Scribd
Upload
56 views2 pages

Supplementary Materials

part of a github project about blood pressure classification

Uploaded by

Farah Omri
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
56 views2 pages

Supplementary Materials

part of a github project about blood pressure classification

Uploaded by

Farah Omri
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

Supplementary Material

1 DATASET GUIDELINES
PulseDB is available for download from the GitHub repository at https://github.com/pulselabteam/PulseDB.
The training, calibration, and testing subsets summarized in Table 4, as well as the supplementary
training, calibration, and testing subsets described in Section 4.3 and Table 7, can be reproduced from
the script provided in the GitHub repository. The supplementary subsets are also available from Kaggle
at https://doi.org/10.34740/KAGGLE/DS/2447469. Here, we provide information about how to use this
dataset.
1.1 File Structure
The PulseDB dataset is released as 5, 361 MATLAB® data files (MAT-File version 7.3), with each
file corresponding to all segments belonging to one subject. The files were put into two folders,
“PulseDB MIMIC” and “PulseDB Vital”, separating subjects in the MIMIC-III matched subset from
subjects in the VitalDB dataset.
Each of these files is an 1-D array of MATLAB structures. Each structure corresponds to a 10-s segment,
which includes the following fields:
• SubjectID: It takes the format of “pXXXXXX”, in which the 6 digits after “p” is the original subject
ID used by the MIMIC-III matched subset and the VitalDB dataset. Segments in the same data file
share the same SubjectID.
• CaseID: Identifier of record. Segments in the same data file with different CaseID come from different
records belonging to the same subject.
• SegmentID: Sequence of segments. For segments with same SubjectID and CaseID, a segment with
smaller SegmentID occurs prior to another segment with larger SegmentID temporally.
• ECG Raw, PPG Raw, ABP Raw: Raw, unfiltered physiological signals directly cropped from
records in the MIMIC-III matched subset and the VitalDB database. In each segment, the amplitude of
ECG Raw and PPG Raw signals were linearly remapped between 0 and 1, while the amplitude of
ABP Raw signal was not modified, such that the absolute SBP and DBP values were preserved. These
raw signals can be filtered with user-defined settings to be used as inputs or outputs that fit best to the
desired BP estimation method.
• ECG F, PPG F, ABP F: Filtered ECG, PPG and ABP signals. The ECG signal was band-pass filtered
at [0.5, 40] Hz with a Butterworth filter, while the PPG and ABP signals were band-pass filtered at
[0.5, 8] Hz using a Chebyshev-II filter, as described in Section 2.4. We suggest using the ECG F and
the PPG F signals as inputs of BP estimation models, if there is no specific requirement for processing
the input signals of the model.
• ABP Lag, PPG ABP Corr: ABP Lag is the lag that yields the highest cross correlation between
ABP F and PPG F. The maximum possible lag is limited to ± 125 samples, or ±1 s. The value of
Pearson’s correlation coefficient between ABP F and PPG F aligned using this lag is PPG ABP Corr.
• ECG RPeaks, PPG SPeaks, PPG Turns, ABP SPeaks, ABP Turns: Positions of extracted
characteristic points, specified as indices. For example, indexing ECG Raw using ECG RPeaks
identify the amplitudes of all R-peaks of the ECG signal in this segment.

1
Supplementary Material

• SegSBP, SegDBP: the segment-averaged SBP and DBP values. We suggest using these values as
reference SBP and DBP labels for training sequence-to-label BP estimation models (e.g. CNN), while
using ABP Raw for training sequence-to-sequence models (e.g. LSTM).
• Age, Gender, Height, Weight, BMI: Demographic information of the subject from which the segment
was retrieved. The Height, Weight and BMI fields are only available for segments derived from the
VitalDB database, since the MIMIC-III matched subset does not record these information.
1.2 Proxy Files for Statistical Analysis
To enhance data loading efficiency and to reduce memory cost for analyzing a large dataset, proxy files
are generated for PulseDB to include only the demographic information and the reference SBP and DBP
labels in each segment for basic statistical analysis, without requiring the effort of loading the waveform
dataset.
Three types of proxy files are generated and included as part of the dataset. The first type, named
as “PulseDB Info”, is a summary of demographic information and reference BP values for all
subjects in the PulseDB dataset. The second type, named as “Train Info”, “CalBased Test Info”,
“CalFree Test Info”,“AAMI Test Info” and “AAMI Cal Info”, corresponds to each of the training,
calibration, and testing subsets of segments summarized in Table 4. Since these proxy files may
record information of subjects from both MIMIC and VitalDB datasets, the field Subj Name
is used to identify different subjects, which consists of an additional digit of 0 or 1 after
the SubjectID in the data files, to distinguish subjects in the MIMIC-III matched subset from
subjects in the VitalDB dataset. Moreover, “VitalDB Train Info”, “VitalDB CalBased Test Info”,
“VitalDB CalFree Test Info”,“VitalDB AAMI Test Info” and “VitalDB AAMI Cal Info” correspond
to the supplementary subsets generated from only VitalDB subjects, described in Section 4.3 and Table 7.
1.3 Generation of Python-Friendly Training, Calibration and Testing Subsets
A MATLAB function is provided to generate the training, calibration, and testing subsets summarized in
Tables 4, 7 and Section 4.3, using the subset separation settings defined in the proxy files. To maximize
training and testing efficiency, the function fetches required data from the MATLAB data files that
store the signal segments, and concatenates them into multi-dimensional arrays. These arrays can be
effectively loaded into Python environments for training and evaluating deep learning models. The generated
supplementary training, calibration, and testing subsets are available from Kaggle.

You might also like