Non-invasive method and system for estimating blood pressure from photoplethysmogram using statistical post-processing

A system for estimating blood pressure (BP) using a photoplethysmogram (PPG), the system comprising:

• a PPG sensor (102) configured to capture a PPG signal;
• a memory (104);
• a processor (106) in communication with the memory (104);
• a preprocessing module (108), coupled to the processor (106), configured to preprocess the captured PPG signal by performing band pass filtering, baseline correction, noise removal, and cycle selection based on similar pulse height and periodicity to take care of changes in PPG signal amplitude due to variable finger pressure and discontinuities due to motion artifacts, wherein the band pass filtering is performed in the range of 0.75-8 Hz;
• a feature selection module (110), coupled to the processor (106), configured to extract a plurality of features from the preprocessed PPG signal, wherein the plurality of features are extracted by evaluating linear correlation between the plurality of features and ground truth BP in a training dataset, wherein the ground truth BP is measured using a digital BP monitoring device immediately after capturing the PPG signal, and wherein the plurality of features extracted include Ts, Ts/Td, As/Ad, and 1/Ts2, and wherein Ts is systolic upstroke time, Td is diastolic time, As is systolic area, Ad is diastolic area of the PPG signal;
• a training model generation module (112), coupled to the processor (106), configured to create a training model for estimating the BP using the plurality of features and a ground truth value of the BP, wherein the training model is created using a gradient descent learning based linear regression;
• a regression module (114), coupled to the processor (106), configured to generate a BP distribution using the training model, wherein the BP distribution is part of a set of BP distributions generated from different subjects;
• and
• a post-processing module (116), coupled to the processor (106), configured to post process the set of BP distributions to:
• reject unacceptable BP distributions out of the set of BP distributions based on a predefined criteria, wherein the predefined criteria includes clustering the set of BP distributions into clusters of similar BP based on probability density and rejecting those BP distributions where there are i) multiple equally dominant clusters ii) two clusters that are too far apart or iii) unacceptable cluster formation, wherein the multiple equally dominant clusters are the clusters with a range of cluster populousness below a threshold, wherein the set of BP distributions is checked for the multiple equally dominant clusters when number of clusters in systolic BP or diastolic BP is greater than or equal to three, to identify if too many small clusters exist uniformly all over the set of BP distributions, wherein when the set of BP distributions comprises two clusters, then distribution width between the two clusters is checked to identify if the clusters are too far apart, and wherein the clusters of similar BP are determined based on best Bayesian Information Criterion (BIC);
• and
• provide a statistical average for the accepted BP distributions to estimate the BP.

One or more non-transitory computer readable information medium having embodied thereon a computer program for executing a method for estimating blood pressure (BP) using a photoplethysmogram (PPG), wherein the method comprises:

• capturing a PPG signal using a PPG sensor (102);
• preprocessing the PPG signal by performing band pass filtering, baseline correction, noise removal, and cycle selection based on similar pulse height and periodicity to take care of changes in PPG signal amplitude due to variable finger pressure and discontinuities due to motion artifacts, wherein the band pass filtering is performed in the range of 0.75-8 Hz;
• extracting a plurality of features from the preprocessed PPG signal, wherein the plurality of features are extracted by evaluating linear correlation between the plurality of features and ground truth BP in a training dataset, wherein the ground truth BP is measured using a digital BP monitoring device immediately after capturing the PPG signal, and wherein the plurality of features extracted include Ts, Ts/Td, As/Ad, and 1/Ts2, and wherein Ts is systolic upstroke time, Td is diastolic time, As is systolic area, Ad is diastolic area of the PPG signal;
• creating a training model for estimating the BP using the plurality of features of the PPG signal and a ground truth value of the BP, wherein the training model is created using a gradient descent learning based linear regression;
• generating a BP distribution using the training model, wherein the BP distribution is part of a set of BP distributions generated from different subjects;
• and
• post processing the set of BP distributions to estimate the BP, wherein the post processing comprises:
• rejecting unacceptable BP distributions out of the set of BP distributions based on a predefined criteria, wherein the predefined criteria includes clustering the set of BP distributions into clusters of similar BP based on probability density and rejecting those BP distributions where there are i) multiple equally dominant clusters ii) two clusters that are too far apart or iii) unacceptable cluster formation, wherein the multiple equally dominant clusters are the clusters with a range of cluster populousness below a threshold, wherein the set of BP distributions is checked for the multiple equally dominant clusters when number of clusters in systolic BP or diastolic BP is greater than or equal to three, to identify if too many small clusters exist uniformly all over the set of BP distributions, wherein when the set of BP distributions comprises two clusters, then distribution width between the two clusters is checked to identify if the clusters are too far apart, and wherein the clusters of similar BP are determined based on best Bayesian Information Criterion (BIC);
• and
• providing a statistical average for the accepted BP distributions to estimate the BP.