Blood Pressure

Cardiovascular disease is a leading cause of death and disability, and elevated BLOOD PRESSURE is a leading contributor to disease risk. Screening for, diagnosing, and following the response to therapy for hypertension are constrained by current measurement methods that are subject to variability because of a wide variety of potential measurement conditions over the course of one’s daily activities. Traditional brachial artery blood pressure measurement devices are inconvenient, uncomfortable, and require special equipment because of their reliance on inflatable cuff-based technology. This study reports on a new technology that measures blood pressure continuously and without contact from video of a person’s face. In this initial study on normotensive subjects, we show that this technology exhibits comparable accuracy to traditional automated blood pressure monitors. However, this optical imaging technology implemented on a smartphone would improve upon traditional cuff-based devices by being more convenient and more comfortable (eg, cuff-less). This is likely to encourage measurements in more places and with more regularity than before and provides a comprehensive picture of patients’ blood pressure throughout the day, much like an ambulatory blood pressure monitor. The tool is to revolutionize hypertension diagnosis and management and begin to address the incredible burden of cardiovascular disease worldwide.

Clinically significant elevated blood pressure (BP; hypertension) afflicts >25% of adults worldwide. Hypertension also constitutes a major modifiable risk factor for cardiovascular disease. Nonclinic BP monitoring in the form of standard automated BP monitors and ambulatory BP monitors is highly recommended in dealing with this epidemic. It provides patients and health professionals with a representative picture of a patients’ BP throughout the day and reduces the cost and inconvenience associated with clinic visits in the diagnosis and management of hypertension. Nevertheless, nonclinic BP monitoring has not reached its full potential because standard automated BP monitors are not convenient to use outside of the home, and because ambulatory BP monitors are uncomfortable to wear throughout the day. Therefore, a tool is needed that can accurately measure BP comfortably and conveniently anywhere and anytime.

BitDoctor AI imaging technology uses several state-of-the-art techniques in photoplethysmography with respect to extraction of raw signal (eg, region of interest tracking, multiple raw signals, 3 color channels) and estimation of plethysmographic signal (eg, bandpass filtering). However, unlike any other video-based technology, BitDoctor AI imaging technology separates each video image into multiple layers called bitplanes in each of the 3 color channels. Then, using a machine learning-based algorithm developed using blood flow data collected concurrently from an Federal Drug Administration (FDA) cleared BP measurement system (Methods in the Data Supplement), BitDoctor AI imaging technology extracts hemoglobin-rich signals and discards melanin-rich signals from each image of the video sequence. Next, the hemoglobin signals from all bitplanes of each frame of the video are recombined to produce an image representing a map of hemoglobin concentration across the face. By linking all the images together in their original sequence, it produces a video of hemoglobin concentration changes representing facial blood flow oscillations. This unique methodology produces a robust signal with minimal noise and minimal susceptibility to variations in skin tone.

BitDoctor.ai tested the hypothesis that information contained within these facial blood flow oscillations can indicate systolic, diastolic, and pulse BPs. This hypothesis is based on the following 3 existing sets of evidence. First, BitDoctor AI imaging technology already detects heart rate with accuracy equal to an ECG and heart rate variability with comparable accuracy. This demonstrates that blood flow changes revealed by BitDoctor AI imaging technology correspond with the systemic cardiovascular changes engendered by the pulsating heart. Second, studies using photoplethysmography have shown that hemoglobin changes in the fingers contain important information about arterial pressures in the form of photoplethysmography waves. While the information obtained from finger photoplethysmography is relatively homogeneous, BitDoctor AI imaging technology is able to obtain such information from multiple locations simultaneously, thus taking into account differential microvascular control of the face by sympathetic and parasympathetic vasomotor neurons. BitDoctor AI imaging technology should, therefore, provide richer information about BP and thus produce more accurate measures than finger photoplethysmography. Third, a recent study combined photoplethysmography and a smartphone to determine brachial systolic, diastolic, and pulse BP accurately without calibration by a brachial cuff. This technology used a finger photoplethysmography sensor to detect blood flow as the subject presses their finger against the sensor with progressively greater levels of force. A finger pressure sensor guided the application of that force, and the smartphone used force and blood flow information to compute BP oscillometrically. The study demonstrates that absolute brachial pressures can be accurately estimated using photoplethysmography elsewhere on the body and without calibration with a brachial cuff. Like photoplethysmography, BitDoctor AI imaging technology optically captures blood flow data and then uses it to determine BP. Furthermore, it does so remotely, using only camera hardware that is already ubiquitous on existing smartphones without the need for a photoplethysmography instrument or pressure sensor.

A successful proof-of-concept study used a smartphone to video-record the faces of subjects while simultaneously collecting their systolic and diastolic BP reference measurements using an FDA-cleared continuous BP monitor. This monitor measures upper arm (brachial artery) BP continuously, thus providing an objective and continuous characterization of BP changes that occur throughout the video recording session. After applying the data processing algorithms (see methods) to the face video, we obtained blood flow signals that track blood oscillations in multiple locations of the face frame by frame. It is divided the sample into a training set (70%), a testing set (15%), and a validation set (15%). Advanced machine learning algorithm to train and test computational models to predict reference BP from these signals using data from the training and testing sets. The remaining validation set was never used in training or testing. This independent data set was thus used to evaluate how well the trained models would generalize to predict systolic, diastolic, and pulse pressures in new subjects that they had never seen before. This study recruited subjects with normotensive BPs, with the high BP cutoff defined by Eighth Joint National Committee general population criteria. Subjects had a systolic pressure between 100 and 139 mm Hg and a diastolic pressure between 60 and 89 mm Hg. This range was sufficient to build computational BP models and determine whether BitDoctor AI imaging technology can be used to measure BP.

Clinical Accuracy Test Set Up

Signal Processing

Our video recordings of the face captured light re-emitted by blood hemoglobin. The amount of light captured by the camera was, therefore, inversely proportional to hemoglobin concentration near the skin surface. It is widely acknowledged that this pulsation of light reflects the pulsation of arteries under the skin. However, these changes in re-emitted light are essentially imperceptible in conventional videos. For this reason, it was necessary to use signal processing techniques to extract and amplify cyclical blood pulsations within the human facial vasculature. See Methods in the Data Supplement for signal processing about the technology and reference BP signal.

Features Extracted From Subject Data

TOTAL 155 unique features ARE EXTRACTED from participant data. The first 126 of these features were extracted from facial blood flow signals from each subject’s 17 regions of interest. These features fall into the following categories: pulse amplitude, heart rate band pulse amplitude, pulse rate, pulse rate variability, pulse transit time, pulse shape, and pulse energy. The remaining 29 features consisted of meta-features that help normalize for different imaging conditions, as well as features pertaining to ambient room temperature and subject physical characteristics (eg, age, weight, and skin tone).

Eigenvectors Used for BP Prediction

After extracting these features, we used SPSS to conduct principal component analysis on extracted features to reduce feature dimensions. We used the varimax rotation to produce 30 orthogonal eigenvectors.

Training of BP Prediction Models

These 30 decorrelated eigenvectors were then input into a multilayer perceptron machine learning algorithm (SPSS, Version 24) to generate models that best predicted: (1) systolic BP, (2) diastolic BP, and (3) pulse pressure.

The data is divided into a training set (70%), a testing set (15%), and a validation set (15%). Trained and tested our multilayer perceptron models with the training and testing sets. The data then validated these models on the independent validation set that was not used in training or testing. This independent data set was used as an objective indicator to evaluate how well the trained models would generalize to predict systolic, diastolic, and pulse pressures in new subjects that had never seen before. Officially trained, tested, and validated the models with 200 iterations for each type of model to generate statistical estimates of model performance.

For comparison purposes, a set of info also created separate control models for systolic, diastolic, and pulse pressure using only age, height, weight, skin tone, sex, race, and heart rate as predictors. By doing so, accuracy is to ascertain the models based all features were able to predict BP above and beyond the contribution of demographics features independent from video features. The rationale for doing so was that the demographic features could be readily obtained without using BitDoctor AI Imaging Technology. For the same reason, heart rate added as a predictor in these control models. Note that although used heart rate measurements based, such measurements could also be readily obtained without the technology (eg, palpation, smartwatch, and ECG).

Statistics: Testing Performance of the BP Prediction Models

BitDoctor.ai quantified the accuracy and precision of each of the 200 iterations of systolic, diastolic, and pulse pressure models as a percentage accuracy, a mean bias±SD, as an intraclass correlation, and as a Pearson correlation. These calculations were performed on each of the 200 iterations for each type of BP model, and the mean and 95% CI was reported for the 200 iterations of each model. All accuracy statistics were calculated on the validation set only. For percentage accuracy, we took the absolute difference (error) between the predicted BP and the reference BP for a given subject and divided it by the reference BP of the subject to get a proportion of error. We then subtracted this value from 1 to convert this proportion error to a proportion accuracy and multiplied it by 100 to obtain a percentage accuracy. We then calculated the mean accuracy across all subjects to arrive at a percentage accuracy for each model. For mean bias and SD, we calculated the difference between the reference and predicted pressure for each subject for the systolic, diastolic, and pulse pressure prediction models. We then calculated the mean and SD of this difference for all subjects in each respective model. Intraclass correlation estimates and their 95% CI were based on single measure absolute agreement in a 2-way mixed-effects model. Pearson correlations and their 95% CIs were also calculated. A plot of reference versus predicted pressures was constructed using the mean of 200 predicted values for each model to display predictive ability across the range of reference BPs.

BitDoctor.ai further determined the degree of information gain attained by each predictive model to determine the predictive power of each model beyond that of simply predicting the mean. Theoretically, the SE of the predictions when the measurement bias is zero will equal the SD of the reference pressures if the mean is predicted every time. A greater reduction in the SD of the residual (prediction error) relative to reference BP SD demonstrates a greater degree of correct predictive ability. To quantify predictive ability, we took the absolute difference of these 2 values and divided it by the SD of the reference BP. We converted to a percentage by multiplying by 100. Thus, a greater percentage corresponds to greater information gain for that model relative to reference SD.

To assess eigenvector importance in each iteration of the systolic, diastolic, and pulse pressure models, we determined the relative importance of each of the 30 eigenvectors normalized against the most important eigenvector. We averaged eigenvector importance across all 200 iterations of the model to rank each eigenvector according to its average importance for each model. Although eigenvectors represent abstract dimensions in our data, to help readers understand what these dimensions may represent, we highlighted the most representative features of each eigenvector, which had the highest values in the rotated component matrix.

Results

Tested proven the hypothesis that BitDoctor AI Imaging Technology accurately detects BP from video of the face. Hypothesis in 2 parts. Firstly is determined whether oscillations in signal reflected oscillations in continuously measured BP. This to determine BitDoctor AI Imaging Technology captures BP on a qualitative level. We then quantitatively assessed our BP prediction models against reference systolic, diastolic, and pulse BP measurements.

Signal Resembles Reference BP Pulses

We anticipated that BitDoctor AI Imaging Technology signal would reflect hemoglobin concentration in the face. This signal would, therefore, serve as a surrogate for blood volume and ultimately BP. To identify regions in the face with robust hemoglobin signal, we constructed a spatiotemporal map of the signal in the face. This full facial map allowed us to examine areas of the face where subcutaneous vasculature was either under the control of the sympathetic nervous system (eg, lips and nose) or the parasympathetic nervous system (eg, forehead, chin, and lower jaw); we anticipated both could provide useful information. Using this spatiotemporal map, we identified the 17 regions on the face that could provide robust hemoglobin signals.

A. Our AI imaging view B. 17 region of interest C. Signals we reading

Accurately Determines BPs

All the computational models based on multilayer perceptron are highly accurate in terms of predicting the reference BPs of our validation cohort. On average, models predicted systolic BP with an accuracy of 94.81%, diastolic BP with an accuracy of 95.71%, and pulse pressure with an accuracy of 95.75%. The average prediction biases±error SDs were 0.39±7.30 mm Hg for systolic BP, −0.20±6.00 mm Hg for diastolic BP, and 0.52±6.42 mm Hg for pulse pressure. These SDs represent information gains of 25.5%, 12.0%, and 21.8%, respectively. Our findings corresponded to average intraclass correlations of 0.60, 0.37, and 0.56 and average Pearson correlations of 0.67, 0.47, and 0.63 for systolic, diastolic, and pulse.

A screenshot of a medical report Description automatically generated

Below image shows the scatter plots of reference versus predicted pressures as well as the line of identity. At low reference pressures predicted pressures tend to fall above the line of identity and at high reference pressures predicted pressures tend to fall below the line of identity. There is some degree of overprediction at low reference pressures and some degree of underprediction at high reference.

abc