PhysioFT Blood Pressure HRV Sp

About PhysioFT Blood Pressure HRV Sp

PhysioFT - STRICTLY EXPERIMENTAL app to estimate BLOOD PRESSURE (BP), HEART RATE VARIABILITY (HRV), SpO2 using phone camera video
Blood pressure (BP) and blood glucose levels are probably the two most commonly monitored cardiovascular parameters. Many people monitor their own BP outside of healthcare facilities and thus rely on mobile devices.
PhysioFT presents an algorithm which can run on a common smartphone, using the camera to capture a photoplethysmogram (PPG), which is analyzed via Fast Fourier Transform (FFT), feature extraction from the FFT, and MultiLayer Perceptron (MLP) training on the features. It yields acceptable values for diastolic BP (DBP), BP range (RBP) and sometimes systolic BP (SBP). The use of time-consuming MLP restricts the usability of the smartphone app.
As byproducts it also estimates heartrate variability (HRV) via the FFT, and SpO2 via empirical linear regression.
The ALGORITHM which evolved between 2016 and 2019 is
1. capture a PPG from the smartphone camera while holding a finger over it, with moderate background (sunlight) illumination. For training simultaneously record BP with a cuff.
2. perform a FFT on the captured PPG
3. derive 35 “features” of the FFT power spectrum
4. train a MLP on at least 100 sets of the 35 features with known BP values
USAGE: 1. To record a PPG the user presses “Start” (Screenshot 1), waits until at least 512 frames have been analyzed, then presses “Stop” (Screenshot 2). For a training sample, the SBP, DBP and HR must be simultaneously recorded with a BP cuff, e.g. with the cuff on the left arm and the phone in the right hand with the right index finger over the camera.
2. To run the FFT and estimate HR and SpO2, the user presses “CalcHR FT” (Screenshot 3). HR is estimated by both peak counting and maximum FFT amplitude peak in the range 0.5-2.5 Hz; if the 2 results differ, the peak counting method is usually more accurate.
3. The “Save” button (Screenshot 4) saves measured parameters and FFT-derived features to the file “PhysFT_Feat41.csv” in the device “Downloads” folder.
4. To train the MLPs the user presses “Train BP” - THIS STEP CAN TAKE MANY HOURS. Due to the huge number of MLP floating-point calculations, the phone may become very warm, and the battery may be drained if the phone is not plugged into a power source. This step may also “freeze” the phone if it uses all available cores. It does not display any values. The best sets of node-node weights are saved in "Downloads" as PhysFT_WgtSBP.csv, PhysFT_WgtDBP.csv, PhysFT_WgtRBP.csv
5. To estimate a new, unknown BP the user presses “Guess BP” (Screenshot 5). This displays estimates of all “unknown” BP recordings (those where the SBP and DBP were not set).
Re SpO2: the brightness is not the only feature of a smartphone video that oscillates with time. The hue (shade of red) of the image varies as the relative portions of oxygenated versus deoxygenated blood vary with each heartbeat. An empirical linear regression can estimate SpO2. Of course this is valid only for the test person.
Re: heartrate variability (HRV): The average HR is an average of several similar frequencies. The amplitudes of the lesser peaks relative to the main peak are a way to measure HRV.
The barrier to the use of the app for BP estimation presented here is the time required for the MLP calculations.
Future work: There are technical issues which can and should be improved.
1) For BP, the number of nodes in the MLP hidden layers should be optimized. Starting weights of the node connections can be chosen rather than initialized randomly. The amount of training data from each individual must be much greater.
2) For HRV, the simple weighting formula is clearly just a first step.
3) For SpO2, the linear regression was based on data from one individual, and the altitude factor was fixed at 40 m above sea level. Clearly, more data from many individuals must be collected.