Leednet: a lightweight network for event detection in EEG signals

Detection of diagnostically relevant events within an EEG recording requires making classification decisions about small event windows of a few seconds duration each. For long-duration EEG recordings, this can translate into thousands of decisions leading to a high computational load, which must be reduced for field deployments of EEG triage and decision support systems, especially in resource-constrained healthcare systems. In this work, we explore three avenues for enhancing computational efficiency: (1) purpose-built architectures; (2) length of the decision window/segment; and (3) lightweight preprocessing and feature extraction. For the architecture, we present LEEDNet (Lightweight EEG Event Detection Network), a lightweight convolutional architecture that is purpose-built for prompt classification of short EEG events into three categories (Normal, Slow Waves, or Spike and Sharp Waves). LEEDNet is evaluated on the expert-annotated NMT-Events dataset with subject-wise splits. Compared to four state-of-the-art baselines, LEEDNet delivers the best performance, on three out of four, performance evaluation metrics while being much more deployment friendly (requiring only 0.25 M parameters and 37 MFLOPs per segment, enabling inference in < 5 ms on a standard CPU). For the segment length, ablation results show a length of 2 s to be the best performing, with longer segments degrading due to label dilution and background dominance. We also examine the impact of different feature engineering configurations and whether the increased computational overhead associated with extraction of complex features leads to substantial performance gains. In particular, we examine three input modalities—raw signals, FFT-based spectra, and wavelets—both independently (single-branch) and in a modular, multi-branch fusion model. Our results indicate that the single-branch LEEDNet, based on Raw EEG data, achieves the best overall performance, with 81.2% accuracy and 77.5% macro-F1. No consistent gains were observed after the introduction of modalities based on Fourier and Wavelet features indicating that, for our architecture, a data-driven approach (based on raw EEG waveforms) is better suited compared to a handcrafted feature engineering approach. This is welcomed from a computation overhead perspective since complex feature engineering can be eliminated in favor of raw samples as input features without sacrificing performance.