SEU-AM-ECG is a database towards ambiguous ECGs created by Wearable SHE Lab in Southeast University. The ECG episodes with ambiguous heartbeats include real-world and simulated records. The ambiguous data contains data considered too noisy to be identified as ECGs and ECGs that are misidentified as noise interference due to artifacts that cause partial waveforms to be unstably identified. All the records are chosen from the long-term out-of-hospital ECG signals of 200 patients aged 18 to 80 years with a history of arrhythmia disease using a wearable ECG monitoring device. The simulated ambiguous beats are generated on the basis of the training set from CPSC2019. The aim of constructing this database is to develop the algorithms suited to wearable ECG analysis.

This database was created aiming at ambiguous ECGs and QRS-like artifacts in the context of the explosion of wearable ECG devices. The majority of wearable ECGs have high levels of artifactual noise interference, resulting in QRS-like artifacts and ambiguous episodes. So far in Holter ECG analysis, the artifacts are determined referring to multi-lead information and heartbeat clusters, yet these methods cannot be migrated directly to wearable ECGs, particularly single-lead ECGs. In most situations, the morphology of QRS-like artifacts tends to be confusing. In practice, over detection or ignorance of such ambiguous data may cause incorrect identification of abnormal heartbeats (ventricular premature and atrial premature) or stray RR intervals, which may be adverse to the subsequent tasks. Importantly, pathological information may be concealed within ambiguous ECGs or QRS-complexes, such as QRS-complex variation and recurring “artifacts”. A feasible solution is to mark the ambiguous ECGs and QRS-complexes so that physicians can clearly attribute the anomaly in the RR interval sequence to specific or ambiguous heartbeats and evaluate their clinical significance based on variations and trends in RR intervals. Different from the signal quality database in the traditional ECG monitoring scenario, the normality of wearable ECGs consists of noisy and ambiguous episodes. Simple grading for the wearable ECGs according to the existing signal-quality-assessment (SQA) criteria is likely to cause the ignorance of valuable information, such as the ECGs collected during physical exercise. The purpose of constructing SEU-AM-ECG database is to solve the problem of how to capture successive recurring morphology patterns with sufficient consideration of ambiguous episodes.

To build ambiguous ECG database, we recorded the 24-hour out-of-hospital ECG signals of 200 patients with a history of arrhythmia disease using a wearable ECG monitoring device. The recorded ECGs are preliminary screened for the candidate ambiguous beats by the automatic analysis and annotation software supported by NaLong Health Technology Co., LTD. Two physiologists then label the candidate ambiguous beats as artifacts or QRS-complex. In the end, the heartbeats that the two physiologists provide inconsistent labels are annotated as ambiguous ones. Since it is expensive to prepare the real-world ECG episodes with ambiguous heartbeats, we provide a simulation approach as the supplement. The simulation of ambiguous beats is to insert a complete QRS-complex selected from another random ECG episode. This approach will ensure a portion of ECG episodes containing artifacts, yet they may be incorporated into the selected ECG episode as a subordinate QRS-complex when the inserted artifact is consistent with the contextual morphology pattern. The automatically screened noise episodes are manually double-checked, and the manually reannotated ECG episodes are accepted as ambiguous ECGs into the SEU-AM-ECG database. We also include some invalid noises containing sporatic QRS-complexes into the database as ambiguous episodes.

All of the records in SEU-AM-ECG database have been standardized as 10-second episodes sampled at 500 Hz. There are four subcatalogs including ambiguous ECG containing 523 episodes, ambiguous noise containing 2154 episodes, real-world ambiguous QRS-complex containing 1083 episodes and simulated ambiguous QRS-complex containing 2001 episodes. Also we provide the original training set in CPSC2019 for reference. All the records are wrapped into wfdb format and attached with the beat annotation according to the standard annotation codes from https://archive.physionet.org/physiobank/annotations.shtml#aux. To avoid misunderstanding, it should be notified that all the certain QRS-complex are annotated with Code “N” and the ambiguous ones are annotated with the Code “|”, which is described as “isolated QRS-like artifact” in the standard.

The database is intended for the development of ECG analysis method in application scenarios of wearable ECG monitoring, especially for long-term single-lead ECGs. The simulation approach for QRS-like artifacts is provided as sim_artifact.py.