Presentation: 2025 ND EPSCoR Annual conference 

October 21, 2025, NDSU Memorial Union, Fargo, North Dakota

Sleep Stage Classification in Real-Time Using Single-Channel Frontal EEG System with a Memory-Augmented Neural Network

We propose a lightweight neural network for real-time sleep stage classification using a single frontal EEG channel (Fp1–Fp2), optimized for deployment on portable, resource-constrained devices. Unlike traditional deep learning models that rely heavily on recurrent layers and large memory footprints, our approach uses a hybrid architecture combining feedforward neural networks and temporal context through a memory-augmented transition vector. This vector is derived by multiplying the previous epoch’s softmax prediction with a learned transition probability matrix. To capture temporal dynamics in sleep architecture, we analyze autocovariance functions across stages using the MASS SS5 dataset. Our results reveal prolonged dependencies in N3 and REM stages, validating the inclusion of memory features. We train two models: a no–memory feedforward model and a memory-augmented variant. During inference, a confidence-based selector chooses the more certain prediction if the model confidence exceeds 0.7–a threshold systematically tuned to improve classification performance while minimizing the number of rejected epochs. This combined approach achieves 85.4% accuracy and a Cohen’s kappa of 0.79 while rejecting just 9.4% of uncertain epochs, offering a compelling solution for low-power, wearable EEG systems.