Extraction of novel signatures to improve the diagnosis of obstructive sleep apnoea
Dr Manasi Nandi
Reader in Integrative Pharmacology
Cancer and Pharmaceutical Science, King’s College London
Professor Joerg Steier (Consultant in Respiratory Medicine KCL and GSTFT);
Professor Philip Aston (Professor of Mathematics, University of Surrey)
Dr Ged Rafferty (Reader, Human Physiology, King’s College London)
This project will involve the analysis of thousands of pre-collected respiratory and pulse waveforms from patients with and without a common disease called obstructive sleep apnoea (OSA). The student will apply a novel mathematical method (symmetric project attractor reconstruction) which transforms complex waveforms into simpler images. These images allow the shape and the variation of any input waveform to be readily quantified. By extracting these new features from conventional waveforms, we can improve the sensitivity of detecting subtle changes in the body. The aim is to establish whether this mathematical method can more accurately discriminate between patients with OSA from those with other respiratory disorders, ensuring more rapid diagnosis and timely treatment. This project would provide proof of concept data to develop the technology as a diagnostic device.
OSA prevalence increases with age and affects 2-10% of adults. Individuals experience repeated temporary cessations of breathing during sleep, due to obstruction of the upper airway. This results in sleep deprivation and daytime drowsiness and is associated with significant cardiovascular morbidity and mortality. Undiagnosed OSA could also result in peri-operative morbidity or mortality in patients undergoing anaesthesia or sedation. OSA is treatable, once diagnosed.
Diagnosis is normally made in specialised sleep centres – where continuous high-fidelity physiological waveforms (e.g. ventilatory flow, pulse oximetry) are simultaneously collected from patients. Analysis is complex, often taking expert technicians 1-2 weeks to complete. Given the high prevalence of OSA in the adult population and the limited number of bed spaces in sleep centres, it remains an underdiagnosed pathology. Therefore, efforts are being made to develop cost efficient technologies to a) roll out initial diagnosis in the primary care setting – achieved via wearables on patients in their homes b) speed up diagnosis in patients subsequently referred to sleep centres. The student will therefore primarily focus on pulse oximetry waveforms, as these can be collected from wearables and sleep centre devices.
This student will be part of an interdisciplinary team comprising physiologists, mathematicians, clinicians and sleep centre technologists. St Thomas’ sleep centre has collected over 5000 patient records. These will be fully anonymised and waveform data extracted and split into training data and test set data comprising mixture of pre-diagnosed OSA and non-OSA patients. The training waveforms will be processed via existing mathematical coding to generate corresponding images of the waveform data. Features of these images will be quantified through development of new coding. The student will establish whether there is a ‘signature’ that discriminates OSA from non-OSA patients. The student will optimise machine learning methods to determine the accuracy of binary classification of OSA and non-OSA. The student will also be encouraged to develop deep learning approaches to classify the images. OSA detection algorithms will be tested using unseen test set data. At the end of the project, we hope to develop a bespoke algorithm for OSA detection which can be easily used by clinicians/technicians.
The student will regularly engage with the interdisciplinary team to ensure the correct clinical questions are being asked and the appropriate mathematical methods are employed. Attendance to relevant training courses, conferences, networking and public engagement events will be encouraged and supported.
Key ref. links:
This is a data study using retrospective non-sensitive patient data. St Thomas’ sleep centre has collected thousands of patient datasets. The high fidelity numerical waveform data (pulse oximetry, respiratory flow) can be fully anonymised. We are already in engagement with GSTFT to ensure GDPR and ethical compliance. We will also engage with the Caldicott Guardian and the HRA as and when required. We will ensure any ethical approvals are in place prior to commencing the project.
Machine learning, sleep disorder diagnosis, waveform analysis