introduces the KAN-EEG model, a novel architecture based on the Kolmogorov-Arnold Network (KAN) for efficient seizure detection using Electroencephalogram (EEG) signals. The KAN model is designed to replace traditional Multilayer Perceptron (MLP) models, offering improved accuracy, efficiency, and interpretability. The study evaluates the KAN-EEG model on three diverse datasets from the USA, Europe, and Oceania, demonstrating its ability to generalize across different geographical regions. The empirical findings show that while both KAN and MLP models perform well in seizure detection, the KAN model exhibits superior out-of-sample generalization and computational efficiency. The KAN architecture's resilience to model size reduction and shallow network configurations further highlights its versatility and potential for real-world applications. The study underscores the promising role of KANs in medical diagnostics, particularly in epilepsy detection, and suggests that KAN-EEG could be a viable alternative to MLP-based models.