New data-driven algorithms for intelligent remote patient monitoring systems from AIMLab at the Technion Faculty of Biomedical Engineering.
With billions of mobile devices worldwide and the relatively low cost of connected medical sensors, recording and transmitting medical data has become easier and faster than ever. Continuous and long-term dynamic physiological data can now be easily obtained. However, this ‘wealth’ of physiological data has seen very limited successes in being harnessed to provide actionable clinical information.
Part of the challenge is due to the high variability in data quality, the lack of standards for data representation (e.g., resolution, sampling frequency, and metadata), and the development, in many studies, of relatively small datasets which fail to capture the vast range of variability across patients and time.
Another part of the challenge is the lack of smart and robust algorithms that can decrypt the information contained in a large number of data points recorded over time, what is known as physiological time series. The development of machine learning algorithms combined with existing and novel wearable biosensors offers an unprecedented opportunity to improve the screening and tracking of an individual’s health, and support the management of patients’ conditions, particularly through remote health monitoring. Remote health monitoring relates to the monitoring of individuals outside the classical hospital environment, typically in their home.
The Technion Artificial Intelligence in Medicine Laboratory (AIMLab.) directed by Assistant Professor Joachim Behar develops innovative pattern recognition algorithms to exploit the information encrypted within large datasets of physiological time series. The AIMLab leverages these new data-driven algorithms toward the creation of novel intelligent remote patient monitoring systems.
Within its research scope, AIMLab recently published two innovative research works looking at continuous remote monitoring of heart and lung diseases:
y is a non-invasive method routinely used to monitor blood oxygen saturation levels. Low oxygen levels in the blood mean low oxygen in the tissues, which can ultimately lead to organ failure. Yet, unlike heart rate variability measures, a field that has seen the development of stable standards and advanced toolboxes and software, no such standards and open tools exist for continuous oxygen saturation analysis.
AIMLab offers, for the first time, such as standardization and an open-source toolbox for digital oximetry biomarkers. Using this new resource, the researchers demonstrated the value of such biomarkers for the diagnosis of obstructive sleep apnea – a common sleep-related breathing disorder, causing the upper airways to temporarily close off during sleep. The lab is now also applying this technology to the diagnosis and management of other respiratory diseases including COVID-19.
In a second study, AIMLab. evaluated, for the first time, the ability to estimate the percentage of time a patient spends in atrial fibrillation over long-term continuous recordings, using a Deep Recurrent Neural Network (DRNN) approach. This serves to better diagnose atrial fibrillation – a highly prevalent arrhythmia. The DRNN significantly outperformed traditional algorithms. Using continuous recordings for a 24-hour window versus one that Is just a few seconds long, which is the current standard method of operation, improved the diagnosis performance of the system by 16%.
The work described in this research is performed within Asst. Prof. Behar’s laboratory by students Armand Chocron (M.Sc. candidate, EE), Shany Biton (MSc candidate, BME), Jeremy Levy (M.Sc. candidate, EE), and Jonathan Sobel (Postdoctoral fellow, BME).
Associated publications of the lab: