1. Introduction

Recent studies have shown that sleep disorder patients continue to increase ^[1], and people with sleep disorders are more vulnerable to various motor diseases ^[1] or disorders such as Alzheimer's disease ^[3,4]. The most important part for the treatment of sleep disorders is estimation of the sleep stages. Based on the sleep stages, a doctor can determine the overall sleep quality and amount and then conduct a sleep diagnosis based on the analysis. However, the estimation of sleep stage can only proceed with the approach of manual sleep estimation by sleep specialists at a medical site ^[5], which results in high costs of economic resources and time. In addition, recent research reports that sleep experts’ sleep stage estimation results vary wildly from one to another ^[6] since the sleep stage is determined by their subjective judgment.

To solve these problems, various methods of estimating the sleep stages using neural networks have recently been proposed. Supratak et al. ^[7] proposed a sleep stage automation architecture using convolution network layers ^[8] and long short-term memory (LSTM) cells ^[8]. However, there is a black-box problem of not being able to explain the resulting process due to the use of a deep neural network ^[9]. Results reported for 21 limited subjects have led to insufficient generalization performance and validity of the model.

In 2019, Mousavi et al. ^[10] proposed a sleep stage estimation model using a convolutional neural network (CNN) and bidirectional recurrent neural network (Bi-RNN) ^[8] for seq2seq learning ^[11,16]. The results showed a slight improvement in accuracy compared to a previously proposed method, but the problem of using only a small number of subjects was not solved. Also, the black-box problem was further complicated because they used multiple Bi-RNNs.

The medical community is increasingly demanding explainable artificial intelligence (XAI) ^[12,13]. Thus, an explainable artificial neural network model should be proposed for medical decision support systems. However, no previously proposed neural networks ^[7,10] at an explainable level have been proposed for sleep automation models through artificial neural network algorithms.

In this study, we propose a deep neural network-based explainable model that estimates five sleep stages based on an EEG band component. The attention layer in the model updates its attention weights to learn which EEG band components are useful for an estimate of a specific sleep stage. Through these methods, the model learns from multiple -decomposed EEG band components, which sleep specialists suggest may help to estimate more accurate sleep stages ^[27,30]. In addition to these proposals, we demonstrate the generalization performance and increase the robustness of the model by using at least five times more subject data (80 subjects) than previous studies ^[7,10].

2. Methodology

2.2 Model Specifications

The attention layer ^[16] is located before the overall input of the model architecture, as shown in Fig. 1. How much each signal component is learned is reflected in the final label, which ranges from -1.0 to +1.0. The model can analyze the attention weight for each post-learning estimation stage by the attention layer structure. After learning the input data, analysis of the channel weights of the relevant attention layer allows us to analyze which EEG frequency component the model relies on more for estimating specific sleep stages.

To predict the sleep stage through the EEG data, the model performs representation learning and temporal learning. In the representation learning parts, 30 seconds of decomposed EEG signals were given as input to the 1D convolutional neural network (CNN). Small neural networks for each EEG frequency component were designed and constructed by a 1D CNN ^[8] while pooling layers eight times. Based on this structure, features for each EEG component can be learned appropriately for the input data for each CNN structure.

To learn about the time series elements of EEG data, temporal learning is proposed through Bi-GRU. Bi-GRU enables the learning of an epoch time unit with every 10 time steps for single-channel EEG information for 30 seconds of a sleep EEG signal. A model with this structure can learn sequential information in the time domain and extract features from the input EEG frequency components. The model hyper parameter is optimized based on previous sleep stage estimation algorithms ^[7,10], and parameters are updated by an adaptive moment estimation (ADAM) optimizer ^[8].

Fig. 1. Illustration of the proposed model. Model has three different parts: attention learning with attention layers, representation learning with convolution layers, and temporal learning parts with bidirectional-gated recurrent unit (Bi-GRU) cells.

Table 1. Representation learning architecture with convolution layers.

3. Results

To evaluate the proposed model, performance measurement of the model was performed using the F1-score ^[26], the most used indicator in sleep stage analysis ^[7,10]. The results of 5-fold cross-validation ^[25] of the sleep stage estimation model for 80 subjects’ single-channel EEG sleep data with the proposed method are shown in Table 2. Cross-validation of the model showed that the average F1-score was 70.61% (${\pm}$26.34).

To analyze the major EEG components from the EEG decomposition using the model in estimating sleep stages, attention weights were used to calculate which EEG frequency components have more weights to track their update rate for each sleep stage. The attention-weight analysis demonstrates the channel-specific attention weights used to estimate each sleep stage among the five sleep-stage estimation labels, as shown in Fig. 2. The results in Fig. 2 confirm the previous findings that different valid EEG signal components are used to estimate the sleep stages ^[27-30]. In light sleep, such as N1 and N2, and REM, high-frequency components have more attention rates, as shown in Fig. 2. This shows that the model estimates the sleep stage mainly based on the high-frequency components, whereas deep sleep (N3) shows more attention weight in the low-frequency component, showing that deep sleep is estimated mainly based on the low-frequency component.

Table 2. Average model evaluation results through 5-fold cross validation.

4. Discussion

Analysis of the attention weights of the proposed model confirms that the model learns through different channel weights corresponding to different EEG components. This aligns with previous knowledge for sleep stage estimation that different brainwave components are used for estimating different sleep stages ^[28-30]. The proposed model has more clearly described which component of the EEG is more useful to detect specific sleep stages based on the selective focus learned from different frequency components of the input data to estimate a particular sleep stage. In addition, the results for the attention weights of each EEG component show similar results to those of useful EEG components reported in previous papers analyzing sleep stage through EEG channels. This shows the model might learn how to evaluate the sleep stage around the low-frequency component for deep sleep and the high-frequency component for light sleep, similar to what sleep specialists do for their estimation of the sleep stages.

Furthermore, considering the characteristics of sleep-stage estimation models, which show worse performance as the number of subjects increases, the model still shows similar F1-scores, even when subject numbers increase by more than 5 times compared to previously reported sleep-stage estimation algorithms. This implies that attention mechanisms in sleep stage analysis could be effective in estimating the sleep stages and designing an explainable model.

Fig. 2. Attention weight for each sleep stage. (a)-(d) shows high attention weight for high-frequency band, (e) shows high attention weight in low-frequency band.

5. Conclusion

In this paper, we proposed an explainable artificial intelligence model centered on brainwave components through attention algorithms. Based on that model, we were able to show which brainwave components were useful in predicting specific sleep stages, and the accuracy was also similar to that of previously proposed sleep stage estimation models. However, the problem of chronic sleep stage estimation bias caused by an imbalance of the sleep stage can be still observed. This is particularly shown in the N1 sleep stage.

To solve this problem, we propose increasing the learning volume through the synthetic minority oversampling technique (SMOTE) ^[31] and data augmentation ^[32] to improve the performance of estimating the N1 sleep stage. In addition, there are still some black-box problems because the level of the explainable model cannot show the learning progress by time frame. To solve this problem, we propose a study of a deeper-level explainable model through time-unit attention ^[33].