Fusion of End-to-End Deep Learning Models for Sequence-to-Sequence Sleep Staging

A Review of Recurrent Neural Network-Based Methods in Computational Physiology

Artificial intelligence and machine learning techniques have progressed dramatically and become powerful tools required to solve complicated tasks, such as computer vision, speech recognition, and natural language processing. Since these techniques have provided promising and evident results in these fields, they emerged as valuable methods for applications in human physiology and healthcare. General physiological recordings are time-related expressions of bodily processes associated with health or morbidity. Sequence classification, anomaly detection, decision making, and future status prediction drive the learning algorithms to focus on the temporal pattern and model the nonstationary dynamics of the human body. These practical requirements give birth to the use of recurrent neural networks (RNNs), which offer a tractable solution in dealing with physiological time series and provide a way to understand complex time variations and dependencies. The primary objective of this article is to provide an overview of current applications of RNNs in the area of human physiology for automated prediction and diagnosis within different fields. Finally, we highlight some pathways of future RNN developments for human physiology.

XSleepNet: Multi-View Sequential Model for Automatic Sleep Staging

Automating sleep staging is vital to scale up sleep assessment and diagnosis to serve millions experiencing sleep deprivation and disorders and enable longitudinal sleep monitoring in home environments. Learning from raw polysomnography signals and their derived time-frequency image representations has been prevalent. However, learning from multi-view inputs (e.g., both the raw signals and the time-frequency images) for sleep staging is difficult and not well understood. This work proposes a sequence-to-sequence sleep staging model, XSleepNet,¹ that is capable of learning a joint representation from both raw signals and time-frequency images. Since different views may generalize or overfit at different rates, the proposed network is trained such that the learning pace on each view is adapted based on their generalization/overfitting behavior. In simple terms, the learning on a particular view is speeded up when it is generalizing well and slowed down when it is overfitting. View-specific generalization/overfitting measures are computed on-the-fly during the training course and used to derive weights to blend the gradients from different views. As a result, the network is able to retain the representation power of different views in the joint features which represent the underlying distribution better than those learned by each individual view alone. Furthermore, the XSleepNet architecture is principally designed to gain robustness to the amount of training data and to increase the complementarity between the input views. Experimental results on five databases of different sizes show that XSleepNet consistently outperforms the single-view baselines and the multi-view baseline with a simple fusion strategy. Finally, XSleepNet also outperforms prior sleep staging methods and improves previous state-of-the-art results on the experimental databases.

Automatic sleep scoring with LSTM networks: impact of time granularity and input signals

Supervised automatic sleep scoring algorithms are usually trained using sleep stage labels manually annotated on 30 s epochs of PSG data. In this study, we investigate the impact of using shorter epochs with various PSG input signals for training and testing a Long Short Term Memory (LSTM) neural network. An LSTM model is evaluated on the provided 30 s epoch sleep stage labels from a publicly available dataset, as well as on 10 s subdivisions. Additionally, three independent scorers re-labeled a subset of the dataset on shorter time windows. The automatic sleep scoring experiments were repeated on the re-annotated subset.The highest performance is achieved on features extracted from 30 s epochs of a single channel frontal EEG. The resulting accuracy, precision and recall were of 92.22%, 67.58% and 66.00% respectively. When using a shorter epoch as input, the performance decreased by approximately 20%. Re-annotating a subset of the dataset on shorter time epochs did not improve the results and further altered the sleep stage detection performance. Our results show that our feature-based LSTM classification algorithm performs better on 30 s PSG epochs when compared to 10 s epochs used as input. Future work could be oriented to determining whether varying the epoch size improves classification outcomes for different types of classification algorithms.

End-to-End Sleep Staging Using Nocturnal Sounds from Microphone Chips for Mobile Devices

PURPOSE

Nocturnal sounds contain numerous information and are easily obtainable by a non-contact manner. Sleep staging using nocturnal sounds recorded from common mobile devices may allow daily at-home sleep tracking. The objective of this study is to introduce an end-to-end (sound-to-sleep stages) deep learning model for sound-based sleep staging designed to work with audio from microphone chips, which are essential in mobile devices such as modern smartphones.

PATIENTS AND METHODS

Two different audio datasets were used: audio data routinely recorded by a solitary microphone chip during polysomnography (PSG dataset, N=1154) and audio data recorded by a smartphone (smartphone dataset, N=327). The audio was converted into Mel spectrogram to detect latent temporal frequency patterns of breathing and body movement from ambient noise. The proposed neural network model learns to first extract features from each 30-second epoch and then analyze inter-epoch relationships of extracted features to finally classify the epochs into sleep stages.

RESULTS

Our model achieved 70% epoch-by-epoch agreement for 4-class (wake, light, deep, REM) sleep stage classification and robust performance across various signal-to-noise conditions. The model performance was not considerably affected by sleep apnea or periodic limb movement. External validation with smartphone dataset also showed 68% epoch-by-epoch agreement.

CONCLUSION

The proposed end-to-end deep learning model shows potential of low-quality sounds recorded from microphone chips to be utilized for sleep staging. Future study using nocturnal sounds recorded from mobile devices at home environment may further confirm the use of mobile device recording as an at-home sleep tracker.

Towards More Accurate Automatic Sleep Staging via Deep Transfer Learning

BACKGROUND

Despite recent significant progress in the development of automatic sleep staging methods, building a good model still remains a big challenge for sleep studies with a small cohort due to the data-variability and data-inefficiency issues. This work presents a deep transfer learning approach to overcome these issues and enable transferring knowledge from a large dataset to a small cohort for automatic sleep staging.

METHODS

We start from a generic end-to-end deep learning framework for sequence-to-sequence sleep staging and derive two networks as the means for transfer learning. The networks are first trained in the source domain (i.e. the large database). The pretrained networks are then finetuned in the target domain (i.e. the small cohort) to complete knowledge transfer. We employ the Montreal Archive of Sleep Studies (MASS) database consisting of 200 subjects as the source domain and study deep transfer learning on three different target domains: the Sleep Cassette subset and the Sleep Telemetry subset of the Sleep-EDF Expanded database, and the Surrey-cEEGrid database. The target domains are purposely adopted to cover different degrees of data mismatch to the source domains.

RESULTS

Our experimental results show significant performance improvement on automatic sleep staging on the target domains achieved with the proposed deep transfer learning approach.

CONCLUSIONS

These results suggest the efficacy of the proposed approach in addressing the above-mentioned data-variability and data-inefficiency issues.

SIGNIFICANCE

As a consequence, it would enable one to improve the quality of automatic sleep staging models when the amount of data is relatively small.¹¹The source code and the pretrained models are published at https://github.com/pquochuy/sleep_transfer_learning.

Confidence-Based Framework Using Deep Learning for Automated Sleep Stage Scoring

STUDY OBJECTIVES

Automated sleep stage scoring is not yet vigorously used in practice because of the black-box nature and the risk of wrong predictions. The objective of this study was to introduce a confidence-based framework to detect the possibly wrong predictions that would inform clinicians about which epochs would require a manual review and investigate the potential to improve accuracy for automated sleep stage scoring.

METHODS

We used 702 polysomnography studies from a local clinical dataset (SNUBH dataset) and 2804 from an open dataset (SHHS dataset) for experiments. We adapted the state-of-the-art TinySleepNet architecture to train the classifier and modified the ConfidNet architecture to train an auxiliary confidence model. For the confidence model, we developed a novel method, Dropout Correct Rate (DCR), and the performance of it was compared with other existing methods.

RESULTS

Confidence estimates (0.754) reflected accuracy (0.758) well in general. The best performance for differentiating correct and wrong predictions was shown when using the DCR method (AUROC: 0.812) compared to the existing approaches which largely failed to detect wrong predictions. By reviewing only 20% of epochs that received the lowest confidence values, the overall accuracy of sleep stage scoring was improved from 76% to 87%. For patients with reduced accuracy (ie, individuals with obesity or severe sleep apnea), the possible improvement range after applying confidence estimation was even greater.

CONCLUSION

To the best of our knowledge, this is the first study applying confidence estimation on automated sleep stage scoring. Reliable confidence estimates by the DCR method help screen out most of the wrong predictions, which would increase the reliability and interpretability of automated sleep stage scoring.