The COVID-19 global pandemic, which has infected over 35 million people worldwide, has undoubtedly had one of the most significant impacts on the modern healthcare structure, accelerating the adoption of Machine Learning (ML) powered solutions. Whether ML can live up to its hype and become a transformational force in healthcare depends heavily on platforms on which these applications are deployed. The key to success lies in prioritizing human needs and concerns first, ensuring high performance and quality of service along with the low cost and the necessary patient privacy.

Problems Faced in Healthcare Systems

Fast adoption of ML-based medical image analysis and other tools which can revolutionize medical practice is inevitable, as we all experienced during the last several months. The tragedy of rapid spikes in confirmed COVID-19 virus cases has revealed weaknesses of healthcare systems, including overloaded and tired practitioners, late diagnoses and a lack of tools for self-evaluation and support during quarantines. Many of these problems can be mitigated by automating and accelerating medical processes at the edge. That’s why we observe a big shift in ML-based healthcare towards embedded solutions which can process critical workloads regardless of cloud connectivity and operability.

Embedded systems are emerging as the preferred way to deliver medical services with very low latency and minimal power consumption while preserving data security and associated privacy. The use of specialized ML embedded systems is crucial to quickly respond to extreme situations like the world is facing right now.

Performance and Low Power at the Edge: Key to Cost-efficient and Timely Diagnosis

As the COVID-19 virus spreads, we are all forced to face the new reality and look for innovations in medical solutions. Availability of datasets was often a limitation for ML research, but recently more and more imaging resources have been compiled together to promote active studies on automated image analysis solutions [1]. The promising findings of several studies on Computed Tomography (CT) scans have highlighted the growing role of this approach for detection of COVID-19 infection cases [2].

The use of ML has a big impact in this area, allowing for analysis of large amounts of data within seconds vs. up to two days of standard tests [3]. Yet, as specified by Philips Research [4], the big challenge for imaging systems remains performance. Cloud computing, which is frequently used for processing, suffers from limitations of conventional data transfer technology that affect large data volumes, e.g. raw imaging modalities data or genomics files.

In order to ensure real-time feedback and address the increased demand for timely diagnosis, it’s important to expedite data analysis at the edge. Applications at medical imaging workstations often rely on dedicated graphics processing units (GPUs) [5], but the use of specialized ML edge platforms is even more beneficial due to the additional performance gain at a lower cost and decreased power consumption.

Real-time Processing Enables Accurate, Multimodal Personalized Healthcare at the Edge

Apart from faster and more efficient data analysis, embedded systems are crucial for ensuring our mind and body wellness during crisis situations. One of the problems revealed during the pandemic is associated with delivering healthcare and preserving social distancing at the same time.

Boston Dynamics proposed to solve this problem with their dog-like robot, Spot, which allows for measuring vital signs in a contactless way [6]. Another patient support system, PETRA, was developed by Merck AB in collaboration with Furhat Robotic [7]. With these solutions, the health of physicians doesn’t have to be put at risk, as they can perform diagnosis without personal contact with patients. Nowadays, cloud data centers are a popular choice as a computational backend for robotics due to limitations of local compute resources. In the future, with the increased computational capabilities at the edge, it will be possible to become independent of cloud connectivity.

COVID-19 has also greatly affected our mental wellness. Self-isolation has an especially huge impact on elderly people, causing deterioration of their health and raising the risks of various diseases and heart problems [8]. This creates a demand for self-monitoring systems and autonomous talking robots like the Pepper robot [9], which can become our daily assistants.

This kind of medical support is crucial during the pandemic, as many people feel confused and don’t have access to tools which would allow them to perform an initial self-evaluation of symptoms, such as elevated temperature or change in breathing patterns estimated from thermal camera streams [10] Continuous processing and contactless monitoring of medical data collected from various cameras and other sensors (e.g. pulse oximeters, spirometers, thermometers, medication monitors, etc.) help people feel less anxious and in more control, having a positive impact on mental health. Analysis of multimodalities is also essential at medical facilities in order to provide more personalized treatment options and ensure comprehensive responses.

Currently, power utilization is one of the most important limitations for different ML applications. Reducing power comes at the cost of decreased performance, which is unacceptable for health monitoring applications where every second counts. Specialized low-power ML at the edge is necessary to deliver low latency responses to users and what’s needed are architectures with software that are better than 10x of what is available today in the market for performance in frames per second. In addition, it’s important to fully utilize capabilities of multi-sensor platforms and enable ensemble reasoning using multiple workloads for improved accuracy without sacrificing performance. Thus, the solution needed will address this problem by supporting real-time execution of parallel pipelines at the edge. Power efficiency is another key performance indicator of the desirable solution which should ensure real-time, accurate communications and decrease the number of required battery recharges, improving user experience.

Security Features in ML Embedded Platforms Making Remote, Real-time Healthcare Possible

The COVID-19 pandemic has spurned a tremendous amount of interest in home-based and non-contact diagnostics and propelled the healthcare industry to adopt remote diagnostics at warp speed. This will forever change the healthcare landscape as noted in McKinsey’s report [11], citing the post-COVID-19 reality will include the use of telehealth solutions on a daily basis, and going forward this branch of healthcare could potentially account for $250B of business for the healthcare industry. The overall willingness to use telemedicine has also changed, permanently shifting the standard medicine model to virtual services. 60% of people surveyed by Accenture expressed the need for telemedicine utilization in the post-COVID-19 future [12].

This indicates that remote medical solutions will be in greater demand as people become accustomed to using them during the current reality. This is why it’s so important to boost the capabilities of remote healthcare by deploying high performance solutions that have excellent quality of service while having the necessary security to give users that confidence to use these in their homes. Once these are in place, these innovative high performance embedded systems will act as home medical assistants, personal emergency response systems or gateways aimed at processing data from multiple transmitters at the edge [13]. Key to enabling these revolutionary advancements are low-power high-performance embedded systems capable of processing multiple inputs, which will prove their capabilities in these medical use cases.

However, as the adoption rates of telehealth skyrocket, new problems have emerged and these factors may potentially stall progress [14]. The first problem lies in the availability of remote healthcare services. As specified in The Rural Telehealth Initiative stemming from President Trump’s executive order [15], telemedicine is expected to support people living in rural areas even after the pandemic ends. However, 14.5 million people still lack access to the fixed broadband internet service [16].

Embedded systems are crucial for providing medical solutions to rural areas, eliminating the need for cloud connectivity and providing necessary diagnostics even when there is no internet service available. However, according to the Accenture 2020 [17], the increased telemedicine popularity could slip in the future due to concerns about data security and the fact that consumers will revert to prior expectations about quality of deployed health IT systems after the pandemic. Healthcare systems have recently faced many more security breaches and leaks of sensitive patient information [18]. According to the U.S. Department of Health and Human Services [19], the number of cases under investigation from March 2020 until now has doubled compared to the same period of time last year, e.g. the recent failures of computer systems for Universal Health Services – one of the largest cyberattacks in US history [20]. This might have been caused by the need for fast development of IT healthcare solutions where the rapid introduction created some security issues. The Department of Health and Human Services, Office for Civil Rights decided to temporarily waive penalties for noncompliance with the regulatory requirements under the HIPAA Rules [21] as long as good faith provisions of telehealth are applied. However, to make sure the growth of the telehealth trend is not derailed, there is a need for addressing security and privacy concerns.

Although HIPAA regulations were temporarily loosened for telehealth because of the pandemic, the need to comply with them is expected to be enforced again. Embedded systems are emerging as an appropriate safeguard to ensure security of collected data, protecting patient privacy, and reducing the risk of data breaches. To make sure the idea of local and secure medical data analysis comes true, it’s crucial to deliver high performance edge architecture and software with encryption, local data processing, user authentication, control over what code is authorized to run, support for rolling logging, and minimal number of required data transfers, e.g. only in emergency situations or in case of firmware updates.

The Future

The problems facing our healthcare systems won’t be easy to fix overnight, especially as the pandemic continues. However, the good news is is developing an advanced machine learning platform that integrates high performance, security and low power capabilities targeted for ML-enabled computer vision workloads. This unique technology will be a key for the success of various healthcare applications.

Specialized ML-powered embedded systems will promote a better patient experience by enabling the smooth processing of medical data, providing efficient power utilization and improved data security, and accelerating adoption of automated healthcare. We know embedded technologies have the opportunity to really change lives. In the future, we will further explore these applications and see how supporting high performance ML at the edge can continue to deliver many benefits in the medical market.


About the author: Alicja Kwasniewska is a software architect at specializing in Artificial Intelligence applications for embedded edge. Her expertise in the areas of Computer Science and Deep Learning for Biomedical Engineering are foundational as she works to develop solutions for autonomous driving, smart home and remote healthcare applications. She is also a co-founder of the International Summer School on Deep Learning educating future generations on the fundamentals of deep learning methods. She frequently presents her work at international conferences and has 20+ publications in the field, which were awarded with best paper and best young professional paper awards. In her PhD dissertation she proposed AI solutions aimed at improving the accuracy of the contactless vital signs estimation from low resolution thermal sequences.