Long-term electrocardiogram and wrist-based photoplethysmogram recordings with annotated atrial fibrillation episodes

Rationale Atrial fibrillation (AF) has emerged as a cardiovascular epidemic affecting more than 33 million individuals worldwide and carrying a 5-fold increased risk of brain stroke and a 3-fold increased risk of heart failure (Hindricks et al. 2021). AF is a progressive disease, with primary paroxysmal episodes being self-terminating; therefore, the success of complication management highly depends on early arrhythmia detection, which often requires long-term AF monitoring (Keach et al. 2015). Unfortunately, existing devices for long-term AF monitoring are either expensive (implantable cardiac monitors) or inconvenient due to skin irritation (Holter monitors, electrocardiogram (ECG) patches). Thus, it is desirable to develop inexpensive technologies ensuring wearing comfort. Recently, biooptical photoplethysmography (PPG) signal has emerged as such technology with immense potential for convenient long-term AF monitoring (Pereira et al. 2020). However, due to the lack of guidelines for arrhythmia interpretation in PPG, simultaneous ECG recording is needed for verification of the episodes detected in PPG. The present dataset contains simultaneously acquired wrist-based PPG and reference ECG signals with annotated AF episodes, and thus, is particularly suitable for use in the development and testing of automatic PPG-based AF detectors. Subjects and data acquisition protocol The dataset contains long-term ECG and PPG signals from 39 patients with suspected AF monitored for 5 to 8 days (261 days in total). Detailed demographic (sex, age, height, weight) and clinical (diagnosed comorbidities, medications) characteristics of the patients are provided in the supplementary file subject_info.xlsx. The acquisition of the PPG and ECG signals was started at Vilnius University Hospital Santaros Klinikos (Vilnius, Lithuania) and continued for a week at the patient’s home. The PPG signal was acquired at a sampling frequency of 100 Hz using a green LED embedded in a wrist-worn device developed at the Biomedical Engineering Institute (Kaunas, Lithuania). The reference ECG signal was acquired at a sampling frequency of 500 Hz using the Bittium Faros™ 180 ECG device together with the Bittium OmegaSnap™ patch electrode (Oulu, Finland). Additionally, triaxial acceleration signals were acquired with both devices at sampling frequencies of 50 and 25 Hz using wrist-worn and reference ECG devices, respectively. The occurrence times of QRS-complexes in ECG signals were obtained using an open-source QRS detector (Moeyersons et al. 2020), and initial AF episodes were automatically detected using a low-complexity AF detector relying on rhythm irregularity information (Petrėnas et al. 2015). Then, the AF detector output was visually inspected and manually corrected by medical specialists experienced in arrhythmia diagnosis with the aim to find undetected and discard falsely detected episodes. The PPG and acceleration signals of the wrist-worn device were aligned to the time of the ECG device, as described in ECG–PPG synchronization below. The data acquisition protocol was in accordance with the ethical principles of the Declaration of Helsinki and was approved by Vilnius Region Biomedical Research Ethics Committee (No. 158200-18/7-1052-557). All patients gave written informed consent to participate. In this version of the database, AF annotations were further reviewed and manually refined. Subject identifiers were also reassigned and may differ from those used in previous versions. Users should use the subject identifiers and corresponding metadata provided with this version when linking recordings to subject information. The PPG signals were also re-synchronized to the ECG. ECG–PPG synchronization The ECG device (Bittium Faros 180) and the wrist-worn device had separate clocks. Their offsets ranged from seconds to several hours (e.g. time-zone or daylight-saving settings). In this version, all PPG and wrist-worn device accelerometer data are aligned to the ECG clock, and the start times (recording_starttime, recording_startday) refer to the ECG clock. Synchronization procedure First, lost data packets (filled with NaN) and clock drift within each wrist-worn device file were corrected using the device’s internal time stamps. Each PPG segment was then aligned to the ECG in two stages. A coarse alignment matched the movement recorded by the wrist and chest accelerometers. Where possible, a fine alignment followed, matching PPG pulses to ECG R-peaks (assuming a 0.3 s pulse delay) and, during walking, matching step impacts recorded by both accelerometers. Only time offsets confirmed by several consistent estimates were used, and slow drift between the two clocks was corrected within each segment. Because a 0.3 s pulse delay was assumed, the R-peak-to-PPG delay in these data is not an independent measurement. Each PPG segment has a sync_method value. beat means the timing was confirmed at beat level (about ±0.05 s). coarse means it was confirmed by movement only (about ±1 s). Segments were excluded if their timing could not be confirmed, if they lay outside the ECG recording or in a gap without ECG, or if no PPG was recorded. Segments partly outside the ECG were trimmed. Segment numbering may therefore differ from earlier versions. For some subjects no usable PPG is available, and only the ECG file is provided (subjects 024, 025, 034 and 037). The alignment was checked by: the stability of the R-peak-to-PPG delay across minutes; the matching of ectopic beats in the ECG and PPG; AF episodes appearing at the same times in both signals; agreement between the PPG-derived and ECG-derived heart rate per minute. Signal quality Wrist PPG is very sensitive to movement and to how the wrist-worn device sits on the wrist. Large parts of the recordings, especially during daytime activity, are dominated by motion artifacts: individual pulses cannot be seen and the PPG rhythm does not follow the ECG. This is normal for wrist PPG recorded in daily life and is not a synchronization error. The time alignment applies to the whole segment. For beat segments it was confirmed in the clean parts of the segment, so beat does not mean that every pulse is visible. Rhythmic movements can also produce regular PPG waves that look like pulses but are not (their rate differs from the ECG heart rate). The accelerometer signals can be used to identify such periods. The ECG also contains periods of noise or lost electrode contact, where no R-peaks were detected. Technical details The acquired signals are provided in MAT-files named XXX_YYY.mat, where XXX is the patient ID and YYY is ECG for the signals from the Bittium Faros™ 180 ECG device and PPG for the signals from the wrist-worn device. For each patient, there is one ECG file and, if usable PPG is available, one PPG file containing one or more PPG segments. The PPG acquisition could be interrupted for a short time due to technical reasons, or for a longer time to allow battery charging of the wrist-worn device. If the ECG was recorded in two parts, they are joined in one file, with the interval between them filled with NaN. In addition to PPG, ECG, and acceleration signals, each file contains a signal header, the day when the recording started with respect to the first monitoring day of the patient, and the time of day when the recording started (on the ECG clock). The ECG files also contain QRS time indices and calculated RR intervals, together with AF annotations on a beat-to-beat basis. In the PPG files, the signals, recording_starttime, recording_startday and sync_method are cell arrays with one cell per PPG segment, in the same order. In the subject_info.xlsx file, physical inactivity is defined as < 5000 steps/day or < 150 min/week of moderate-intensity exercise or < 75 min/week of high-intensity exercise, excessive physical activity is defined as > 750 min/week of moderate-intensity exercise, and hypertension is classified into stages based on systolic/diastolic blood pressure: stage I corresponding to 140/90–159/99 mmHg, stage II to 160/100–179/109 mmHg, and stage III to ≥ 180/100 mmHg. Limitations Users should be aware that the timing accuracy of the PPG differs between segments (about ±0.05 s for beat and about ±1 s for coarse segments; see sync_method), and that the R-peak-to-PPG delay reflects the assumed 0.3 s pulse delay rather than an independent measurement. Large parts of the PPG are affected by motion artifacts (see Signal quality). The sampling frequencies of the devices can also vary slightly.