Advanced machine learning techniques to improve the quality of the PPG signal

The quality of the PPG signal is crucial to achieve the highest diagnostic accuracy possible and to and to avoid misleading analyses as much. The challenge of the PPG signal is the susceptibility to noise artefacts that impact the quality of signal, and therefore can result in the absence or extra presence of peaks. Therefore a unique method of signal processing was developed at FibriCheck that was submitted for a patent. This technique takes an advanced look at the camera image that is acquired. In the background of the FibriCheck application the camera is recording the signal and the information of each individual camera pixel is analysed. Based on this information the signal processing is tuned to obtain only good quality signals which are used to construct the final signal. 

A peak under the hood: The camera pixels are analysed to only detect good quality data

This results in improved interpretation, as can be seen in the following graph. For the validation of the signal processing, an ECG signal was simultaneously recorded with the smartphone-derived PPG signal. In the ECG signal, it is clear that there are two short-coupled supraventicular ectopic beats. However, in the native signal of the PPG these coupled beats cannot be detected due to the absence of the pulse. Typically, the volumetric response of that heartbeat is compromised resulting in an unmeasurable result.

Enhancing the signal quality with traditional filtering and signal processing techniques slightly improves the quality of the signal but does not reveal the presence of those short-coupled ectopic beats. Using the novel method of FibriCheck not only enables to improve the general quality of the signal, but also reveals the presence of these beats, resulting in a more accurate assessment of the heart rhythm. This technique is a fundamental component to achieve medical grade quality, and to ensure that arrhythmias are correctly detected.

In case of regular heart rhythms, the signal quality is relatively easy to determine, typically amplitude and predictive RR-intervals are used as a quality metric. However, in the case of irregular heart rhythms, these methodologies are not applicable. Therefore, new and dimensionless parameters are required to identify and differentiate bad quality data from arrhythmias. Below an example of an atrial fibrillation patient can be found. In the top and middle graphs,  the raw signal and after conventional signal processing still contain some doubtful segments that cannot result in a high quality RR-interval analysis. When applying the FibriCheck processing, a high-quality PPG signal is obtained enabling in-depth analysis of the true signal and defining the correct heart rhythm.