Designing AI for defibrillators, UB students rank highly in TinyML competition

The UBPercept team earned fifth place out of more than 150 entrants in the TinyML Design Contest, a multi-month research and development competition sponsored by the Institute of Electrical and Electronics Engineers (IEEE) and the Association for Computing Machinery (ACM). It concluded with an awards presentation last month at the IEEE/ACM International Conference on Computer Aided-Design (ICCAD) in San Diego, Calif. 

Teams were tasked with developing machine-learning algorithms to help implantable cardioverter-defibrillators (ICDs) better detect ventricular arrhythmias, an abnormal heartbeat that can cause sudden cardiac death. 

“The goal of this competition was to challenge researchers worldwide to design a new machine-learning algorithm that can potentially revolutionize the design of ICDs,” says Jinjun Xiong, one of the UBPercept team’s advisors and a SUNY Empire Innovation Professor in the Department of Computer Science and Engineering. “I believe this provided a great learning experience for our students, and sparked a new round of innovation on this important topic.”

ICDs are battery-powered and approximately the size of a pocket watch. They’re implanted in the upper chest of those at high risk for cardiac arrest and send an electric pulse to the heart when an irregular heartbeat is detected. They can also record and store data on a patient’s heart rhythms. Approximately 800,000 Americans have an ICD and at least 10,000 more receive them every month. 

A U.K.-based study recently published by the British Medical Association found artificial intelligence could eventually increase ICD speed and reduce the number of inappropriate shocks, leading to improved quality of life for patients. Even small improvements could save hundreds of lives, the study said. 

However, ICDs and other medical devices currently use a simpler form of A.I. called rule-based A.I., which can’t adapt well to an individual patient’s needs.

“But machine-learning-based A.I. can learn the hidden patterns and features from a patient’s data automatically,” Xiong says. “This has the potential to greatly improve accuracy and reduce delays.”

The TinyML Design Contest required each team’s machine-learning model to differentiate irregular heartbeats within thousands of intracardiac electrograms collected from real ICDs. Each team’s submission was measured by three metrics: accuracy, latency and memory footprint. 

The UBPercept team had the best latency – the time it takes a computer to act after receiving a signal – of any team in the competition, while ranking third in memory footprint and ninth in accuracy. 

“Generally, there is a trade-off among accuracy, memory footprint and speed,” says computer science PhD student Changjae Lee. “Our team decided to submit a design that strikes the appropriate balance among those metrics.”

The key was a simplistic design, adds Pranay Meshram, who is pursuing an MS in robotics. 

“Though we started with pitching complex ideas, at the end of the day the simplest solutions outperformed,” Meshram says. 

The team also included Tianchen Yu, who is also pursuing an MS in robotics. The team was co-advised by Karthik Dantu, an associate professor in the Department of Computer Science and Engineering. Xiong advised Lee and Yu, while Dantu advised Meshram. 

Dantu is a recent awardee of the Stephen Still Institute for Sustainable Transportation and Logistics' Sponsored Pilot Research Program. The results of the TinyML Design Contest are related to Dantu's ongoing project scope.

Although machine learning in ICDs has enormous potential, there are some limitations, including high processing power and cost.

“When the technology is more mature, I expect more medical devices will overcome these stringent resource constraints and be driven by advanced A.I. algorithms,” Xiong says.