KAN-EEG: towards replacing backbone-MLP for an effective seizure detection system

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Medical Diagnostics

Background:

• Artificial intelligence (AI) research is evolving, with multilayer perceptron (MLP) models forming the basis of many AI systems.
• Epilepsy seizure detection using electroencephalogram (EEG) data is a critical application area for AI.
• Existing AI architectures may face challenges in generalizing across diverse datasets.

Purpose of the Study:

• To introduce and evaluate the Kolmogorov-Arnold network (KAN) as a novel AI architecture.
• To develop and test a KAN-electroencephalogram (EEG) model for efficient seizure detection.
• To assess the KAN model's generalization capabilities across varied datasets and its resilience to architectural changes.

Main Methods:

• Implementation of a KAN-based neural network architecture tailored for EEG data analysis.
• Rigorous testing and validation of the KAN-EEG model on three distinct epilepsy seizure datasets from different geographical regions (USA, Europe, Oceania).
• Comparative performance evaluation against traditional MLP architectures, focusing on out-of-sample generalization and model size reduction.

Main Results:

• The KAN model demonstrated high-level out-of-sample generalization across diverse, multi-regional EEG datasets.
• KAN architecture showed resilience to model size reduction and shallow network configurations, preventing overfitting.
• Both KAN and MLP architectures performed commendably, but KAN exhibited superior adaptability and efficiency.

Conclusions:

• The Kolmogorov-Arnold network (KAN) represents a significant advancement in AI architecture, offering enhanced generalization capabilities.
• The KAN-EEG model is a promising tool for efficient and adaptable seizure detection in epilepsy patients.
• KAN's versatility and efficiency position it as a potentially pioneering architecture for critical applications like medical diagnostics.