DeepQ Arrhythmia Database: A Large-Scale Dataset for Arrhythmia Detector Evaluation

Abstract

DeepQ Arrhythmia Database, the first generally available large-scale dataset for arrhythmia detector evaluation, contains 897 annotated single-lead ECG recordings from 299 unique patients. DeepQ includes beat-by-beat, rhythm episodes, and heartbeats fiducial points annotations. Each patient was engaged in a sequence of lying down, sitting, and walking activities during the ECG measurement and contributed three five-minute records to the database. Annotations were manually labeled by a group of certified cardiographic technicians and audited by a cardiologist at Taipei Veteran General Hospital, Taiwan. The aim of this database is in three folds. First, from the scale perspective, we build this database to be the largest representative reference set with greater number of unique patients and more variety of arrhythmic heartbeats. Second, from the diversity perspective, our database contains fully annotated ECG measures from three different activity modes and facilitates the arrhythmia classifier training for wearable ECG patches and AAMI assessment. Thirdly, from the quality point of view, it serves as a complement to the MIT-BIH Arrhythmia Database in the development and evaluation of the arrhythmia detector. The addition of this dataset can help facilitate the exhaustive studies using machine learning models and deep neural networks, and address the inter-patient variability. Further, we describe the development and annotation procedure of this database, as well as our on-going enhancement. We plan to make DeepQ database publicly available to advance medical research in developing outpatient, mobile arrhythmia detectors.