Time Series Pattern Matching Technique for Low-Energy Arrhythmia Classification in Implantable Cardiac Devices

Implantable cardiac devices (ICDs) are often used as an effective treatment for arrhythmia. Although these devices have access to a live Electrocardiogram (ECG) stream, currently they do not offer on-device classification of arrhythmia due to the limited computing capability and severe power constraints. In this paper we propose a low-energy computing method for extracting shape-based features from ECG in combination with machine learning techniques for classifying nine different cases of arrhythmia. This is achieved by using a Localized Longest Common Subsequence (LLCS) algorithm which has low computational requirements that allows on-device execution. The proposed method strongly focuses on maintaining minimal energy and computational footprint, in line with the operating constraints of implantable devices. To demonstrate the energy efficiency and low computation load of the proposed method we implement the classification pipeline on a low-power RISC microcontroller and compare its performance with existing classification techniques. The classification accuracy and energy of the proposed method is compared with state-of-the art research in arrhythmia classification.