In this work, we report on an innovative pattern recognition algorithm to perform the post-processing of PhotoPlethysmoGraphy (PPG) signals acquired via a multichannel ECG+PPG combo portable system. The functional parts of the system are the probes, the Analog Front-End, and the main PPG/ECG-Embedded Subsystem. Each PPG probe includes a 940 nm LED and a Silicon Photomultiplier detector allowing good responsivity and high gain. The use of Texas Instruments ADS1194 biopotential measurements sampling family offers great scalability and up to 4 simultaneous sampling channels (16-bit resolution, 1kSPS sampling rate). Thanks to our system, it is possible to perform a real-time monitoring of cardiovascular parameters by simultaneously acquiring 2 PPG waveforms at different body locations and 1 ECG lead. Preliminary measurements demonstrate good quality of the acquired PPG signals, being even possible to extract the Augmentation Index from a single measurement on the wrist (rather than from a pressure wave) to estimate the arterial stiffness.
Herein, a brief description of the pattern recognition algorithm is illustrated. The collected PPG raw signal is preliminary filtered through classical IIR filter composed by low/high pass filters and then processed via ad-hoc self-adaptive pipeline using a nonlinear system. However, the filtered PPG signal needs further processing being still affected by noise and signal-distortion due to breath activity, motion artifacts, and micro-vibrations. Preliminary first and second derivative computation is performed to detect relative maximum and minimum values. Afterwards, the proposed pipeline performs normalization in [0, 1] and a basic segmentation of the collected PPG timeserie supposing to find compliant PPG waveform between two subsequent minimum values. A self-adaptive nonlinear oscillator is properly configured to generate a compliant PPG waveform according to a novel mathematical model of the PPG signal proposed to address the issue of PPG corrupted signal. For each collected PPG segmented waveform, the reference PPG signal is temporally rescaled using “nearest” algorithm to get both waveforms (reference and collected) time-comparable. Ad-hoc sample cross-correlation analysis between rescaled-normalized PPG waveforms will be finally performed by the proposed pattern recognition pipeline. High correlated PPG waveforms will be accepted building a robust and cleaned PPG timeserie, while the low-correlated ones will be discarded. The collected results obtained by using the proposed pattern recognition pipeline confirms the robustness and efficiency in terms of sensibility/specificity ratio of the proposed approach.
This activity was supported by Advancing Smart Optical Imaging and Sensing for Health (ASTONISH) Project (Grant no. 692470), funded H2020-EU.220.127.116.11. – ECSEL programme.