Internet of Things (IoT) for healthy sleep
< Project overview >
Sleep apnoea-hypopnea syndrome is a sleep disorder which is common in many children and adults. It is characterised by abnormal breath pauses or shallow breathing during sleep. It is a common cause of daytime sleepiness and neurocognitive impairment and is related to cardiovascular disease as well. Therefore, it is important to diagnose the syndrome as early as possible.
Overnight polysomnography (PSG) has been recognised as the gold standard method for a definitive diagnosis of sleep apnoea. However, it requires the sleeping process of patients to be monitored in laboratories with much professional equipment, which increases the cost and limits its application.
Additionally, the obtained overnight data need to be analysed by professional doctors or clinical technicians, which is usually boring and time-consuming.
The study of automatic sleep apnoea detection has attracted significant attention. In the IoT4 Healthy Sleep project we focused on automatic detection of apnoea with PSG, video and other data from portable devices.
Studies of numerous sleep disorders suggest that they appear to pre-empt the onset of numerous neurological disorders. These include rapid-eye movement (REM) sleep behaviour disorder (RBD), where mounting evidence suggests that this disorder predicts Parkinson’s disease (PD) by years, potentially decades. This predictive ability provides an opportunity to explore preventative medicine and better understand how neurodegenerative disorders develop over time.
Consequently there is a growing demand to identify and study individuals with RBD, but this process is hampered by manually-laborious diagnostic procedures and overburdened sleep clinics.
The International Classification of Sleep Disorders (ICSD3) requires polysomnography (PSG) evidence of REM sleep without atonia (RSWA) to make a diagnosis. The absence of atonia (the temporary paralysis of muscles) during REM is evaluated by clinicians that observe electromyogram (EMG) sensors, where muscle movement during REM sleep produces signal activity. Clinicians are taught to visually identify EMG activity without a clear and precise definition.
As a result our research has focused on developing automated diagnostic tools that include automated sleep staging in wearable recordings and the use of EMG sensors on limbs to automate limb movement detection and identify individuals with RBD.
What was done?
The project undertook a number of activities including gaining ethical approval, data collection, development of machine learning and IoT methods and their validation.
More details about the technical work can be found in the results section below. Two videos about the project activities were recorded in September, in which the three project partners (Sheffield, Sheffield Children’s Hospital (SCH), and Oxford) collaborated. Details of these and other deliverables can be found in the deliverables section.
The IoT4 Healthy Sleep project developed:
1. A recursive method which can be applied to the videos of patients and achieves a high level of detection and classification accuracy.
2. A framework for autonomous detection of sleep apnoea, using peripheral blood haemoglobin oxygen saturation (SpO2) data based on the fusion of multiple features and machine learning methods. The SpO2 signals are segmented into overlapping sub-sequences and several features are extracted from each segment. The distributions of features extracted from disorder and normal segments are modelled by two Dirichlet process mixture models, respectively.
3. Automatic apnoea detection by combining features extracted from both single-lead Electrocardiogram (ECG) and SpO2 signals using different algorithms of fusion and feature selection for accurate detection performance.
In our work, many features are extracted from ECG signals and several most informative ones are selected using the maximum relevance minimum redundancy (MRMR) algorithm to reduce the computational complexity. With the selected features of ECG and those extracted from SpO2 signals, three fusion methods are proposed for apnoea classification with a high detection accuracy, ie multivariate vector, expert vote and multiplied probabilities, together with the Dirichlet process mixture model.
4. An exploration of automated sleep staging from sleep recordings made by wearables. An evaluation of a Random Forest (RF) model for automated sleep staging was assessed on new data (simultaneous PSG and wearable data from a mixed cohort of participants with a mixed array of sleep disorders). Evaluation on limb movement that correspond to various sleep stages was performed.
Included in our datasets are 36 participants (diagnosed with RBD) from the John Radcliffe hospital, with two nights of clinical PSG recordings that include detailed notes on all movements and arousals.
A Dirichlet process and Gaussian mixture models were developed to automate the detection of limb movement using EMG sensors on the left and right limbs. This required an extraction of all text found in the annotations and the confirmation of notes associated with limb movement. These annotations were then used to extract EMG signal segments associated with limb movement to quantify the nature of limb movement duration and activity.
Our findings yielded a list of unique limb movement descriptors along with the average movement duration of 10s that occurs roughly 2s before and 8s after the annotation. This quantitative analysis allows us to extract and identify features to develop models to automate movement detection.
5. A framework for autonomous detection of limb movement. Signals from the left and right limb EMG sensors were used to segment ten second mini-epochs, which have been annotated as either having limb-movement or not (based on clinical notes).
Features extracted from each 10s mini-epoch included many previously used EMG features (time and frequency domain) along with the manually annotated sleep stage. These features were then used to train two Dirichlet process Gaussian mixture models that define a model associated with limb movement and one without limb movement.
For each model the log-likelihood for the left limb EMG was added to the log-likelihood of the right limb EMG signal, as both signals are considered independent. Feature selection was achieved through minimum redundancy maximum likelihood (mRMR). Using 10-fold cross validation this technique achieved an accuracy, sensitivity, and specificity of 95%, 50%, 96%, respectively for limb movement detection.
The low sensitivity can be attributed to notes on minor limb movement which were not strong enough to be captured by the EMG sensors. Nonetheless, specificity remained high and achieved a precision of nearly 25%. These numbers mean that movement detection could be achieved with a certain level of precision and in turn could be used to identify participants with RBD.
6. Movement detection as a metric for RBD identification. From analysing the distribution of automated limb movement, there was a clear distinction between healthy controls and RBD individuals. Simple metrics that measured movements per hour, movement per hour during REM, and total instances of movement were significantly different between a cohort of RBD and healthy control participants. Future work will look to include these metrics in automated RBD detection.
Deliverables and other tangible outputs
Project videos on YouTube:
Li, Z, Mihaylova, L, Arvaneh, M, Elphick, H and Kingshot, R. ‘Autonomous detection of obstructive sleep apnoea events using Dirichlet process mixture model and multiple features from single-lead electrocardiogram signals’, in preparation for a journal.
Cooray, N, Andreotti, F, Lo, C, Symmonds, M and Hu, M T M. ‘Proof of concept: screening for REM sleep behaviour disorder with a minimal set of sensors’, Clin. Neurophysiology, in preparation.
Li, Z, Mihaylova, L, Arvaneh, M, Elphick, H and Kingshot, R (2020) ‘Autonomous sleep apnea detection with a Dirichlet process mixture model and oxygen saturation data’. Proceedings of the International Conference on Information Fusion (Fusion 2020), ISIF, South Africa, 6–9 July 2020.
Cooray, N, Li, Z, Wang, J, Lo, C, Hu, M, Arvanez, M and Mihaylova, L (2021) ‘Automated movement detection with a Dirichlet process gaussian mixture model and electromyography’. Proceedings of the International Conference on Information Fusion, ISIF, South Africa, 12–15 July 2021, in preparation.
Li, Z, Arvaneh, M, Elphick, H, Kingshot, R, Cooray, N, Hu, M and Mihaylova, L (2021) ‘Autonomous classification of sleep disorder patterns with oximetry and ECG data’, Child Health Technology, March 2021.
Codes available on ORDA for the conference paper ‘A Dirichlet process mixture model for autonomous sleep apnea detection using oxygen saturation data’.
Hospitals across the UK conduct studies to investigate and treat sleep disorders. Brain signals, oxygen levels and other data are collected mainly in hospital environments and used for diagnostics. However, the data collection and data processing to a large extent is manual. This project is about how the Internet of Things and machine learning can make a step change and support both patients and clinicians.
This project was a great opportunity to establish a collaboration of the University of Sheffield, Sheffield Children’s Hospital (SCH) and the University of Oxford. The two research groups have successfully coordinated their efforts to disseminate their respective expertise in new healthcare applications. Not only has this venture increased the knowledge and skill-set of each group but provided new applications to explore, which have the potential to provide novel healthcare solutions.
We have received ethics approval for the sharing of data from children with apnoea collected by Sheffield Children’s Hospital. We developed algorithms that can be used for automatic detection and classification of sleep disorders and have demonstrated them on real data. If implemented in practice, the methods could reduce significantly the clinician time needed to analyse multiple different sources of data – EEG, ECG, oximetry and video.
The study was expanded to rapid-eye movement (REM) sleep behaviour disorder (RBD), where mounting evidence suggests that this disorder predicts Parkinson’s disease (PD) by years, potentially decades. Hence, these two areas have the potential to efficiently analyse and predict disease before it is fully revealed. This could have a significant impact on diagnostic processes, on medical treatments and, ultimately, on patient outcomes.
Hence, facilitating data collection processes could make a huge difference in this case: transferring the monitoring process to easier technology that can be used in the home could represent a significant benefit to children and their families. Moreover, the technology that we are developing is applicable both to children and adults. It could also contribute to bringing about behaviour changes.
The project aims not only to detect sleep patterns, but is part of a bigger study that aims to identify whether there is a health problem, where and how severe it is, and what behaviour change is necessary to mitigate its effects.
Data collection currently requires data to be collected during overnight stays in hospitals and clinics. For children in particular this is a big challenge, exacerbated by the fact that they need to wear electrodes on their heads.
Machine learning and Internet of Things can make a huge difference and we have shown it in this project. In the two videos that we created we demonstrated what the key challenges are and how we suggest to solve them.
The publications that have been written and the publicly available code stimulate reproducible research.
The work acts as a significant demonstration of IoT capability in the area.
This work was accomplished during an challenging time in human history, during a worldwide Covid-19. This probably constituted the greatest challenge to the project, to work within the limitations and confined of the pandemic. It’s a great testament to the collaborative teams and modern technology that this challenge was overcome and did not falter during these testing times.
The process of gaining ethical approval from the NHS for the sharing of data took nearly a year. It is not clear what could have been done differently to speed this process, but it is thanks to having SCH collaborators that it was a success, and that we have now the data and the plan is to use them in future projects.
What has Pitch-In done for you?
Pitch-In gave us fantastic opportunities to create new co-operations. This multidisciplinary project provided amazing opportunities for collaborations between data scientists, system engineers, as well as clinicians, medical doctors and psychologists. With the support of Pitch-In, funded by Research England, this collaboration has happened.
In this project, we also created a framework about how to address the barrier of acceptance of IoT by clinicians and patients through a co-design and co-creation approach, where the stakeholders get familiar with this technology and discuss their concerns and expectations, including data privacy. Hence the project has a big potential to make impacts in the health, social and economic areas.
Professor Lyudmila Mihaylova – the University of Sheffield