A tinyML solution won the honorable mention at the 2022 ICCAD TinyML Design Contest

On Nov.2 Luigi Capogrosso, PhD student on tiny ML, received the good news. The first challenge on Tiny Machine Learning has seen INTELLIGO Labs as the 8th team over 150 candidates all over the world! The challenge: put a deep learning classifier on a extremely low power budget, memory footprint, low latency and high accuracy on extremely resource-constraint embedded platforms. the classification challenge was to to discriminate life-threatening ventricular
arrhythmias (VT & VF) over each 2s r intracardiac electrograms (IEGM) segment

Two months before the deadline: we submitted the code to be run on a flash memory, with the following constraints <128K, power <1 uW, latency < 5ms, accuracy > 90% 

“The first ICCAD TinyML contest on healthcare application. The first challenge for INTELLIGO Labs on Tiny Machine Learning. Achieving the 8th place + honorable mention over 150 participants has been rewarding”

Heart disease is the leading cause of death in US, according to CDC 2020 Mortality Report. More specifically, Ventricular Tachycardia (VT) and Ventricular Fibrillation (VF are the two most frequent recorded shockable rhythms causing Sudden Cardiac Death (SCD). The treatment: as primary Prevention, Medical or interventional therapy undertaken to prevent SCD in patients who have not experienced symptomatic life-threatening sustained VT/VF. 

As secondary intervention, and where we act, is to implant a Cardioverter-Defibrillator (ICD), a small battery-powered device to deliver defibrillation for patients who are with a history of life-threatening ventricular arrhythmias. 

Existing industry practice is simple rule based, and has not changed over the past a few decades. AI/ML can potentially revolutionized ICD design by extracting features not easily identifiable by or even unknown to human experts. While many machine learning competitions exist, they mostly focus on computer vision only and does not utilize hardware platforms to such an extreme limit (tinyML).

Our solution was based on a 1D CNN with many tricks, which will be subject of a paper we are currently writing. A simple idea, but definitely effective, to pave the way toward future advancement on tiny ML.

Lascia un commento

Il tuo indirizzo email non sarà pubblicato. I campi obbligatori sono contrassegnati *