Research and development process

Cardiox offers a unique solution that provides both a photo-plethysmogram and a cardiogram at the same time. By combining ECG results and cutting-edge math formulas it became possible to estimate blood pressure without a cuff. We have developed an API to integrate external ECG data sources with the CardioX mathematical core.

We have acquired hundreds of thousands of ECG records of both men and women, aged from 15 to 93 y.o. Along with ECG, the records contain a person’s Blood Pressure readings taken at the same time as ECG, as well as a number of individual metrics, such as gender, age, physical health, etc., and load it into our Neural Network.

The Neural Network analyzes a 10-second ECG interval, with the discrete frequency of 1000 counts/sec. Therefore, we have 10000 points for each ECG interval. Then it uses mathematical algorithms to measure blood pressure.

CURRENT PERFORMANCE

CardioX neural network is structured as an interconnected group of smaller neural networks, which estimate 2 parameters: systolic and diastolic blood pressure. In the table, you can see the accuracy of the BP estimates, based on the random test sampling of 15% of the total number of records.

The probability of the CardioX neural network to estimate systolic BP with <10 points deviation equals 72%, and for diastolic BP – 92%.

We have acquired hundreds of thousands of ECG records, Blood Pressure readings, and individual metrics, such as gender, age, and physical health, all taken at the same time. All parameters mentioned above and many more were used in our machine learning models. We eliminated all irrelevant and questionable records. Then, we removed all noise from ECG signals and straightened the isoline.

Noise Filtering

Many ECG intervals were distorted by noise. Since we needed this data to teach the neural network, and considering that the ECG signal from our CardioX ECG monitor and other external and wearable devices could be imperfect, we used a number of mathematical algorithms which allowed us to extract a clean ECG signal from the distorted one.

Images below show the succession of ECG signal processing to remove noise and bring it to the  final version of usable ECG interval.

Rate Variability (HRV)

We analyze the difference between each interval and their statistical average, a number of intervals that are significantly different vs a number of intervals that are very similar. Based on these values we can build a histogram. It indicates the presence of arrhythmia, which can affect persons Blood pressure levels.

ECG Signal Transformation

Even though quite a few ECG intervals we had were relatively free of high-frequency noise, they were still somewhat distorted, most likely by patients’ movements during ECG. To remove baseline drift from ECG signals we created and applied techniques that allowed us to remove low-frequency artifacts causing baseline drift. The images below show the results of using these techniques.

Physiological Parameters

By developing and applying the algorithms for detecting peaks, we could calculate over 50 derivative parameters, that we used as features for machine learning.

Statistical Parameters

ECG interval can also be viewed as a systematic distribution of parameters that can be analyzed statistically. This approach allows us to apply statistical methods to calculate not only simple values like sum, average, median, mad, root mean square deviation, etc. as well as more complex parameters. Moreover, we could also calculate fractal-based self-similarity coefficient of ECG signal.

Now the CardioX team is working in a number of areas, including the development of a device in the form factor of bracelets for both hands, as well as  a mobile app for the rapid evaluation of ECG photos.