A new database with annotations of P waves in ECGs with various types of arrhythmias

Electrocardiography (ECG) is still the most available and widely used method for cardiovascular system examination (World Health Organisatio ). The ECG signal reflects the electrical activity of the heart and provides a significant amount of information about the heart function (Mendis et al 2011). Accurate detection of ECG components, such as the P wave, QRS complex and T wave, are fundamental steps in ECG analysis and subsequent cardiac pathological events detection. In practice, automated evaluation of ECG records using software is necessary (Wagner et al 2009). The detection of QRS complexes and T waves is usually efficient. However, methods for P-wave detection are not so successful in physiological signals and especially in pathological signals. This fact applies in both real practice and research (Portet et al 2008, Hossain et al 2019, Maršánová et al 2019a, Kalyakulina et al 2020).

One factor that prevents progress in this field is the lack of publicly available datasets with correct P-wave annotations suitable for training and testing detection algorithms (Leutheuser et al 2016, Maršánová et al 2019a).

Methods are usually tested either on part of a publicly available QT database (Laguna et al 1997, Goldberger et al 2000) or on a CSE database (Willems et al 1990) that is not publicly available, both with manual P-waves annotations. In addition to these, there are two new databases, namely the MIT-BIH Arrhythmia Database P-Wave Annotations (Goldberger et al 2000, Maršánová et al 2019a, 2019b) and the Lobachevsky University Electrocardiography Database (Goldberger et al 2000, Kalyakulina et al 2020), which is not used frequently yet. There are also two publicly available databases with P-waves annotations that contain mistakes: the P-waves annotations of the MIT-BIH Arrhythmia Database by Elgendi et al (Elgendi et al 2016). and the automatically annotated part of the QT database (Laguna et al 1997). Thus, these annotations cannot be recommended for use in the testing of P-wave detection algorithms.

The most commonly used databases, the QT database and the Common Standards for Electrocardiography (CSE) database, contain predominantly physiological ECG records or pathologies that do not affect P-wave detection. However, the content of the pathologies in databases is very important for the objective testing of P-wave detection algorithms. During pathological function of the heart, information about the positions of P waves is very important in determining the diagnosis. Unfortunately, current algorithms are not able to detect P waves in pathological signals reliably.

Therefore, we fill this gap and introduce a new database of ECG signals with manually annotated P waves. The Brno University of Technology ECG Signal Database with Annotations of P Wave (BUT PDB) consists of 50 two-minute two-lead ECG signals. The records contain pathologies during which it is usually difficult to detect P waves by automatic algorithms (Portet et al 2008, Hossain et al 2019, Maršánová et al 2019a, Kalyakulina et al 2020). The P-waves positions were manually annotated by two ECG experts with seven years of practical experience evaluating Holter ECG records in cardiovascular ambulance (Maršánová et al 2019a).

The ECG signals were selected by two ECG experts from three existing databases: the MIT-BIH Arrhythmia (MIT-A) Database (Goldberger et al 2000, Moody et al 2001), the MIT-BIH Supraventricular Arrhythmia (MIT-S) Database (Greenwald et al 1991, Goldberger et al 2000) and the Long Term Atrial Fibrillation (LT-AF) Database (Goldberger et al 2000, Petrutiu et al 2007). All three databases contain ECG signals and annotations of positions and types of QRS complexes. The MIT-A database additionally contains annotations of types of arrhythmias present in records. All of these databases contain 2 leads of ECG signals and do not contain significant noise. Detailed information about these databases is available on Physionet (Goldberger et al 2000) or in articles (Greenwald et al 1991, Petrutiu et al 2007, Moody et al 2001).

Two ECG experts went through the databases and selected records with diagnoses which affect the P wave. From the record they selected interesting two-minute sections with higher incidence of pathologies. The signals were chosen to represent as many types of pathologies present in records as possible due to the contents of the databases used. The heartbeats influenced by pathology are not present in these signals consistently, so the ECG signals also contain physiological heartbeats. The final database consists of 50 two-minute, two-lead ECG signals with various types of pathologies. The number and type of pathologies chosen represent a real sample of data from medical practice. From MIT-A database, 38 signals were selected. From MIT-S database, 5 signals were selected. From LT-AF database, 7 signals were selected. The process of signal selection signals together with information about the number of heartbeats and P waves present in database is ilustrated in figure 1.

Zoom In Zoom Out Reset image size

The P-waves positions were manually annotated by two ECG experts with seven years of practical experience evaluating Holter ECG records in cardiovascular ambulance (Smíšek et al 2017, Maršánová et al 2019a). The ECG data was not pre-processed. The first expert provided manual annotations of peaks of P waves, and the second manually checked them. Unclear parts of the records were discussed by both experts until a consensus was reached. Everything was conducted manually without the use of automated annotation software. To facilitate the work of the ECG experts, a free software tool—SignalPlant (Plesinger et al 2016)—was used for manual marking of peaks of P waves.

Each record also contains annotation of the diagnosis (pathology) and types of QRS complexes (from the original databases) (Goldberger et al 2000, Moody et al 2001). The information about pathologies present in records was checked by ECG experts—the original annotations were found correct—then taken from the original databases and supplemented by the experts in cases where information was missing (all signals from MIT-S). The information about types of QRS complexes was taken from the original databases.

Types of pathologies with their abbreviations used in the BUT PDB, number of cases (records), number of heartbeats with specific pathology and IDs of the records in which the pathology is present are listed in table 1. The column 'Number of heartbeats with specific pathology' also contains the number of P waves which are present extra - without QRS complex. This is in case of AVB II and AVB III. The column provides also information that numbers of SVTA is included in number of heartbeats A, as same as B, T, IVR, VP, VFL in V - this is due the fact that these pathologies have same origin and morphologic shape of QRS but are present multipletimes (e.g. in pair, trigeminy, tachycardia). BI is included in L, because BI was present in same beats as L (diagnose is left bundle branch block with atrioventricular block 1st degree).

The BUT PDB database contains annotation of P waves in ECG signal with 23 different types of pathologies. The database includes 7,638 QRS complexes, of which 2,120 are without P wave (e.g. atrial fibrillation, ventricular beats or during nodal rhythm) and 5518 following P wave (normal timing). In BUT BDP is 5,599 P waves of which 81 are present without QRS complex (e.g. AVB II).

On Physionet (Goldberger et al 2000, Maršánová et al 2021), where database is published, information about each record from BUT PDB is listed. The information contains: ID of the record, types of pathologies present in each record, types of heart beats, number of heart beats, number of P waves, name of original database, ID of the record from the original database, the start and the end (in samples) of the selected segment and the sampling frequency. The column 'Types of pathologies' contains information about pathological beats as same as pathological rhythm. The column 'Types of heartbeats' contains infrormations about all types of heartbeats present in record, these annotation are taken over from original database and supplemented by missing information by ECG experts (e.g. AFIB, VFL, PREX beats).

The examples of records with various pathologies (i.e. second atrioventricular block (rec. 01), ventricular premature contraction, ventricular tachycardia, atrial fibrillation (rec. 48) and newly annotated P waves are shown in figure 2.

Zoom In Zoom Out Reset image size

The BUT PDB is available on Physionet (Goldberger et al 2000, Maršánová et al 2021) (https://physionet.org/content/but-pdb/1.0.0/). All data is provided in the WaveForm Database (WFDB) format, which is supported by the WFDB Software Package (Goldberger et al 2000). The IDs of the recordings are numbers from 01 to 50. The ECG signals are stored in files with suffix *.dat; the annotations of P waves are stored in files with suffix *.pwave; the positions of QRS complexes, their types and the sampling frequency of each ECG signal are stored in files with suffix *.qrs. The exact types of pathologies present in each signal are described in the text file with the name README.txt.

The codes needed for loading the data are available on Physionet (Goldberger et al 2000) (https://archive.physionet.org/physiotools/wfdb.shtml). In the case of using Matlab, to load files *.dat with ECG records, use function 'rdsamp': e.g. [signal, fs, tm] = rdsamp ('17.dat'). To load files with annotations of P waves or QRS complexes and their types, use function 'rdann': e.g. [P] = rdann ('17', 'pwave') or [QRS, typeQRS] = rdann ('17', 'qrs').

BUT PDB was created for the development, testing and objective comparison of algorithms for P-waves detection. For objective comparison of algorithms, the prerequisite is that the researchers will use the entire database for testing their algorithms and that they will not select and/or shorten the signals. The BUT PDB includes a representative sample of pathologies. The BUT PDB will help develop new, more accurate and robust methods for processing and analysing ECG records in the sense of P wave detection. These algorithms will be usable for implementation to the software for ECG signals analysis in real medical practice and thus will help cardiologists to evaluate ECG records.

The authors wish to thank LCDR Joshua Swift from Office of Naval Research (ONR) Code 342 and Dr Stephen O'Regan from ONR Global Central and Eastern European Office for their support. The funders had no role in the design of the study; in the collection, analyses or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

There are no conflicts of interest.

This work has been funded by the United States Office of Naval Research (ONR) Global, award number N62909-19-1-2006. The authors wish to thank LCDR Joshua Swift from ONR Code 342 and Dr Martina Siwek from ONR Global Central and Eastern European Office for their support.