VATML: Towards On Device Ventricular Arrhythmia Detection Using TinyML

Vipin Gautam, Sharad Sinha, Shitala Prasad

Abstract: Cardiac arrhythmias pose a significant health risk, underscoring the critical need for precise detection. This paper introduces “VATML: Ventricular Arrhythmia Detection using TinyML” demonstrating TinyML's potential to enable detection within implantable cardioverter defibrillators (ICDs), a common medical device. TinyML i.e. deploying machine learning on resource-constrained edge devices, holds promise for healthcare transformation. VATML features a lightweight 1D Convolutional Neural Network (CNN), tailored for life-threatening Ventricular Arrhythmias (VAs), with a minimal 17.33 KiB memory footprint and low Multiply-Accumulate (MAC) complexity of 11.51K, meeting ICD operational demands. Experimental results showcase VATML's notable performance, achieving a high $\mathbf{F}_{1}$ score of 98.33 % and a generalization score of 93.75 % on the STM32F303K8 board with an ARM Cortex-M4 processor compared to competing methods.

External IDs:dblp:conf/ises/GautamSP24