Abstract Introduction Arousal events are transient interruptions of sleep caused by abrupt changes in neural and physiological activity, and accurate detection is essential for sleep staging and evaluating sleep-disordered breathing. However, manual scoring can be inconsistent due to subjective differences between scorers. To address this, we developed a spectrogram-based deep learning model that detects arousal events as a temporal segmentation task. This study presents the model’s architecture, training strategy, and performance. Methods We developed an arousal detection model based on a two-dimensional U-Net architecture, using nine-channel spectrogram tensors of size (9, Frequency, Time) as input. Spectrograms were selected for their robustness to amplitude variability and their suitability for capturing the abrupt spectral changes characteristic of arousal events. The architecture comprises four key components: (1) Encoder–decoder, which reframes arousal detection as a segmentation task; (2) Bottleneck, which evaluates the relative importance of compressed feature channels; (3) Deep supervision, which stabilizes intermediate representations through auxiliary losses; and (4) Frequency convolution, which applies a final 1×1 convolution to infer the reliability of individual frequency components rather than using simple mean aggregation. The final Sigmoid layer outputs a 0–1 probability map indicating arousal likelihood. An optional freq = True mode aggregates the frequency axis using a learnable 1D Conv1d, but this approach yielded negligible performance gains and was therefore omitted from the final model. Without frequency convolution, the model produces an output tensor of shape (1, Freq, Time), which is averaged along the frequency axis during inference to yield a final (1, 1, Time) arousal prediction sequence. Results We conducted a clinical trial with 99 participants. Golden reference data were generated using a two-thirds voting approach, and agreement analysis showed that arousal detection achieved a PPA of 82.2% and an NPA of 95%, outperforming typical human-scorers agreement. Conclusion The spectrogram-based U-Net effectively detects arousal events as a temporal segmentation task. It faithfully models the physiological definition of arousal, operates robustly, and is currently deployed in the SOMNUM. Support (if any) This work was supported by the Technology development Program (RS-2023-00321754) funded by the Ministry of SMEs and Startups (MSS, Korea).
Choi et al. (Fri,) studied this question.