Cardiorespiratory parameters monitoring by the use of contactless technology

The present work aims at demonstrating the feasibility of fully automated method to monitor cardiorespiratory parameters by remote system. The system consists of a camera that detects reflected light changes proportional to blood volume pulse changes from subject's face. Three different scenarios were conducted: Experiment 1: A rest-to-stand experiment was conducted on 60 healthy subjects to compare the sympathetic activation through HRV obtained from ECG RR intervals and PRV obtained from vPPG pulse-to-pulse-intervals (PPI). Experiment 2: A rest experiment was conducted to evaluate the breathing modulation on PPG signal, asking the subjects to breath at controlled respiratory frequencies [0.1-0.5] Hz with frequency step of 0.1 Hz. The breathing system modifies the PPG waveform in terms of amplitude, width and pulse time occurrence. Therefore a set of methods based on pulse amplitude variability (PAV), pulse width variability (PWV) and pulse rate variability (PRV) was exploited to derive real-time breathing rate estimation. A second dataset is recorded to test the breathing rate measurement accuracy during breathing exercise: 2 minutes of spontaneous respiration followed by apnea and successive 2 minutes of recovery stage. Experiment 3: A pool of subjects with diff erent cardiac rhythms (atrial fibrillation, AF, atrial flutter, AFL, and normal sinus rhythm, SR) was recruited to test the feasibility of vPPG system as method to detect cardiac arrhythmic. events. A set of features was extracted from PPI series: SDNN, RMSSD, sample entropy and the ratio SD1SD2 that expresses the spatial distribution of PPI in Lorenz plot. Beside them, two novel parameters were introduced: the power harmonic strength (PHS), index of percentage of energy located around the dominant frequency of vPPG spectrum respected to the total spectrum energy, shape similarity (SS) that measures the PPG wave similarity along time running windows. A feed forward neural network (FFNN) classif er was trained using the most relevant features (selected by the ANOVA statistical method) in order to discriminate the three di fferent categories of rhythm. The results are presented according to each experiment: Experiment 1: The postural change from rest to stand provokes a slight decrease both of signal-to-noise ratio of vPPG signal (average decrease of 6.75 dB) and of the total accuracy of pulse detection (0.984±0.021 % at rest vs 0.979 ±0.023 % in a standing position with no statistical signi ficance). The main reason is the increase of motion noise introduced by the subject to maintain the balance. Pearson correlation was used to compare temporal and frequency domain parameters between PRV and HRV. In rest condition, both temporal and frequency features maintained a correlation greater than the 0.9 threshold. In the stand ing condition, only SDNN had a high correlation, while the other parameters (RMSSD, LF, HF) had a drop in concordance. The decrease of correlation between the two systems in the standing condition explains the moderate change in the LF/HF ratio found in PRV (0.49±0.34 in rest vs 0.70± 0.51 in standing) while in HRV the sympathetic activation is more evident (0.51 ±0.38 in rest vs 1.17±0.79 in stand). Experiment 2: The results are expressed as median and interquartile ranges of relative error calculated between the estimated breathing frequency by vPPG and the gold standard respiratory belt. In the rest database a good perfomance in the estimation of breathing rate was found in the respiratory frequency range [0.2-0.4] Hz (0.89/3.03 % expressed as median/IQR). The relative error increased dramatically at the respiratory frequencies boundaries (fR = 0.1 Hz with 8.80/35.69 % and at fR = 0.5 Hz with 38.65/36.55 %). In the second database the relative error remained acceptable both during spontaneous breathing (0.43/2.11 %) and during the recovery stage (0.46/4.81 %). Experiment 3: Among all the features used, only PHS and SS displayed a statistically signifi cant di fference between the three classes SR, AF and OA (class containing AFL rhythm). The general performance of the FFNN is 85.9 %, that represents the percentage of the total correct predictions. The main drop of performance regarded the misclassi cation between the two classes AF and OA. Among 55 patients su ering of AF or AFL, 12% of subjects are labeled incorrectly. Reducing the complexity of classi cation to only two classes, SR and arrhythmias, the classi er improves its performance to 95%.

La seguente tesi di dottorato ha come scopo principale la presentazione di un sistema contactless che, attraverso l'utilizzo di una normale telecamera, rileva parametri cardiorespiratori dal volto del soggetto. Le variazioni di luce riflessa dal viso percepite dalla telecamera, sono proporzionali alla variazione di sangue ematico che scorre nelle arterie periferiche. L'applicazione di tale tecnologia sarà determinante nel monitoraggio di soggetti sensibili al contatto, come anziani o bambini, o su soggetti con ustioni cutanee.