OCRNN: An orthogonal constrained recurrent neural network for sleep analysis based on EEG data

Abstract

This paper introduced an end-to-end mixed deep learning model for automatic sleep analysis based on the EEG signal. Unlike some existing machine learning models for EEG analysis, we did not rely on any hand-crafted feature engineering and elaborate pipeline design. Furthermore, apart from some existing deep learning frameworks based on some off-the-self existing modules, we introduced the orthogonal constrained recurrent neural network (OCRNN) as the downstream module after the spatial-temporal expansion and representation provided by the one dimensional convolutional neural networks. We evaluated our model using the EEG-based sleep datasets for sleep stage scoring. We compared the performances of four types of RNN frameworks, where three of them are OCRNNs. The results show that OCRNN can achieve competitive better F1 score, accuracy and AUC score compared to the previous baseline results. Moreover, our model (orRNN and pdRNN) can achieve the above results with less number of parameters and less number of training epochs, which demonstrate its potential usage to launch approximate real-time medical diagnosis.

Introduction

Electroencephalograph (EEG) is a complex signal that contains information from different frequency bands and different channels (multiple probes for different brain functional areas). This multi-modality signal attracts researchers to find approaches to disclose the underlying information so that we can leverage the information to solve problems such as epilepsy, sleep, brain computer interfacing, and cognitive monitoring [1]. The challenging task of making classification based on EEG signals lies on the fact that EEG is a non-stationary and low signal-to-noise ratio (SNR) signal, so that some single hand-crafted signal processing approach meets the bottleneck [2]. EEG has the high temporal resolution due to the high-speed propagation of the electric field but EEG has a low spatial resolution so that the signals are highly correlated. This increases the difficulty to decompose and expand the details of the signal.

Thanks to the development of the modern computing power to trigger the potential of data science and artificial intelligence, researchers seek to handle these problems with more powerful and automatic toolboxes to understand the EEG signal. One of the most adopted off-the-self strategies is as follows. The Component analysis such as PCA, ICA and neighborhood component analysis [3], [4], [5] serve as the pre-processing steps for the downstream steps like SVM, k-means [6], [7], [8]. However, this kind of framework belong to the traditional hand-crafted machine learning algorithms which highly rely on the domain knowledge and they may be subject to the limitation of the applicability and flexibility. Deep learning, as one categories of the machine learning families, provide the hierarchical representations of input data complicate non-linear activation and flexible connections. Since the proven performance of the deep learning in areas like images, videos, acoustic signal, natural language processing, we witness the rise of the interest for deployment of the deep learning model in EEG understanding as well.

Among the deep learning models, recurrent neural network, as a naturally time-series based deep learning model, mostly will serve as the first-try model for EEG related tasks. For instance, a special kind of RNN named echo state network was utilized to classify the REM Behavior Disorder (RBD) from 118 subjects (with healthy control) based on the collected EEG signals [9]. A more complex multi-task RNN framework was proposed for motion intention recognition based on features extracted from segregated EEG signals [10]. An end-to-end RNN learning framework was designed to learn the EEG signal after removal of the artifact for the movement related cortical potentials (MRCP) in order to recognize lane change decision making [11]. More RNN based on system-level application demonstrate the suitability of RNN in broader applications such as autonomous whole-arm exoskeleton control [12] and decoding of motor imagery movements [13].

However, according to the statistics by Roy et al. [14], around 40% of the studies used convolutional neural networks (CNNs), while 13% used recurrent neural networks (RNNs). The previous studies trained their neural networks more on preprocessed EEG signals. The benefit of the mixed model is that it can leverage the advantages of each submodule of the neural network while also minimize the incompatibility. Some previous studies have demonstrated the mixed model for raw time-series data in other applications such as soil moisture retrieval [15], [16] and signal detection [17], [18], especially the mixed model or different neural network, such as CNN-RNN mixed framework [19]. In terms of the EEG-related mixed model, most of the models focus on the end-to-end downstream pipeline framework with hand-crafted pre-processing techniques like short-time Fourier transform (STFT), mel-frequency cepstral coefficient (MFCC) and low-pass filter (LPF). These models still will be constrained by physical limitations due to the preprocessing steps so that we seek to more automatic and flexible framework. A one dimensional CNN plus bidirectional long short term memory (LSTM) model is proposed as the fully mixed deep learning framework [20], but we do not find too much related examples like this.

In this paper, we propose 1dCNN-OCRNN model, an one dimensional CNN followed by the orthogonal constrained RNN for the EEG signal based sleep stage scoring task. The highlights of our work are:

• We use the 1dCNN to generate the multi channel feature maps to achieve the similar function as the time-frequency function, but we reduced the complexity of model by adjusting the output and eliminate repeated convolution layer with fewer channels.

• We design our own orthogonal constraints based RNNs with its related coordinate-descent like iteration algorithm for training. These type of RNNs can mitigate the long-existing vanishing and exploding problems by leveraging the self-adjusted weights. Therefore, important information from the EEG multi-channel feature maps are stored in the ”memory” of the RNN.

• We demonstrate the results use the real-world sleep staging classification dataset based on EEG signals. We provide solid comparison of the performance of the three OCRNNs and the standard LSTM. Compared to the early results provided in [20], we achieve competitive slightly better results using approximate 1/6 total number of parameters and 1/3 of training epochs. Therefore, by saving the computation cost in both spatial and temporal, our model has the potential to be broadly deployed in real scenario.

The rest of the paper is organized as follows. In Section 2, the preliminaries of the OCRNNs and its iterative training algorithms are introduced. In Section 3, the 1dCNN-OCRNN model is introduced and we compared different combinations of the mixture model. In Section 4, the experiment of the EEG sleep stage analysis is introduced and the solid performance comparison is analyzed and discussed. Finally, we conclude in Section 5.