LSTM-based approach for A-phase detection from single-channel EEG in whole sleep recordings

Identifying short and recurrent changes in brain activity during sleep, known as A-phases, is critical to sleep diagnosis because they correlate with sleep quality and sleep disorders. A-phases are the basic components of the well-established sleep pattern known as Cyclic Alternating Pattern (CAP) and are used as clinical indicators of sleep instability. An automated method for identifying A-phases in overnight electroencephalography (EEG) recordings is presented, and the CAP rate between expert and model annotations is evaluated. The classification model was constructed with fully connected layers of Long Short-Time Memory (LSTM) cells and dense layers to identify the time associated with the occurrence of A-phases on a second-by-second basis. The full dataset available, 108 recordings with 50 554 A-phases, from the publicly accessible CAP Sleep Database on Physionet was used to train and test the model under the Leave-One-Out scheme. The model was trained and evaluated to detect A-phases in the entire recording, including Wake, REM, and NREM sleep. The performance of the test reaches an F-score of 59.40 ± 8.10 and 64.02 ± 7.56 for the whole sleep and NREM sleep respectively. The CAP rate results showed that the model annotations are in agreement with the expert annotations (correlation, ρ=0.7). The A-phase identification method showed performance comparable to state-of-the-art. Notably, unlike previous studies, the method does not require sleep stage selection. This paves the way for its use in both clinical and home settings, minimizing the need for human intervention and additional algorithms for noise reduction or segment selection.