methods using Photoplethysmography (PPG) signals, which can achieve a simple and
accurate blood pressure measurement. Teng et al. [2] established a linear regression
model using four extracted features of PPG signals for BP estimation: a width of 2/3
pulse amplitude, a width of 1/2 pulse amplitude, systolic upstroke time, and diastolic
time. Kurylyak [3] has done a lot more in the characteristics of the interval, a total of 21
parameters as the input of the BP network, in order to predict blood pressure. Gao et al.
[4] developed a method for BP estimation using the regression support vector machine
(RSVM) method, with RBF kernel and discrete wavelet transform, and obtained better
performance. Wang et al. [5] used MTM method for feature extraction, and an artificial
neural network (ANN) is used for BP estimation. Wu [6] used SWT to decompose the
pulse wave and extracted ten characteristic parameters as the input of ANN network.
Wen et al. [7] used the sum of two Gaussian functions as the model to fit the pulse
wave envelope, and the parameters of the Gaussian function after fitting were the input,
and the determination of systolic and diastolic pressure was performed through two
feed-forward neural networks. The methods mentioned above all need to perform
feature extraction on the original pulse wave data. The operation of the feature
extraction is burdensome and inevitably added in human subjective consciousness, so
that the nonlinear relation between the pulse wave and the blood pressure cannot be
extracted very well.
This paper presents an end-to-end method to obtain blood pressure. Using a single
complete pulse wave as the input of the neural network, the estimated blood pressure
values can be obtained. In this study, we extract the signal from the MIMIC database [8]
for network training and testing. Finally, the average absolute error on 15 datasets is
3.95 ± 4.38 mmHg for systolic blood pressure (SBP) and 2.14 ± 2.40 mmHg for
diastolic blood pressure (DBP) when using 6000 transfer learning data, fulfilling the
international standard of blood pressure measurement, and the work avoids the complicated feature extraction work, which shows the proposed method is simple, efficient,
and accurate. In summary, the main contribution of this paper is proposing an end-toend neural network for the measurement of blood pressure which proved to be effective.
In the rest of this paper, Sect. 2 introduces the MIMIC database, the details of the
method are recommended in Sects. 3 and 4 introduce the implementation and the
results, a discussion is given in Sect. 5 finally summarize the article.
2 Data Source
The MIMIC database contains data recorded on 90 ICU patients. The data for each case
includes the signals obtained from the bedside monitor and periodic measurements and
clinical data obtained from the patient’s medical records. The length of these records
varies, with almost all records being at least 20 h and many being 40 h or longer. In
short, the database contains nearly 200 days of real-time signals and related data.
The data for each case is kept in one folder, usually containing hundreds of separate
files, each containing a 10-minute recording signal, and then it can be assembled
without clearance and forming a continuous record. The types of information contained
in the records of different cases are slightly different, including roughly arterial blood
pressure (BP), heart rate (HR), PPG, pulmonary arterial pressure (PAP), etc. [9, 10].
An End-to-End Neural Network Model for Blood Pressure …
263
accurate blood pressure measurement. Teng et al. [2] established a linear regression
model using four extracted features of PPG signals for BP estimation: a width of 2/3
pulse amplitude, a width of 1/2 pulse amplitude, systolic upstroke time, and diastolic
time. Kurylyak [3] has done a lot more in the characteristics of the interval, a total of 21
parameters as the input of the BP network, in order to predict blood pressure. Gao et al.
[4] developed a method for BP estimation using the regression support vector machine
(RSVM) method, with RBF kernel and discrete wavelet transform, and obtained better
performance. Wang et al. [5] used MTM method for feature extraction, and an artificial
neural network (ANN) is used for BP estimation. Wu [6] used SWT to decompose the
pulse wave and extracted ten characteristic parameters as the input of ANN network.
Wen et al. [7] used the sum of two Gaussian functions as the model to fit the pulse
wave envelope, and the parameters of the Gaussian function after fitting were the input,
and the determination of systolic and diastolic pressure was performed through two
feed-forward neural networks. The methods mentioned above all need to perform
feature extraction on the original pulse wave data. The operation of the feature
extraction is burdensome and inevitably added in human subjective consciousness, so
that the nonlinear relation between the pulse wave and the blood pressure cannot be
extracted very well.
This paper presents an end-to-end method to obtain blood pressure. Using a single
complete pulse wave as the input of the neural network, the estimated blood pressure
values can be obtained. In this study, we extract the signal from the MIMIC database [8]
for network training and testing. Finally, the average absolute error on 15 datasets is
3.95 ± 4.38 mmHg for systolic blood pressure (SBP) and 2.14 ± 2.40 mmHg for
diastolic blood pressure (DBP) when using 6000 transfer learning data, fulfilling the
international standard of blood pressure measurement, and the work avoids the complicated feature extraction work, which shows the proposed method is simple, efficient,
and accurate. In summary, the main contribution of this paper is proposing an end-toend neural network for the measurement of blood pressure which proved to be effective.
In the rest of this paper, Sect. 2 introduces the MIMIC database, the details of the
method are recommended in Sects. 3 and 4 introduce the implementation and the
results, a discussion is given in Sect. 5 finally summarize the article.
2 Data Source
The MIMIC database contains data recorded on 90 ICU patients. The data for each case
includes the signals obtained from the bedside monitor and periodic measurements and
clinical data obtained from the patient’s medical records. The length of these records
varies, with almost all records being at least 20 h and many being 40 h or longer. In
short, the database contains nearly 200 days of real-time signals and related data.
The data for each case is kept in one folder, usually containing hundreds of separate
files, each containing a 10-minute recording signal, and then it can be assembled
without clearance and forming a continuous record. The types of information contained
in the records of different cases are slightly different, including roughly arterial blood
pressure (BP), heart rate (HR), PPG, pulmonary arterial pressure (PAP), etc. [9, 10].
An End-to-End Neural Network Model for Blood Pressure …
263