Abstract
Heart arrhythmias, resulting from disturbances in the rate, regularity, or conduction of cardiac electrical pulses, are often sporadic and difficult to diagnose in clinical settings. While wearable electrocardiogram (ECG) devices offer a promising solution for long-term monitoring, the lack of high-quality, annotated datasets reflecting the noise profiles of wearable hardware has hindered algorithm development. This paper presents a new open-access database containing 2,000 30-second recordings from over 200 patients, featuring manual annotations by cardiologists and a dedicated signal quality assessment database.
The Challenge of Wearable ECG Monitoring
Traditional "gold standard" datasets, such as the MIT-BIH Arrhythmia Database, were primarily recorded using clinical-grade equipment in controlled environments. However, wearable devices typically utilize dry electrodes and are subject to intense motion artifacts and environmental interference. Algorithms trained on clean clinical data often fail when deployed on the "messy" signals generated by mobile hardware. There is a critical need for data that captures these unique signal challenges to improve the reliability of intelligent diagnosis.
System Architecture and Methodology
To address this gap, researchers developed a comprehensive pipeline for data acquisition and processing, utilizing a custom cloud-based infrastructure.
Hardware and Infrastructure
The study employed an FDA-certified, bipolar limb two-lead wearable device recording at a sampling rate of 400 Hz with 12-bit resolution. Data is transmitted to a five-part cloud platform designed to handle information management, communication, storage, expert annotation, and automated computation.
Annotation Workflow
Ensuring clinical accuracy required a rigorous three-stage process:
- Automated Pass: Initial labels were generated using standard QRS detection (Pan-Tompkins) and a neural network classifier (CNN + LSTM).
- Expert Correction: Two clinical cardiologists independently reviewed and corrected the automated labels.
- Tie-breaking: A third senior expert resolved any disagreements between the primary reviewers.
The Databases
The researchers have released two distinct databases to support different aspects of ECG analysis, accessible at http://www.shelab.cn/Data.
Signal Quality Database
Recognizing that Signal Quality Assessment (SQA) is a prerequisite for accurate diagnosis in wearables, the authors provided 300 recordings (10s each) categorized into three levels:
- Level A (Good): Clear P-QRS-T waves.
- Level B (Medium): Interpretable but contains significant noise.
- Level C (Poor): Unacceptable for clinical diagnosis.
Arrhythmias Database
This core dataset contains 2,000 30-second recordings from 200+ patients (ages 18–82). It is subdivided into 18 categories across three major classes: Sinus Rhythms (e.g., tachycardia, bradycardia), Atrial Rhythms (e.g., AFib, PACs), and Ventricular Rhythms (e.g., PVCs, ventricular couplets).
Key Insights and Clinical Significance
A major contribution of this work is the emphasis on SQA as a "first-class citizen" in arrhythmia detection. By identifying "Level C" signals, algorithms can avoid generating false positives from unusable data. Furthermore, by providing 18 specific subcategories with detailed morphological descriptions, the authors aim to bridge the gap between clinical cardiology and the data science community.
Limitations and Future Work
While the database is a significant step forward, some rare arrhythmia subtypes remain underrepresented. Future efforts will focus on expanding the dataset's diversity and leveraging the cloud platform for the labor-intensive task of beat-by-beat annotation for long-term (24–72 hour) dynamic recordings.
Summary updated