Method of estimating blood pressure of a subject
The present invention relates to a method and system for estimating blood pressure of a subject. In particular, but not exclusively, the method involves receiving a photoplethysmogram (PPG) signal from a light-based Pulse-Plethysmography sensor applied to the skin of a subject and reconstructing a pulse blood pressure waveform between systolic and diastolic blood pressure of the subject. Additionally, but not exclusively, the method involves processing the pulse blood pressure waveform and reconstructing an absolute blood pressure waveform of the subject.
1. A method of estimating blood pressure of a subject, the method including:
receiving a photoplethysmogram (PPG) signal from a light-based Pulse-Plethysmography sensor applied to the skin of the subject;
processing the PPG signal using a Wavelet transform algorithm to derive PPG wavelet coefficients in a plurality of frequency bands;
estimating blood pressure coefficients by processing the PPG wavelet coefficients using a machine learning algorithm that has been trained on training data including PPG wavelet coefficients derived from PPGs from light-based Pulse-Plethysmography sensors applied to the skin of test subjects correlated with Invasive Arterial Blood Pressure (ABP) wavelet coefficients derived from simultaneously obtained ABP measurements of systolic and diastolic blood pressure of the test subjects; and
reconstructing a pulse blood pressure waveform between systolic and diastolic blood pressure of the subject from the blood pressure coefficients.
2. A method according to claim 1 , further including processing selected ones of the PPG wavelet coefficients for processing using the machine learning algorithm based on an energy level present at each of the frequency bands exceeding a threshold level.
3. A method according to claim 2 , further including providing the selected ones of the PPG wavelet coefficients in a 2-Dimensional matrix of features to the machine learning algorithm.
4. A method according to claim 3 , wherein each column of the 2-Dimensional matrix is autoregressive incorporating near past and near future samples of the selected ones of the PPG wavelet coefficients.
5. A method according to claim 1 , wherein the machine learning algorithm is a Recurrent Neural Network (RNN) algorithm.
6. A method according to claim 5 , wherein the RNN algorithm is a Long short-term memory (LSTM) model.
7. A method according to claim 6 , further including processing the PPG wavelet coefficients using multiple LSTM models.
8. A method according to claim 7 , further including reconstructing output of the multiple LSTM models to form the pulse blood pressure waveform using an inverse Wavelet transform algorithm.
9. A method according to claim 8 , wherein the Wavelet transform algorithm is a Maximally Overlapped Discrete Wavelet Transform (MODWT) algorithm, and the method further includes reconstructing the pulse blood pressure waveform using an inverse MODWT algorithm.
10. A method according to claim 1 , further including pre-processing the PPG signal before processing using the machine learning algorithm by low-pass filtering the PPG signal with a cut-off frequency for the frequency bands.
11. A method according to claim 1 , further including:
processing the pulse blood pressure waveform using a further Wavelet transform algorithm to derive pulse blood pressure wavelet coefficients in a plurality of frequency bands;
extracting features from the pulse blood pressure wavelet coefficients in the plurality of frequency bands;
estimating mean arterial blood pressure (MAP), systolic blood pressure (SBP) or diastolic blood pressure (DBP) coefficients by processing the features using a further machine learning algorithm that has been trained on the training data; and
reconstructing the MAP, SBP and or DBP waveforms from the MAP, SBP and or and/or DBP coefficients, respectively.
12. A method according to claim 11 , further including combining the pulse blood pressure waveform and one or more of the MAP, SBP or DBP waveforms to reconstruct an absolute blood pressure waveform of the subject.
13. A method according to claim 11 , wherein the further machine learning algorithm is a Convolutional Long short-term memory (ConvLSTM) network and the features are a 2-Dimensional feature matrix including temporal and spatial features of the pulse blood pressure wavelet coefficients.
14. A method according to claim 13 , further including estimating intermediate MAP, SBP or DBP coefficients using the ConvLSTM network.
15. A method according to claim 14 , further including:
selecting features of the PPG signal and the pulse blood pressure waveform; and
estimating further intermediate MAP, SBP or DBP coefficients by processing the features of the PPG signal and the pulse blood pressure waveform using deep neural networks, respectively, that have been trained on the training data.
16. A method according to claim 15 , further including: concatenating the intermediate and further intermediate MAP, SBP or DBP coefficients to form a concatenated output vector; and passing the output vector through multiple layers of hidden neurons in an output network to generate the estimated MAP, SBP or DBP coefficients.
17. A method according to claim 1 , wherein the light-based Pulse-Plethysmography sensor includes a light source having an emission wavelength and a photodiode having a detection wavelength, wherein the emission wavelength and the detection wavelength are around an isosbestic wavelength where oxygenated and deoxygenated blood absorbs the same amount of light.
18. A method according to claim 17 , wherein the light source is an Infrared (IR) light source.
19. A method of monitoring blood pressure of a subject, the method including: applying a light-based Plethysmography sensor to the skin of the subject for a period of time; and estimating the blood pressure of the subject according to the method of claim 1 at designated intervals over the period time.
20. A system for estimating blood pressure of a subject, the system including:
a light-based Pulse-Plethysmography sensor configured to be applied to the skin of the subject to generate a photoplethysmogram (PPG) signal;
a processor in data communication with the sensor;
a memory; and
software resident in the memory and accessible to the processor, the software including a series of instructions executable by the processor to configure the processor to:
process the PPG signal using a Wavelet transform algorithm to derive PPG wavelet coefficients in a plurality of frequency bands;
estimate blood pressure coefficients by processing the PPG wavelet coefficients using a machine learning algorithm that has been trained on training data including PPG wavelet coefficients derived from PPGs from light-based Pulse-Plethysmography sensors applied to the skin of test subjects correlated with Invasive Arterial Blood Pressure (ABP) wavelet coefficients derived from simultaneously obtained ABP measurements of systolic and diastolic blood pressure of the test subjects;
reconstruct a pulse blood pressure waveform between systolic and diastolic blood pressure of the subject from the blood pressure coefficients; and
output the pulse blood pressure waveform to a display.