Redefining Blood Pressure Monitoring for Modern Era using MFine App
Traditional blood pressure (BP) monitoring involves using sphygmomanometer with an inflatable cuff and stethoscope. The blood pressure readings are determined by listening to Korotkoff while the cuff is manually inflated and deflated. Then, automated BP monitors were introduced, which use electronic units to automatically inflate and deflate the cuffs to determine BP readings. The use of cuff is non-invasive; however, it is an obtrusive and causes inconvenience to the user, which hinders the regular BP measurement and tracking.
Photoplethysmography (PPG) signal represents the changes in blood volume during each cardiac cycle. The PPG signal can be used to estimate various cardiovascular parameters, such as heart rate, arterial oxygen saturation, and BP. Recent advances in wearable technology have led to the development of compact and portable PPG devices that allows for the continuous monitoring of cardiovascular parameters in real time. PPG based blood pressure monitoring has gained popularity as a non-invasive and cuff-less method for measuring blood pressure. The continuous and non-invasive monitoring of various cardiovascular parameters in a cost effective manner has a potential to revolutionize the way cardiovascular diseases are diagnosed and managed.
At MFine, we have developed a mobile app for health vital measurement based on PPG. The App utilizes the flashlight and camera sensor as a light source and detector, respectively to measure the pulsatile information caused by blood flow. The PPG is extracted from the video signal and machine learning approaches are employed to measure blood pressure. The video below shows the steps involved for BP measurement using the MFine App. The App is available for download from the Google Play Store. The digital signal processing and machine learning steps involved in measuring BP from the PPG signal are discussed in this article.
Extraction of PPG from Image Sequence captured via smartphone camera
Extracting PPG signals from video data collected using smartphone cameras is a relatively new and rapidly developing area of research. The underlying principle behind this approach is akin to traditional PPG measurements, wherein changes in light absorption or reflection caused by pulsatile blood flow are detected. In this case, the smartphone camera is utilised to record video data of a specific area of skin, typically the fingertip or face, which is illuminated using a light source like the smartphone flash or ambient light.
The video signal recorded in the MFine App is utilised to extract the PPG signal using a digital signal processing approach. The pixel intensity in each frame of the video data captures the fluctuations in blood volume caused by the heartbeat. The visual representation of the 1-D PPG signal extraction from the video data is shown below (Figure 1).
The signal processing approach implemented to analyze the variations in pixel intensity in each frame is outlined below:
Let the pixel intensities of red, blue and green channels be represented as Ir, Ig, and Ib respectively. The corresponding PPG signal extracted from the red, blue, and green channel is denoted as PPGr, PPGg, and PPGb respectively. The PPG signal is constructed by averaging the pixel intensities in each frame as given in Equation 1.
In equation 1, i is the frame number, x and y are the pixel location and M is the total number of pixel locations. If there are N frame in the video data then i varies from 1 to N.
The PPG signal contains various frequency components including the fundamental heart rate frequency and its harmonics as well as noise and artifacts. Low frequency noise can arise from various sources such as respiration and changes in skin temperature. The high frequency noise can be caused by powerline interference and motion artifacts. To enhance the signal quality, digital filtering is crucial. The PPG signal is bandpass filtered between 0.1 and 8 Hz to preserve important physiological information while attenuating low and high frequency noise.
Feature Engineering and ML for BP measurement using PPG.
Blood pressure measurement using PPG involves extracting relevant features from the PPG signal that are associated with BP. The contraction and relaxation of heart during each heartbeat result in corresponding increases and decreases in blood pressure, known as systolic and diastolic BP values. Estimating beat-to-beat blood pressure from PPG involves analysing individual pulse segments of the PPG signal.
The first step in estimating the beat-to-beat BP from PPG is to detect the individual pulse segments from the PPG signal. This is typically achieved using a peak detection algorithm, which identifies the local maxima in the signal corresponding to each pulse. The time between successive pulses is known as the pulse interval, which is inversely proportional to the heart rate.
A noise free PPG segment representing the pulsatile information between heart beats consists of four landmark points: onset (O), systole peak (S), Notch (N), and diastole peak (D). The peak detect method is applied once again on a segment of PPG data to detect the above four landmark points. Additionally, the first and second derivatives of PPG signal, referred to as VPG and APG, respectively, are extracted from the PPG segment. Various time domain related features are extracted from the PPG, VPG, and APG segment. The block diagram of BP measurement using PPG signal is shown in Figure 2.
- Timespan Features: The time interval between different landmark points are extracted. For instance, the timespan between onset and systole peak is computed as below
Similar to PPG landmark points, the peaks and troughs of VPG and APG form the landmark points . The timespan between each of these landmark points are extracted as timespan features.
- Amplitude Features: The absolute, relative and the ratio of the amplitudes for all the landmark points of PPG, VPG, and APG segment are extracted
- Waveform Area Features: The area under the curve between two landmark points is extracted as waveform area features. For instance, area features of PPG waveform are:
- Energy Features: The sum of square of the sample values between two landmark points are computed as energy features.
- Slope Features: The ratio of the relative amplitude to the timespan between two landmark points is referred to as slope feature.
After extracting all the above features, two separate Random Forest (RF) model is trained for estimating systolic and diastolic BP values.
Results and Analysis.
In this section, the data used to train the model and the performance of the BP model is discussed. The video samples are recorded using 18 different smartphones and each recording is of 30 sec. In total 1,200 PPG samples were collected from an outward patient department of a partner hospital of MFine. The data distribution of the train and test split is shown in Figure 3 (left and middle). The BP values are non-uniformly in both the train and test datasets. The histogram in Figure 3 (left and middle) suggests that the number of samples are highest in 80–100 mmHg and 120–130 mmHg ranges for diastole and systole BP , respectively.
The MAE achieved on the test dataset using the proposed method is shown in orange and blue color points for diastolic and systolic BP values, respectively in Figure 3 (right). The MAE achieved with the proposed method has minimum absolute error for ~80 mmHg and ~120 mmHg for diastole and systole BP values. The overall MAE achieved using the proposed method is 14.30±13.37𝑚𝑚𝐻𝑔 and 9.11± 8.06𝑚𝑚𝐻𝑔 for systolic and diastolic BP, respectively. Further, the MAE achieved for hypotension (< 120), normal ((120 − 130)/(80 − 90)), and hypertension (> 140/90) are reported in Table I.
The entries of Table I and II are also utilised to benchmark the proposed method against US Association of American Measurement Institute (AAMI) and British Hypertension Institute (BHI). The proposed method meets the criteria of both AAMI and BHI for BP measurement in the range of 80–90 mmHg of diastole and 120–130 mmHg of systole. The predictions errors are slightly higher beyond the above mentioned BP values. It is due to the fact that the dataset is highly imbalanced and has non-uniform BP values.
Concluding Remarks
The development of a mobile App based BP measurement holds immense promise in the field of hypertension management. The convenience and ease of use offered by MFine App for BP measurement make it a valuable tool for individuals seeking to monitor their BP levels regularly. As per the AAMI and BHI standards, the MFine BP measurement tool ensures accurate readings within the systole and diastole range of 120–130 and 80–90, respectively.
While the innovative approach of BP measurement using App shows great potential, further improvements are necessary to expand its effectiveness across the systole and diastole range of values. This can be achieved using training data having uniform distribution of BP values across all the possible range of values. In other words, the training dataset needs to be balanced and have equal representation for all the BP range. However, obtaining a balanced dataset is particularly challenging in the healthcare domain.
Nonetheless, the advent of MFine App based BP measurement solution brings us closer to a future where hypertension management can be seamlessly integrated into our daily lives. The convenience it offers empowers individuals to take control of their health with ease and provides a foundation for proactive monitoring and early intervention. With the ongoing research progress and refinement of the BP measurement using MFine App, this method has the potential to become an indispensable tool for healthcare professionals and individuals alike, aiding in the early detection and effective management of hypertension.
