Sensors and Machine Learning: Glucose Monitoring with An AI Edge

6747
AI algorithms related to glucose sensing improve the accuracy and performance of continuous glucose monitoring devices. (GETTY IMAGES)

Medtronic’s mission is to alleviate pain, restore health, and extend life through the application of biomedical engineering, explains Elaine Gee, PhD, Senior Principal Algorithm Engineer specializing in Artificial Intelligence at Medtronic. It’s a mission Gee is well equipped for. With over 15 years’ experience in modeling, bioinformatics, and engineering, she drives machine learning algorithm development and analytics to support next-generation medical devices for diabetes management.

On behalf of AI Trends, Ben Lakin, from Cambridge Innovation Institute, sat down with Gee to discuss her most recent focus: algorithm development related to glucose sensing to improve the accuracy and performance of continuous glucose monitoring devices, also known as CGMs.

Elaine Gee, PhD, Senior Principal Algorithm Engineer at Medtronic

Editor’s Note: Gee will be giving a featured presentation on Advancing Continuous Glucose Monitoring Sensor Development with Machine Learning at Sensors Summit in San Diego, December 10-12. This conversation has been edited for length and clarity.

AI Trends: CGMs are a hot topic in the medical device field; they’ve revolutionized monitoring for diabetes patients. Where do Medtronic’s CGM algorithms operate (edge computing, smartphone, the Cloud)? How is Medtronic utilizing new technology to benefit their CGM users?

Medtronic transforms raw user data into personalized insights by various computing methods, from computing on device to computing online. The most advanced sensor from Medtronic is the Guardian Sensor 3. This sensor powers the Guardian Connect CGM. An algorithm performs computations on the device to provide readings every five minutes of user glucose levels and sensor health. By performing analytics in physical proximity to the source where data is generated, reliable real-time insights are available to the user.

Users of the Guardian Connect standalone CGM also receive access to the personal Sugar.IQ diabetes smart assistant. This combines artificial intelligence with diabetes expertise in an app that analyzes daily glucose patterns and factors that affect them, such as food intake or exercise. This informs the user in real-time to encourage around-the-clock glycemic control.

Lastly, users of Medtronic’s MiniMed 670G Insulin Pump System can access the secure web-based CareLink software. This software processes sensor and pump information uploaded by the user to generate personalized insights by highlighting data trends and associating relationships between glucose levels and insulin usage, carbohydrate intake, exercise, and medication to help identify patterns for better glycemic control. The CareLink system also allows collaboration between healthcare providers and patients. With the CareLink Pro software, patients may grant access for providers to view and even download data into their EMR.

These devices must meet numerous design requirements to be considered safe and effective. What are some of the greatest engineering or technical challenges in developing successful algorithms?

Multiple elements come into play when developing a medical device that can pass regulatory standards to get to market. The number one challenge in algorithm design is having enough data. Algorithm development relies on data in every aspect of the development cycle, from the very start with prototyping, to training and optimizing the algorithm, through to testing and validation. Algorithm development begins with proper experimental design to ensure that the data collected to train the algorithm contains the necessary information. Machine learning algorithms rely on high-quality data with proper information density to support learning complex models in a way that avoids over or under fitting. The goal is to create a machine learning model that generalizes well to data not encountered during training to ensure safe use on the market. All these steps, in addition to the regulatory review process, are necessary to verify that the trained algorithm meets the rigorous requirements for safe and effective use in medical device applications.

With regards to future upgrades, what do you think are advantages to pursuing software upgrades or algorithm changes, as opposed to hardware improvements? Is there any difference when it comes to regulatory oversight?

With respect to regulatory oversight, both software and hardware upgrades will trigger a regulatory review. As far as product development, each type of upgrade can improve the CGMs performance and usability. Typically, the R&D process for developing a new algorithm may have a faster turnaround time as compared to the process for supporting a hardware update for various reasons.

First, algorithm development requires large datasets to support training and learning the model. In the cases where training data is unavailable, clinical trial studies are required to collect new data for algorithm development. In this case, this time intensive process can add to the algorithm development timeline as clinical trials must be arranged, subjects enrolled, and data collected.

In terms of algorithm design and experimentation, these steps are done in silico and can be sped up by taking advantage of parallel processing on scalable compute infrastructure. Parallel processing can explore changes to the algorithmic architecture and the parameter space more efficiently to reduce the development time.

Compare this process to making changes to CGM sensor chemistry or other hardware elements. Hardware upgrades require both in vitro and in vivo testing to evaluate performance. In addition, there are manufacturing considerations to address, such as sensor assembly, packaging, sterilization, and shelf life.

Implementing and commercializing new algorithms for CGMs can be more straightforward than hardware changes in some cases, but nonetheless creating a new algorithm requires significant time in development, testing, and validation to ensure the analytic performance supports its medical device use claim and to ensure safe and effective use.

A topic that’s frequently in the news is cybersecurity. With even medical devices at risk from the unscrupulous, how is Medtronic protecting patient data?

I’m glad you asked this, because protecting patient data is a high priority for Medtronic. Cybersecurity is a critical element we incorporate into our development process from the beginning. We take necessary measures to ensure patient data is secure during generation, in transit, and when being stored. Our internal processes guide us to use industry best practices and state-of-the-art approaches. We work with security experts to confirm that our products meet our rigorous criteria, and we closely monitor our products and systems to ensure ongoing protection through the lifecycle of the device.

I’m curious about the possibility of wearable sensors being applied to other conditions. Do you think there are other conditions or medical applications that can benefit from continuous monitoring and real-time data analysis?

Wearable sensors have become accessible and widespread. Real-time data is collected in various forms, whether from direct-to-consumer activity trackers like Fitbits, Apple Watches, and smartphones, or collected from continuous monitoring medical devices like the CGM that we have here at Medtronic Diabetes. With the advancement of wearable sensor technology, consumers and healthcare providers have access to a diverse range of continuous physiological signals. This is exciting because it creates opportunities for new applications of real-time data analysis that could benefit users living with chronic conditions. For example, in the case of diabetes, CGMs and insulin pumps have given users greater control in diabetes management by providing actionable, real-time information. Continued development of sophisticated decision support systems that leverage diverse input signals could give users living with chronic conditions even greater control over their health by generating more actionable insights in real-time.

Are there any new algorithms or hardware technologies on your radar? As an industry scientist, what are you excited about right now?

There are so many technological advances I could list, but given our recent discussion on edge computing I will highlight advances in computational processing by AI microchips. Machine learning algorithms rely on complex computations, placing high demands on the compute, memory, and storage of the hardware chipset. For example, deep learning algorithms frequently perform high-dimensional matrix calculations and thus require large on-board memory to support operations with the hyperparameters and model variables as well as high computing density to support calculations for rapid model inference. Typically, portable wearable devices are battery powered putting constraints on power usage. In order to advance the capabilities of AI-driven wearable devices, algorithms must become more efficient in footprint and power usage, and/or chips must become more powerful and efficient without driving up cost. Algorithms that are more efficient at complex modeling can operate on a smaller hardware footprint, while a faster and cheaper microchip would support high-density calculations at reduced power and cost. Research and development in the area of low-powered AI chip design is exciting because advancements in microchip technology on the hardware side will expand the possibilities for heavier machine learning workloads on the software side.

Learn more about AI at Medtronic.