CESAR - Using Complex Event Processing for Low-threshold and Non-intrusive Sleep Apnea Monitoring at Home

The Cesar project (2017-2021) started with the hypothesis that low-cost solutions based on smartphones, wearables, and machine learning (ML) can contribute to decrease the number of undiagnosed sleep apnoea (SA) patients (which is currently around 80%). The project team at Ifi and the Oslo University Hospital (OUS) could give through a clinical study strong evidence that the hypothesis is correct. To achieve this many challenges had to be addressed in the entire pipeline from data acquisition, quantifying the quality of data from wearables, to finally create ML models that can properly classify sleep monitoring data with respect to SA events. With the support from the A3 study at the OUS and St. Olavs Hospital we could achieve some major breakthrough towards this goal. By using the sleep data collected in the A3 study from 579 patients with Polygraphy (PG) we could show that (1) ML based analysis comes rather close to the results of sleep experts, (2) approximately 300 sleep recordings are enough to train modern Deep Learning (DL) to achieve the maximum SA detection performance, and (3) different DL approaches achieve nearly the same maximum performance. But one of them, i.e., Convolutional Neural Networks (CNN), outperform all others because CNN need much less training data to achieve maximum performance and use much less resources for training and testing. Furthermore, we could show that just by using one instead of all four channels in PG, like a single respiration belt, it is possible to come close to maximum performance. This is good news for patients, because wearing a single sensor is much more convenient than wearing four sensors. The last 47 A3 patients volunteered to use in addition to the PG the Flow sensor from the Norwegian company SweetZpot, which is a simple respiration belt intended for athletes at the costs of 2000 NOK. This allowed us to analyze to which degree the Flow sensor can be used for SA detection. With DL we achieved for Flow data an accuracy of SA detection close to 80% accuracy, using the evaluation of the PG data by a sleep expert as the ground truth. In comparison, we achieved with DL for PG data approximately 90% accuracy. Another important result is the fact that to train a DL model for Flow data analysis we do not need a large amount of Flow data, because we can use the PG data instead. This is very important for any future study that investigates a new/improved sensor, because it is not necessary to collect for every new sensor large amounts of data with that new sensor (which is very cumbersome and expensive). Potential improvements of the Flow sensor and the software for data acquisition have been identified such that future versions of this sensor should lead to even better results. Our initial investigation into the use of oxymeter data from recent smart-watches to support the SA detection are promising, but require more work to conclude with thorough scientific results. Access to sufficient training data in terms of quantity and quality is central for ML, but especially hard in the medical domain, because medical experts need to label the data and can make mistakes and most data cannot be shared due to privacy concerns. For both of these challenges the Cesar project has developed new solutions to: (1) better learn and perform with data containing noisy labels, and (2) transfer knowledge learned from a private data set without revealing information that could be used to identify an individual that contributed to the private data set.

More interdisciplinary research for the project members and 22 master students. Increased international visibility of the project members in the domain of machine learning (ML) for sleep medicine. Two follow-up research project proposals have been submitted. Low-cost sleep apnea (SA) detection at home with a respiration belt, a smartphone and ML is possible, which is very patient friendly and promises high compliance for longitudinal studies. The time it takes to diagnose SA patients and the number of undiagnosed SA patients can be substantially reduced. Large scale longitudinal sleep studies can be performed at substantially lower costs and lead to new insights in sleep medicine. Quality of life for SA patients will be improved. Costs in the health sector related to SA can be reduced. Introduction of new respiration belts for SA detection will be easier and more cost efficient. New business opportunities for sensor and app producers.

The project performs interdisciplinary research, with three partners from the medical domain and experts in the areas of Obstructive Sleep Apnea (OSA) and next generation of sensors for medical use; and two partners from computing with expertise in mobile systems, sensor data acquisition and processing, signal processing, data analysis, and event detection. The application requirements are determined by the medical experts that will also perform user studies. An extensible data acquisition system will be implemented with smart phones and sensors, like Shimmer motes and the Bitalino sensor set. This system will be used to collect longitudinal data from sleep monitoring at home and in the sleep laboratory (combined with classical polysomnography to annotate the ground truth). Supervised learning (data mining) techniques will be systematically studied for their use to automatically analyze longitudinal data for OSA detection. These studies will use data from the PhysioNet databases (early project phase), and later-on data that has been collected in user studies with the data acquisition system. Furthermore, we investigate the usefulness of supervised and unsupervised learning (data mining) techniques to identify interesting data patterns that might lead to new knowledge in OSA research and to support the design and engineering of the on-line analysis tool. The design of the on-line analysis tool is driven by the goal to enable individuals with limited computing skill to customize and personalize the on-line analysis. To achieve this goal, the following three principles will be strictly applied: use of a declarative approach with Complex Event Processing, using few powerful abstractions of physical and logical sensors, and a fine granular modularization implemented in sensor hierarchies. Furthermore, the team will build tools to quantify the quality of off-line and on-line data analysis results.