Machine learning techniques for estimating carotid-femoral pulse wave velocity and/or vascular age from single-site arterial waveform measurements
Machine learning techniques for estimating an indication of carotid-femoral pulse wave velocity (CFPWV) and/or vascular age (VA) of a subject from an arterial waveform previously measured at a single site by an arterial waveform acquisition device coupled to the subject, the method comprising using at least one computer hardware processor to perform: obtaining the arterial waveform, the arterial waveform comprising a plurality of portions corresponding to a respective plurality of heartbeats of the subject; generating a representative heartbeat waveform from the arterial waveform; and providing the representative heartbeat waveform as input to a trained machine learning model to produce a corresponding output indicative of the CFPWV of the subject.
1 . A method for using a machine learning model to estimate an indication of carotid-femoral pulse wave velocity (CFPWV) and/or vascular age (VA) of a subject from an arterial waveform previously measured at a single site by an arterial waveform acquisition device coupled to the subject, the method comprising:
using at least one computer hardware processor to perform:
obtaining the arterial waveform, the arterial waveform comprising a plurality of portions corresponding to a respective plurality of heartbeats of the subject, the plurality of portions comprising at least 10 portions, each of the plurality of portions comprising at least 250 data points;
generating a representative heartbeat waveform from the arterial waveform, the generating comprising:
detecting the plurality of portions in the arterial waveform;
individually detrending each of the detected plurality of portions to obtain a plurality of detrended portions,
averaging portions in the plurality of detrended portions to obtain the representative heartbeat waveform; and
processing the representative heartbeat waveform and at least one derivative waveform of the representative heartbeat waveform using a trained one-dimensional (1D) convolutional neural network (CNN) model to produce a corresponding output indicative of the CFPWV of the subject,
wherein the 1D CNN comprises multiple 1D convolutional layers, multiple rectified linear unit layers, multiple pooling layers, and a fully-connected layer,
wherein the 1D CNN comprises at least one million parameters, and
wherein the processing comprises calculating the output indicative of the CFPWV from the representative heartbeat waveform the at least one derivative waveform and the at least one million parameters.
2 . The method of claim 1 , wherein the arterial waveform acquisition device comprises a tonometer, a plethysmographic (PG) device, an echocardiographic device, a digital blood pressure system, an electrical impedance sensor, and/or an ultrasound imaging device.
3 . The method of claim 1 , wherein the arterial waveform acquisition device comprises a wearable device.
4 . The method of claim 1 , wherein the arterial waveform comprises a pressure waveform or an arterial distension waveform.
5 . The method of claim 1 , wherein the arterial waveform comprises an arterial flow waveform.
6 . The method of claim 1 , wherein the plurality of portions comprises at least 100 portions corresponding to a respective plurality of heart beats.
7 . The method of claim 1 , wherein each portion of the plurality of portions comprises at least 500 data points.
8 . The method of claim 1 , further comprising:
removing one or more portions from the detected plurality of portions prior to performing the detrending.
9 . The method of claim 1 , wherein generating the heartbeat waveform from the arterial waveform comprises:
after detecting the plurality of portions in the arterial waveform, removing one or more portions from the detected plurality of portions to obtain a filtered plurality of portions; and
determining a period for the filtered plurality of portions,
wherein the individually detrending is performed on at least some of the filtered plurality of portions and based on the determined period to obtain the plurality of detrended portions.
10 . The method of claim 1 , further comprising:
predicting vascular age (VA) of the subject using a linear regression model different from the trained machine learning model and the output indicative of the CFPWV of the subject.
11 . The method of claim 1 , further comprising:
predicting aortic stiffness of the subject using the output indicative of the CFPWV of the subject.
12 . The method of claim 1 , further comprising:
predicting risk of the subject for developing any of hypertension, diabetes, obesity, myocardial infarction, heart failure, stroke, kidney disease, cognitive impairment and dementia using the output indicative of the CFPWV of the subject.
13 . The method of claim 1 , wherein the corresponding output indicative of the CFPWV of the subject indicates normalized CFPWV (nCFPWV).
14 . A system, comprising:
at least one computer hardware processor; and
at least one non-transitory computer-readable storage medium storing processor-executable instructions that, when executed by the at least one computer hardware processor, cause the at least one computer hardware processor to perform a method for using a machine learning model to estimate an indication of carotid-femoral pulse wave velocity (CFPWV) and/or vascular age (VA) of a subject from an arterial waveform previously measured at a single site by an arterial waveform acquisition device coupled to the subject, the method comprising:
obtaining the arterial waveform, the arterial waveform comprising a plurality of portions corresponding to a respective plurality of heartbeats of the subject, the plurality of portions comprising at least 10 portions, each of the plurality of portions comprising at least 250 data points;
generating a representative heartbeat waveform from the arterial waveform, the generating comprising:
detecting the plurality of portions in the arterial waveform;
individually detrending each of the detected plurality of portions to obtain a plurality of detrended portions,
averaging portions in the plurality of detrended portions to obtain the representative heartbeat waveform; and
processing the representative heartbeat waveform and at least one derivative waveform of the representative heartbeat waveform using a trained one-dimensional (1D) convolutional neural network (CNN) model to produce a corresponding output indicative of the CFPWV of the subject,
wherein the 1D CNN comprises multiple 1D convolutional layers, multiple rectified linear unit layers, multiple pooling layers, and a fully-connected layer,
wherein the 1D CNN comprises at least one million parameters, and
wherein the processing comprises calculating the output indicative of the CFPWV from the representative heartbeat waveform the at least one derivative waveform and the at least one million parameters.
15 . The system of claim 14 , wherein the corresponding output indicative of the CFPWV of the subject indicates normalized CFPWV (nCFPWV).
16 . The system of claim 14 , wherein the plurality of portions comprises at least 100 portions corresponding to a respective plurality of heart beats, and wherein each portion of the plurality of portions comprises at least 500 data points.
17 . The system of claim 14 , wherein the method further comprises:
predicting vascular age (VA) of the subject using a linear regression model different from the trained machine learning model and the output indicative of the CFPWV of the subject; and/or
predicting aortic stiffness of the subject using the output indicative of the CFPWV of the subject.
18 . At least one non-transitory computer readable storage medium storing processor-executable instructions that, when executed by at least one computer hardware processor, cause the at least one computer hardware processor to perform a method for using a machine learning model to estimate an indication of carotid-femoral pulse wave velocity (CFPWV) and/or vascular age (VA) of a subject from an arterial waveform previously measured at a single site by an arterial waveform acquisition device coupled to the subject, the method comprising:
obtaining the arterial waveform, the arterial waveform comprising a plurality of portions corresponding to a respective plurality of heartbeats of the subject, the plurality of portions comprising at least 10 portions, each of the plurality of portions comprising at least 250 data points;
generating a representative heartbeat waveform from the arterial waveform, the generating comprising:
detecting the plurality of portions in the arterial waveform;
individually detrending each of the detected plurality of portions to obtain a plurality of detrended portions,
averaging portions in the plurality of detrended portions to obtain the representative heartbeat waveform; and
processing the representative heartbeat waveform and at least one derivative waveform of the representative heartbeat waveform using a trained one-dimensional (1D) convolutional neural network (CNN) model to produce a corresponding output indicative of the CFPWV of the subject,
wherein the 1D CNN comprises multiple 1D convolutional layers, multiple rectified linear unit layers, multiple pooling layers, and a fully-connected layer, wherein the 1D CNN comprises at least one million parameters, and
wherein the processing comprises calculating the output indicative of the CFPWV from the representative heartbeat waveform the at least one derivative waveform and the at least one million parameters.
19 . The at least one non-transitory computer readable storage medium of claim 18 , wherein the plurality of portions comprises at least 100 portions corresponding to a respective plurality of heart beats, and wherein each portion of the plurality of portions comprises at least 500 data points.
20 . The at least one non-transitory computer readable storage medium of claim 18 , wherein the method further comprises:
predicting vascular age (VA) of the subject using a linear regression model different from the trained machine learning model and the output indicative of the CFPWV of the subject; and/or
predicting aortic stiffness of the subject using the output indicative of the CFPWV of the subject.