Abstract— Electrocardiogram (ECG) is one of the fundamental markers to detect different cardiovascular diseases (CVDs). Owing to the widespread availability of ECG sensors (single lead) as well as smartwatches with ECG recording capability, ECG classification using wearable devices to detect different CVDs has become a basic requirement for a smart healthcare ecosystem. In this paper, we propose a novel method of model compression with robust detection capability for CVDs from ECG signals such that the sophisticated and effective baseline deep neural network model can be optimized for the resource constrained micro-controller platform suitable for wearable devices while minimizing the performance loss. We employ knowledge distillation-based model compression approach where the baseline (teacher) deep neural network model is compressed to a TinyML (student) model using piecewise linear approximation. Our proposed ECG TinyML has achieved ~156x compression factor to suit to the requirement of 100KB memory availability for model deployment on wearable devices. The proposed model requires ~5782 times (estimated) less computational load than state-of-the-art residual neural network (ResNet) model with negligible performance loss (less than 1% loss in test accuracy, test sensitivity, test precision and test F1-score). We further feel that the small footprint model size of ECG TinyML (62.3 KB) can be suitably deployed in implantable devices including implantable loop recorder (ILR). I. INTRODUCTION With the advent of off-the-shelf sensor technologies coupled with advancement and rapid developments of Artificial Intelligence (AI) techniques, Cardiovascular Disease (CVD) detection using single lead ECG is becoming increasingly popular. Proposed techniques are proven to be helpful in diagnosing CVDs including transient infrequently arrhythmias especially Atrial Fibrillation (AF) and rhythm monitoring. Portable ECG devices currently offer an efficient screening option for AF; generating comparable performance to 24 hours Holter monitoring [1]. A study using Kardia band (single lead ECG) demonstrated moderate diagnostic accuracy when compared to 12-lead ECG analysis. The study also concluded that combining the automated device diagnosis with Electrophysiologists’ (EP) interpretation of unclassified tracings yielded improved accuracy. Future improvements in automated algorithms were required with physicians’ involvement when exploring the utility of these devices [2]. Such diagnostic inference on single lead ECG often requires sophisticated deep learning (DL) models. We find two critical problems of running such DL based ECG diagnostics natively on the typical resource-constrained wearables: 1. Depending on the number of layers, size of a DL model may become too high. In order to be wearable ready, the model size requires to be as small as possible (preferably sub-100 KB) 2. The associated battery drain which also in turn demands for smaller and less compute heavy models. In this paper, we propose a piecewise linear approximation of a ResNet [3] based ECG diagnostic inferencing model (ECG TinyML) that takes 156 times less memory than the original Resnet model. The proposed ECG TinyML is 5782 times less computationally intensive compared to the baseline ResNet model, with almost no compromise in classification performance. The final reduced model takes less than 70 KB of memory, making it suitable for embedding into cardiac implantable devices like ILR. II. ON THE PROBLEM OF ECG CLASSIFICATION MODEL FOR WEARABLE DEVICES A. The Application Landscape As illustrated in [4], cardiac rhythm monitoring and management devices have had a large proliferation over the past decade, and hence the on-device detection of cardiac rhythm anomalies is an important problem to solve. Further, [5] provides overview of methods and challenges of analyzing single lead ECG (e.g. KardiaMobile [6]) for clinical outcome, which proves that the problem is non-trivial. The problem is further amplified if we need to perform the detection on a resource constrained device like a wearable or an implantable device that have limited memory and battery power. [4] provides an analysis different cardiac rhythm management devices and associated therapies in Australian market. Though 75% of those devices are traditional pace makers, other devices capable of defibrillation and cardiac resynchronization have found their places in the list. B. The Hardware Landscape ILRs and wearable devices are generally composed of tiny microcontroller units (MCU) and specific sensors [7]. The block diagram of a typical hardware system is shown in Fig. 1. It consists of a main MCU which reads data from connected sensors. A separate MCU or System-On-Chip (SoC) does the communication with external world using Bluetooth Low Energy (BLE) or other low energy communication protocols. Such a hardware setup will have both costlier and faster volatile memory (RAM) and slower but cheaper non-volatile memory such as an external flash. The main MCU in these devices is often severely resource constrained. They typically range from low-end 32-bit ARM Resource Constrained CVD Classification Using Single Lead ECG On Wearable and Implantable Devices Arijit Ukil 1 , Ishan Sahu 1 , Angshul Majumdar 2 , Sai Chander Racha 1 , Gitesh Kulkarni 1 , Anirban Dutta Choudhury 1 , Sundeep Khandelwal 1 , Avik Ghose 1 , Arpan Pal 1 1 TATA Consultancy Services, India, 2 Indraprastha Institute of Information Technology, Delhi, India e-mail 1 : (arijit.ukil, ishan.sahu, sai.racha, gitesh.k, anirban.duttachoudhury, sundeep.khandelwal, avik.ghose, arpan.pal)@tcs.com; email 2 : angshul@iiitd.ac.in 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) Oct 31 - Nov 4, 2021. Virtual Conference 978-1-7281-1178-0/21/$31.00 ©2021 IEEE 886