Towards More Accurate Automatic Sleep Staging via Deep Transfer Learning

DeepSleepNet-Lite: A Simplified Automatic Sleep Stage Scoring Model With Uncertainty Estimates

Deep learning is widely used in the most recent automatic sleep scoring algorithms. Its popularity stems from its excellent performance and from its ability to process raw signals and to learn feature directly from the data. Most of the existing scoring algorithms exploit very computationally demanding architectures, due to their high number of training parameters, and process lengthy time sequences in input (up to 12 minutes). Only few of these architectures provide an estimate of the model uncertainty. In this study we propose DeepSleepNet-Lite, a simplified and lightweight scoring architecture, processing only 90-seconds EEG input sequences. We exploit, for the first time in sleep scoring, the Monte Carlo dropout technique to enhance the performance of the architecture and to also detect the uncertain instances. The evaluation is performed on a single-channel EEG Fpz-Cz from the open source Sleep-EDF expanded database. DeepSleepNet-Lite achieves slightly lower performance, if not on par, compared to the existing state-of-the-art architectures, in overall accuracy, macro F1-score and Cohen's kappa (on Sleep-EDF v1-2013 ±30mins: 84.0%, 78.0%, 0.78; on Sleep-EDF v2-2018 ±30mins: 80.3%, 75.2%, 0.73). Monte Carlo dropout enables the estimate of the uncertain predictions. By rejecting the uncertain instances, the model achieves higher performance on both versions of the database (on Sleep-EDF v1-2013 ±30mins: 86.1.0%, 79.6%, 0.81; on Sleep-EDF v2-2018 ±30mins: 82.3%, 76.7%, 0.76). Our lighter sleep scoring approach paves the way to the application of scoring algorithms for sleep analysis in real-time.

Sleep monitoring using ear-centered setups: Investigating the influence from electrode configurations

Modern sleep monitoring development is shifting towards the use of unobtrusive sensors combined with algorithms for automatic sleep scoring. Many different combinations of wet and dry electrodes, ear-centered, forehead-mounted or headband-inspired designs have been proposed, alongside an ever growing variety of machine learning algorithms for automatic sleep scoring.

OBJECTIVE

Among candidate positions, those in the facial area and around the ears have the benefit of being relatively hairless, and in our view deserve extra attention. In this paper, we seek to determine the limits to sleep monitoring quality within this spatial constraint.

METHODS

We compare 13 different, realistic sensor setups derived from the same data set and analysed with the same pipeline.

RESULTS

All setups which include both a lateral and an EOG derivation show similar, state-of-the-art performance, with average Cohen's kappa values of at least 0.80.

CONCLUSION

If large electrode distances are used, positioning is not critical for achieving large sleep-related signal-to-noise-ratio, and hence accurate sleep scoring.

SIGNIFICANCE

We argue that with the current competitive performance of automated staging approaches, there is a need for establishing an improved benchmark beyond current single human rater scoring.

Transferring Structured Knowledge in Unsupervised Domain Adaptation of a Sleep Staging Network

Automatic sleep staging based on deep learning (DL) has been attracting attention for analyzing sleep quality and determining treatment effects. It is challenging to acquire long-term sleep data from numerous subjects and manually labeling them even though most DL-based models are trained using large-scale sleep data to provide state-of-the-art performance. One way to overcome this data shortage is to create a pre-trained network with an existing large-scale dataset (source domain) that is applicable to small cohorts of datasets (target domain); however, discrepancies in data distribution between the domains prevent successful refinement of this approach. In this paper, we propose an unsupervised domain adaptation method for sleep staging networks to reduce discrepancies by realigning the domains in the same space and producing domain-invariant features. Specifically, in addition to a classical domain discriminator, we introduce local dis-criminators-subject and stage-to maintain the intrinsic structure of sleep data to decrease local misalignments while using adversarial learning to play a minimax game between the feature extractor and discriminators. Moreover, we present several optimization schemes during training because the conventional adversarial learning is not effective to our training scheme. We evaluate the performance of the proposed method by examining the staging performances of a baseline network compared with direct transfer (DT) learning in various conditions. The experimental results demonstrate that the proposed domain adaptation significantly improves the performance though it needs no labeled sleep data in target domain.

Interpretable deep learning for the remote characterisation of ambulation in multiple sclerosis using smartphones

The emergence of digital technologies such as smartphones in healthcare applications have demonstrated the possibility of developing rich, continuous, and objective measures of multiple sclerosis (MS) disability that can be administered remotely and out-of-clinic. Deep Convolutional Neural Networks (DCNN) may capture a richer representation of healthy and MS-related ambulatory characteristics from the raw smartphone-based inertial sensor data than standard feature-based methodologies. To overcome the typical limitations associated with remotely generated health data, such as low subject numbers, sparsity, and heterogeneous data, a transfer learning (TL) model from similar large open-source datasets was proposed. Our TL framework leveraged the ambulatory information learned on human activity recognition (HAR) tasks collected from wearable smartphone sensor data. It was demonstrated that fine-tuning TL DCNN HAR models towards MS disease recognition tasks outperformed previous Support Vector Machine (SVM) feature-based methods, as well as DCNN models trained end-to-end, by upwards of 8-15%. A lack of transparency of "black-box" deep networks remains one of the largest stumbling blocks to the wider acceptance of deep learning for clinical applications. Ensuing work therefore aimed to visualise DCNN decisions attributed by relevance heatmaps using Layer-Wise Relevance Propagation (LRP). Through the LRP framework, the patterns captured from smartphone-based inertial sensor data that were reflective of those who are healthy versus people with MS (PwMS) could begin to be established and understood. Interpretations suggested that cadence-based measures, gait speed, and ambulation-related signal perturbations were distinct characteristics that distinguished MS disability from healthy participants. Robust and interpretable outcomes, generated from high-frequency out-of-clinic assessments, could greatly augment the current in-clinic assessment picture for PwMS, to inform better disease management techniques, and enable the development of better therapeutic interventions.

The power of ECG in multimodal patient-specific seizure monitoring: Added value to an EEG-based detector using limited channels

Objective

Wearable seizure detection devices could provide more reliable seizure documentation outside the hospital compared to seizure self‐reporting by patients, which is the current standard. Previously, during the SeizeIT1 project, we studied seizure detection based on behind‐the‐ear electroencephalography (EEG). However, the obtained sensitivities were too low for practical use, because not all seizures are associated with typical ictal EEG patterns. Therefore, in this paper, we aim to develop a multimodal automated seizure detection algorithm integrating behind‐the‐ear EEG and electrocardiography (ECG) for detecting focal seizures. In this framework, we quantified the added value of ECG to behind‐the‐ear EEG.

Methods

This study analyzed three multicenter databases consisting of 135 patients having focal epilepsy and a total of 896 seizures. A patient‐specific multimodal automated seizure detection algorithm was developed using behind‐the‐ear/temporal EEG and single‐lead ECG. The EEG and ECG data were processed separately using machine learning methods. A late integration approach was applied for fusing those predictions.

Results

The multimodal algorithm outperformed the EEG‐based algorithm in two of three databases, with an increase of 11% and 8% in sensitivity for the same false alarm rate.

Significance

ECG can be of added value to an EEG‐based seizure detection algorithm using only behind‐the‐ear/temporal lobe electrodes for patients with focal epilepsy.

Interrater sleep stage scoring reliability between manual scoring from two European sleep centers and automatic scoring performed by the artificial intelligence-based Stanford-STAGES algorithm

STUDY OBJECTIVES

The objective of this study was to evaluate interrater reliability between manual sleep stage scoring performed in 2 European sleep centers and automatic sleep stage scoring performed by the previously validated artificial intelligence-based Stanford-STAGES algorithm.

METHODS

Full night polysomnographies of 1,066 participants were included. Sleep stages were manually scored in Berlin and Innsbruck sleep centers and automatically scored with the Stanford-STAGES algorithm. For each participant, we compared (1) Innsbruck to Berlin scorings (INN vs BER); (2) Innsbruck to automatic scorings (INN vs AUTO); (3) Berlin to automatic scorings (BER vs AUTO); (4) epochs where scorers from Innsbruck and Berlin had consensus to automatic scoring (CONS vs AUTO); and (5) both Innsbruck and Berlin manual scorings (MAN) to the automatic ones (MAN vs AUTO). Interrater reliability was evaluated with several measures, including overall and sleep stage-specific Cohen's κ.

RESULTS

Overall agreement across participants was substantial for INN vs BER (κ = 0.66 ± 0.13), INN vs AUTO (κ = 0.68 ± 0.14), CONS vs AUTO (κ = 0.73 ± 0.14), and MAN vs AUTO (κ = 0.61 ± 0.14), and moderate for BER vs AUTO (κ = 0.55 ± 0.15). Human scorers had the highest disagreement for N1 sleep (κ_N1 = 0.40 ± 0.16 for INN vs BER). Automatic scoring had lowest agreement with manual scorings for N1 and N3 sleep (κ_N1 = 0.25 ± 0.14 and κ_N3 = 0.42 ± 0.32 for MAN vs AUTO).

CONCLUSIONS

Interrater reliability for sleep stage scoring between human scorers was in line with previous findings, and the algorithm achieved an overall substantial agreement with manual scoring. In this cohort, the Stanford-STAGES algorithm showed similar performances to the ones achieved in the original study, suggesting that it is generalizable to new cohorts. Before its integration in clinical practice, future independent studies should further evaluate it in other cohorts.

MetaSleepLearner: A Pilot Study on Fast Adaptation of Bio-Signals-Based Sleep Stage Classifier to New Individual Subject Using Meta-Learning

Identifying bio-signals based-sleep stages requires time-consuming and tedious labor of skilled clinicians. Deep learning approaches have been introduced in order to challenge the automatic sleep stage classification conundrum. However, the difficulties can be posed in replacing the clinicians with the automatic system due to the differences in many aspects found in individual bio-signals, causing the inconsistency in the performance of the model on every incoming individual. Thus, we aim to explore the feasibility of using a novel approach, capable of assisting the clinicians and lessening the workload. We propose the transfer learning framework, entitled MetaSleepLearner, based on Model Agnostic Meta-Learning (MAML), in order to transfer the acquired sleep staging knowledge from a large dataset to new individual subjects (source code is available at https://github.com/IoBT-VISTEC/MetaSleepLearner). The framework was demonstrated to require the labelling of only a few sleep epochs by the clinicians and allow the remainder to be handled by the system. Layer-wise Relevance Propagation (LRP) was also applied to understand the learning course of our approach. In all acquired datasets, in comparison to the conventional approach, MetaSleepLearner achieved a range of 5.4% to 17.7% improvement with statistical difference in the mean of both approaches. The illustration of the model interpretation after the adaptation to each subject also confirmed that the performance was directed towards reasonable learning. MetaSleepLearner outperformed the conventional approaches as a result from the fine-tuning using the recordings of both healthy subjects and patients. This is the first work that investigated a non-conventional pre-training method, MAML, resulting in a possibility for human-machine collaboration in sleep stage classification and easing the burden of the clinicians in labelling the sleep stages through only several epochs rather than an entire recording.

A Survey on Deep Learning-Based Short/Zero-Calibration Approaches for EEG-Based Brain-Computer Interfaces

Brain-computer interfaces (BCIs) utilizing machine learning techniques are an emerging technology that enables a communication pathway between a user and an external system, such as a computer. Owing to its practicality, electroencephalography (EEG) is one of the most widely used measurements for BCI. However, EEG has complex patterns and EEG-based BCIs mostly involve a cost/time-consuming calibration phase; thus, acquiring sufficient EEG data is rarely possible. Recently, deep learning (DL) has had a theoretical/practical impact on BCI research because of its use in learning representations of complex patterns inherent in EEG. Moreover, algorithmic advances in DL facilitate short/zero-calibration in BCI, thereby suppressing the data acquisition phase. Those advancements include data augmentation (DA), increasing the number of training samples without acquiring additional data, and transfer learning (TL), taking advantage of representative knowledge obtained from one dataset to address the so-called data insufficiency problem in other datasets. In this study, we review DL-based short/zero-calibration methods for BCI. Further, we elaborate methodological/algorithmic trends, highlight intriguing approaches in the literature, and discuss directions for further research. In particular, we search for generative model-based and geometric manipulation-based DA methods. Additionally, we categorize TL techniques in DL-based BCIs into explicit and implicit methods. Our systematization reveals advances in the DA and TL methods. Among the studies reviewed herein, ~45% of DA studies used generative model-based techniques, whereas ~45% of TL studies used explicit knowledge transferring strategy. Moreover, based on our literature review, we recommend an appropriate DA strategy for DL-based BCIs and discuss trends of TLs used in DL-based BCIs.

U-Sleep: resilient high-frequency sleep staging

Sleep disorders affect a large portion of the global population and are strong predictors of morbidity and all-cause mortality. Sleep staging segments a period of sleep into a sequence of phases providing the basis for most clinical decisions in sleep medicine. Manual sleep staging is difficult and time-consuming as experts must evaluate hours of polysomnography (PSG) recordings with electroencephalography (EEG) and electrooculography (EOG) data for each patient. Here, we present U-Sleep, a publicly available, ready-to-use deep-learning-based system for automated sleep staging ( sleep.ai.ku.dk ). U-Sleep is a fully convolutional neural network, which was trained and evaluated on PSG recordings from 15,660 participants of 16 clinical studies. It provides accurate segmentations across a wide range of patient cohorts and PSG protocols not considered when building the system. U-Sleep works for arbitrary combinations of typical EEG and EOG channels, and its special deep learning architecture can label sleep stages at shorter intervals than the typical 30 s periods used during training. We show that these labels can provide additional diagnostic information and lead to new ways of analyzing sleep. U-Sleep performs on par with state-of-the-art automatic sleep staging systems on multiple clinical datasets, even if the other systems were built specifically for the particular data. A comparison with consensus-scores from a previously unseen clinic shows that U-Sleep performs as accurately as the best of the human experts. U-Sleep can support the sleep staging workflow of medical experts, which decreases healthcare costs, and can provide highly accurate segmentations when human expertize is lacking.

Grading hypoxic-ischemic encephalopathy in neonatal EEG with convolutional neural networks and quadratic time-frequency distributions

OBJECTIVE

To develop an automated system to classify the severity of hypoxic-ischaemic encephalopathy injury (HIE) in neonates from the background electroencephalogram (EEG).

METHOD

By combining a quadratic time{frequency distribution (TFD) with a convolutional neural network, we develop a system that classifies 4 EEG grades of HIE. The network learns directly from the two- dimensional TFD through 3 independent layers with convolution in the time, frequency, and time{frequency directions. Computationally efficient algorithms make it feasible to transform each 5 minute epoch to the time-frequency domain by controlling for oversampling to reduce both computation and computer memory. The system is developed on EEG recordings from 54 neonates. Then the system is validated on a large unseen dataset of 338 hours of EEG recordings from 91 neonates obtained across multiple international centres.

RESULTS

The proposed EEG HIE-grading system achieves a leave-one-subject-out testing accuracy of 88.9% and kappa of 0.84 on the development dataset. Accuracy for the large unseen test dataset is 69.5% (95% confidence interval, CI: 65.3 to 73.6%) and kappa of 0.54, which is a significant (P < 0.001) improvement over a state-of-the-art feature-based method with an accuracy of 56.8% (95% CI: 51.4 to 61.7%) and kappa of 0.39. Performance of the proposed system was unaffected when the number of channels in testing was reduced from 8 to 2|accuracy for large validation dataset remained at 69.5% (95% CI: 65.5 to 74.0%).

SIGNIFICANCE

The proposed system outperforms the state-of-the-art machine learning algorithms for EEG grade classification on a large multi-centre unseen dataset, indicating the potential to assist clinical decision making for neonates with HIE.