HeartCycle: A comprehensive dataset of synchronized impedance cardiography and echocardiography for accurate hemodynamic predictions

Abstract

The "HeartCycle" dataset offers a comprehensive collection of synchronized impedance cardiography (ICG) and echocardiography (ECHO) signals, supplemented with finger photoplethysmography (PPG), heart sounds, and electrocardiography (ECG) data from 17 healthy volunteers. Collected during the HeartCycle project (FP7-216695), this dataset aims to address biases in the ICG waveform, particularly the ABEXYOZ complex, where the B and X points do not precisely align with the aortic valve opening and closing notches. The biases in B and X point detection are critical for hemodynamic prediction because these characteristic points are used to calculate essential diagnostic parameters, including systolic time intervals (PEP and LVET), contractility, stroke volume, and cardiac output. By providing synchronized ICG and ECHO signals, researchers can better understand these biases and develop more accurate models for hemodynamic parameter computation. The dataset is stored in HDF5 format, facilitating the storage of complex data structures and easy access to various physiological parameters. It is ideal for developing machine learning models to enhance the detection of characteristic points in ICG signals. For instance, machine learning models can be used to detect characteristic points for improved left ventricular ejection time (LVET) estimation or mapping the ICG signal to the different mechanical events in the cardiac cycle using the ECHO as a reference. Detailed metadata and usage notes are included to support data utilization across different software environments. Ethical approval was obtained from the University of Coimbra Hospital's ethics committee, and informed consent was provided by all participants.

Background

Impedance cardiography (ICG) is one of the reference methods for portable devices in assessing several key hemodynamic descriptors, such as the systolic time intervals (STI) and cardiac output (CO) [1]. The ICG principle is based on the measurement of the thorax impedance variations (dZ/dt) that are influenced by airflow through the lungs, blood flow from the left ventricle to the aorta, and lung perfusion [2]. The assessment of the systolic time intervals requires the determination of the ICG’s characteristic points, which are assumed to be correlated to the opening and closing of the aortic valve [3]. The waveform obtained from the dZ/dt signal presents the ABEXYOZ complex, where B corresponds to the aortic valve opening notch and X to the aortic valve closing notch [4]. However, previous studies conclude there is a bias in the ICG waveform, and B and X points do not exactly align with the notches [5]. While previous datasets, such as the ReBeatICG database [6], have typically provided ICG measurements synchronized with ECG, the simultaneous acquisition of multiple modalities remains unexplored in open access resources. For this reason, this dataset provides researchers with ICG signals synchronized with echocardiography recordings (ECHO) to understand the bias present in the ICG waveforms, and it proposes new models and methods to correct this bias. To the best of our knowledge, this is the first publicly available dataset offering simultaneous ICG, ECHO, ECG, and PPG recordings, enabling comprehensive multimodal analysis of cardiac hemodynamics and validation of ICG-derived parameters against the gold-standard ECHO measurements.

Methods

The data were extracted from physiological studies conducted during the HeartCycle project on healthy subjects. This dataset contains data from 17 volunteers.

The HDF5 data files record the synchronized signals for impedance cardiography (ICG), finger photoplethysmography (PPG), heart sounds, and echocardiography (ECHO). For each one of these modalities, the synchronized signal for electrocardiography (ECG) is also provided. In addition, data files containing the hemodynamic and physiological parameters computed for each record were included. The MATLAB Software [7] was used to process signals and to generate the synchronized HDF5 files. A detailed description of the content of the HDF5 files is provided in the FileMetadata.csv files.

The ICG and ECG signals were recorded using Niccomo^® (TotalMedicalSolutions, Netherlands). Data were exported in .txt format. Vivid Ultrasound from General Electric was used to record ECHO data, and the data were processed using DICOM software, which created images in M-mode and Doppler mode. The ECHO output is stored as an image in the _091 group as a three-dimensional array, where the first dimension corresponds to the channel, the second to time, and the third represents depth or distance from the transducer in M-mode and velocity in Doppler mode. For PPG, sensors from Philips^® V26 Patient Monitor were used to collect the signals. Last, a Meditron stethoscope was used to annotate heart sounds.

Sampling rates depend on the device and the synchronization procedure. For Niccomo, sampling rate is equal to 200 for ECG and ICG signals; for Vivid Ultrasound the sampling rate is 136 for ECHO signals and the synchronized ECG; for the V26 Patient Monitor the sampling rate is 500 for the ECG and PPG signals; and a sampling rate equal to 44,100 Hz was used for phonocardiography and synchronized ECG. All sampling rates are documented in the Rate group inside each one of the groups in the HDF5 files.

The synchronization protocol for handling and processing data in the HeartCycle dataset includes acquisition, organization, and annotation of physiological signals (ECG, ICG, HS, PPG and ECHO). The acquisition process involves recording data using each hardware-specific software, generating files that are then copied to a designated directory structure based on acquisition date and volunteer ID. These raw acquisition files are processed to produce multiple CSV files, which are later imported into MATLAB for further processing. Each acquisition is assumed to generate different signal segments for each modality, corresponding to a different record, as outlined in the acquisition protocol.

Once imported, these signals are organized into MATLAB files named after each volunteer. These files contain three primary structures: aq_info (acquisition details like date and location), vol_info (volunteer demographics and health status), and measure (a matrix organizing ECG, ICG, HS, PPG and ECHO with different hemodynamic parameters collected). Each cell in the measure matrix includes time vectors, signal data, labels, sampling rates, units, run identifiers, and descriptions of the volunteer’s activity during that run. Manual annotation of PPG signals was required, based on visual inspection and protocol-defined intervals, to ensure accurate interpretation and segmentation of physiological responses during each activity.

Data Description

The dataset comprises 2.3 GB of recordings from healthy subjects. The files are systematically named to reflect the subject ID, the date (randomized), and the record ID. For instance, the file CH07_59146237_s0000029.h5 (Table 1) corresponds to the record s0000029 from the subject CH07 and performed on the day 59146237.

Specifically, there are a total of 208 records stored in HDF5 files and distributed in three experiments. There are 32 records in the experiment folder 59146237, 84 records for the experiment folder 59146238, and 92 records for the experiment folder 59146239. The subject distribution is presented in Table 2.

The HDF5 format allows storing complex data structures such as the one presented in this dataset. The structure of this file is summarized in Table 3. Each column represents one of the medical devices used, and each cell stores a vector or matrix with the corresponding data. For instance, ICG data can be accessed at C[4,2] in the HDF5 array – index can vary in function of the programming language. Please note, PPG data are only present in experiment 59146237, as the PPG sensor was incorporated into the acquisition protocol at a later stage of the study. Sampling rates were device-specific: Niccomo (ECG and impedance: 200 Hz), stethoscope (ECG and phonocardiography: 44100 Hz), echocardiogram (ECG and echocardiography: 136 Hz), and PPG (ECG: 500 Hz, plethysmography: 125 Hz).

Table 4 provides a more detailed mapping between .h5 group IDs and the physiological signals/devices, which is also documented in the README.md and GroupMapping.csv. Most of the signals are stored as a 2-dimensional array of shape (1, time), while AVO and AVC are a 1-dimensional array with the time coordinates of the event, and PEP and LVET include the time interval in milliseconds. However, the echo-related group _091 has the shape (3, time, distance/velocity), which differs from the other signals. This three-dimensional array represents three echocardiography signals in three device channels. The last dimension depends on the echocardiography mode, as some files include M-Mode and other files Doppler Mode. For a clearer interpretation, we recommend splitting the array into three matrices and computing the transposed matrix to have time on the X-axis.

Last, some physiological parameters are also recorded from Niccomo, as specified in Table 5.

The dataset is composed of three experiments, namely 59146237, 59146238, and 59146239, stored in directories with the same name. In each directory, there is a subdirectory called measure which contains the HDF5 files with the data, named as indicated before. Two additional files are in each experiment directory, the FileMetadata.csv and the SubjectMetadata.csv file.

Usage Notes

This dataset provides ICG recordings with echocardiography as reference, as well as other techniques, suitable for developing machine learning models to detect the real notches and improve the accuracy of hemodynamic parameter computation from ICG. To utilize the data, researchers can use different data science environments for reading HDF5 data, such as Jupyter, R Studio, or MATLAB, among others. In consequence, this dataset is not software dependent.

The traceability between subjects, files, and experiments is specified in the SubjectMetadata.csv file, where the demographic data of each subject are also summarized. In addition, data quality was included for each record file and specified as the synchronization percentage between two physiological signals or datasets (e.g., ICG and ECHO). It is defined as the proportion of temporally aligned data points or valid overlapping segments relative to the total expected duration of synchronization, expressed as:

Status = 100 - (Error * 10^4)

where Error represents the fraction of temporally misaligned or invalid data segments relative to the total recording duration.

While this dataset offers valuable multimodal synchronized recordings, researchers should note certain limitations. The relatively small sample size may limit generalization across diverse populations, and the controlled laboratory acquisition conditions may not fully represent real-world clinical or ambulatory settings. For this reason, we encourage researchers to use this dataset from a data science perspective for training new AI models, but we recommend avoiding the extraction of physiological conclusions that cannot be extrapolated to other populations.

Further details about how to use and how to get started with the dataset can be found in the README.md file. Furthermore, the script tutorial.py includes some examples on how to load HDF5 data.

import h5py
import numpy as np
import matplotlib.pyplot as plt

f = h5py.File('./59146237/measure/CH07_59146237_s0000029.h5', 'r')

print(f['measure']['value'].keys())

ecg = f['measure']['value']['_030']['value']['data']['value'][0,:]
time = f['measure']['value']['_030']['value']['time']['value'][0,:]

plt.figure(figsize=(12, 5))
plt.plot(time, ecg)
plt.title("ECG signal")
plt.xlabel('Time (ms)')
plt.ylabel('ECG (mV)')
plt.show()

It is important to note that the Niccomo impedance signal stored in the HDF5 file is the raw signal. For most applications, the derivative dZ/dt is required. An easy way to compute this derivative in Python is as follows, where icg_time is the array with the timestamp and icg_record is the array with the raw ICG signal.

import h5py
import numpy as np
import matplotlib.pyplot as plt

f = h5py.File('./59146237/measure/CH07_59146237_s0000029.h5', 'r')

print(f['measure']['value'].keys())

icg = f['measure']['value']['_031']['value']['data']['value'][0,:]
time = f['measure']['value']['_031']['value']['time']['value'][0,:]

dt = np.mean(np.diff(time))
dz = np.gradient(icg, dt)

For echocardiography, a special preprocessing is required to load and visualize the image matrix.

import h5py
import numpy as np
import matplotlib.pyplot as plt

f = h5py.File('./59146237/measure/CH07_59146237_s0000029.h5', 'r')

echo = f['measure']['value']['_091']['value']['data']['value'][0,:,:].transpose()

plt.figure(figsize=(12, 5))
plt.imshow(echo, cmap='viridis', aspect='auto')
plt.title("Echocardiography Image")
plt.xlabel('Time (ms)')
plt.show()

Last, a more complex example is provided in pan_tompkins.py, where some classes and functions were implemented for identifying R peaks in ECG using the Pan–Tompkins algorithm. Please note that any other implementation can be used.

Ethics

The study was approved by the University of Coimbra Hospital's ethics committee under the reference CES-238 and fully complies with the Declaration of Helsinki.

Acknowledgements

This work was supported in part by the EU FP7 project HeartCycle (FP7-216695), funded by the European Commission.

Conflicts of Interest

The authors declare no conflict of interest.

References

1. Kubicek WG, Patterson RP, Witsoe DA. Impedance cardiography as a noninvasive method of monitoring cardiac function and other parameters of the cardiovascular system. Ann N Y Acad Sci. 1970;170(2):724-32.
2. Visser KR, Mook GA, Van der Wall E, Zijlstra WG. Theory of the determination of systolic time intervals by impedance cardiography. Biol Psychol. 1993;36(1-2):43-50.
3. Chan GS, Middleton PM, Celler BG, Wang L, Lovell NH. Automatic detection of left ventricular ejection time from a finger photoplethysmographic pulse oximetry waveform: comparison with Doppler aortic measurement. Physiol Meas. 2007;28(4):439.
4. Benouar S, Hafid A, Attari M, Kedir-Talha M, Seoane F. Systematic variability in ICG recordings results in ICG complex subtypes: steps towards the enhancement of ICG characterization. J Electr Bioimp. 2018;9(1):72.
5. Carvalho P, Paiva RP, Henriques J, Antunes M, Quintal I, Muehlsteff J. Robust characteristic points for ICG-definition and comparative analysis. In: Proceedings of the International Conference on Bio-inspired Systems and Signal Processing; 2011 Jan 26-29; Rome, Italy. Setúbal, Portugal: SCITEPRESS; 2011. p. 161-8.
6. Pale U, Meier D, Muller N, Arza A, Atienza D. ReBeatICG database. Zenodo; 2021. https://doi.org/10.5281/zenodo.4725433
7. The MathWorks, Inc. MATLAB version 9.13.0 R2022b [Internet]. Natick MA: The MathWorks, Inc.; 2022 [cited 2025 Sep 23]. Available from: https://www.mathworks.com