Evaluating xLSTM for EEG seizure classification

Reliable seizure classification from electroencephalography (EEG) remains a challenging problem due to noisy clinical data, various recording conditions, and the need to capture complex temporal–spatial dependencies. This thesis investigates whether the recently proposed Extended Long Short-Term Memory (xLSTM) architecture offers advantages over established biosignal sequence models, particularly the Biosignal Transformer (BIOT), for EEG seizure classification. Using a clinically relevant dataset and a controlled evaluation framework, both architectures are benchmarked across multiple temporal window configurations and training regimes. We evaluate four architectures (Linear Transformer baseline, Matrix Long Short-Term Memory (mLSTM), Scalar Long Short-Term Memory (sLSTM), and their combination) on the Temple University EEG Event Corpus (TUEV) dataset using multiple temporal window configurations. The experimental protocol maintains consistency with the original Biosignal Transformer (BIOT) evaluation framework, replacing only sequence processing components while preserving the tokenization and projection layers. Results demonstrate window-dependent performance patterns with only modest improvements. sLSTM performs best at longer windows, mLSTM at intermediate lengths, while the combined architecture offers stability without achieving peak performance. However, statistical testing reveals that most improvements do not reach statistical significance, with only isolated metrics reaching statistical thresholds. The architectural variants introduce substantial computational overhead without consistent gains across evaluation metrics. This work presents a systematic comparison of Extended Long Short-Term Memory (xLSTM) architectures for clinical Electroencephalogram (EEG) seizure classification. Our findings demonstrate that xLSTM architectures provide limited practical advantages over Linear Transformers in imbalanced biosignal tasks, where dataset characteristics impose fundamental constraints on achievable performance. The results emphasize the importance of rigorous statistical validation beyond numerical comparisons and highlight practical trade-offs between architectural complexity and clinical utility.