# A sleep staging model based on adversarial domain generalized residual attention network

## TL;DR

This paper introduces a new sleep staging model that improves generalization across different subjects using an adversarial domain generalized residual attention network.

## Contribution

The novel ADG-RANet model combines adversarial training with residual attention to enhance generalization in sleep staging.

## Key findings

- The model achieved 82.51% five-classification accuracy on the ISRUC-S3 dataset.
- It outperformed benchmark models in generalization with an m-F1 score of 0.8100 and Kappa coefficient of 0.7748.

## Abstract

To solve the problem of poor generalization ability of the model on unknown data and the difference of physiological signals between different subjects. A sleep staging model based on Adversarial Domain Generalized Residual Attention Network (ADG-RANet) is designed. The model is divided into three parts: feature extractor, domain discriminator and label classifier. In the feature extractor part, the channel attention network is combined with the residual block to selectively enhance the important features and the correlation between multi-channel physiological signals. Inspired by the idea of U-shaped network, the details and context information in the input data are effectively captured through up-sampling and skip connection operations. The Bi-GRU network is used to further extract the deep temporal features. A Gradient Reversal Layer (GRL) is introduced between the domain discriminator and the feature extractor to promote the feature extractor to obtain the invariant features between different subjects through the adversarial training process. The label classifier uses the deep features learned by the feature extractor to perform sleep staging. According to the AASM sleep staging criterion, the five-classification accuracy of the model on the ISRUC-S3 dataset was 82.51%, the m-F1 score was 0.8100, and the Kappa coefficient was 0.7748. By observing the test results of each fold and comparing with the benchmark model, it is verified that the proposed model has better generalization on unknown data.