Finger and Forehead PPG Signal Comparison for Respiratory Rate Estimation Based on Pulse Amplitude Variability

Transcription

1 Finger and Forehead PPG Signal Comarison for Resiratory Rate Estimation Based on Pulse Amlitude Variability Alberto Hernando, María Dolores Peláez, María Teresa Lozano and Montserrat Aiger Centro Universitario de la Defensa (CUD) Academia General Militar (AGM) Zaragoza, Sain Eduardo Gil BSICoS Grou, I3A, IIS Aragón University of Zaragoza Zaragoza, Sain Jesús Lázaro Deartament of Electrical Engineering, (ESAT), STADIUS KU Leuven, Leuven, Belgium IMEC, Leuven, Belgium Abstract Pulse hotoletysmograhic signal (PPG) is modulated by the resiratory rate, so there are some algorithms caable to extract resiratory information from the derived PPG signals, as the Pulse Amlitude Variability (PAV). Previous works have shown that the use of the PPG leads to different results deending on the PPG sensor location (finger and forehead). Therefore, a database recording finger and forehead PPG signals and resiration is done, breathing with fixed frequencies. Results show that while finger PAV signal works correctly, forehead PAV signal has a non resiratory comonent that do not allow to roerly estimate the resiratory rate. I. INTRODUCTION Pulse hotoletysmograhic signal (PPG) is a non-invasive technique widely used to obtain clinic monitoring information [1]. PPG has been alied in many different clinical settings, including the monitoring of blood oxygen saturation, heart rate and its variability, giving information about the autonomic nervous system, blood ressure, cardiac outut and resiration [2], [3]. Focus on the resiratory information, the roosed methods to extract the resiratory rate from the PPG signal are usually based on the modulations induced by the resiration in the ulse rate, amlitude and width variabilities (PRV, PAV, and PWV, resectively). It is known that resiration modulates PPG signal through several effects [4]: PRV is modulated by resiration as heart rate variability (HRV) is, through a henomenon well known as resiratory sinus arrhythmia (RSA); PAV is also modulated by resiration through variations in stroke volume and in blood vessels stiffness [4], and this henomenon in addition to the ressure changes in the thorax during resiratory cycle modulates also the PWV [5]. The algorithm roosed in [5] was described to extract resiratory information based on this three resiratory derived signals. This method allows to extract the resiratory rate using only one signal or with a combination of them and was validated using finger PPG sensor. Nevertheless, finger is not the only ossible location for PPG sensor. Forehead is a widely used lace where PPG sensor can be located deending on the final alication. There are some differences between finger and forehead PPG signals. Light-transmission configuration can be used in the finger but not in the forehead, where lightreflection is the only ossible configuration. This affects to the PPG morhology, obtaining a smoother waveform when the signal is recorded in the forehead [6]. These characteristics in combination with the differences in arterial routes can generate differences in the estimated resiratory rate deending on the sensor location. In this work, a comarison between finger and forehead PPG signals for the estimation of the resiratory rate by means of PAV is resented. II. MATERIALS AND METHODS A. Data collection A database of 10 subjects was recorded to erform the analysis. For these 10 subjects (mean age of 31.0±6.7 years), finger and forehead PPG signals were recorded simultaneously as well as a chest-band resiratory signal that will be used as the gold-standard. Both PPG signals and the resiratory chest-band one were recorded and samled at 250 Hz with the Medicom System, ABP-10 module (Medicom MTD, Ltd, Russia). The used rotocol consisted of 7 different stages with a duration of 3 minutes each one: first, subjects are registered during sontaneous breathing; then a different resiratory rate is imosed in each of the remaining six stages, starting at Hz and ending at Hz in stes of Hz. Only the last 2 minutes of each stage are used to extract the resiratory information of the PPG signal. B. Resiratory signal estimation From the PPG signal ( (n)), artefactual ulses were suressed by using the artefact detector described in [7]. Then, the aex (n Ai ) and the basal (n Bi ) oints of PPG ulses were automatically detected using an algorithm based on a low-ass differentiator filter [8]. Figure 1 reresents a PPG ISBN EURASIP

2 signal (measured in arbitrary units, a.u.) where its more reresentative oints are highlighted x 10 5 n Bi n Ai For both cases (finger and forehead), a ower sectrum density S k (f) is estimated every 5 seconds from the k th 40 s length running window by the Welch eriodogram, using subwindows of 12 s and 50% of overlaing. For each S k (f), the location of the largest eak f I (k) is detected. Then, a reference interval Ω R (k) is established as: Ω R (k) =[F R (k 1) δ, F R (k 1) + 2δ], (2) x n Bi 2.06 n Ai where F R (k 1) is the resiratory frequency estimated from the revious (k 1) window. All eaks larger than 85% of f(k) I inside Ω R (k) are detected, and f II (k) is chosen as the nearest to F R (k 1). Note that f II (k) can be the same f(k) I if the largest eak is also the nearest to F R (k 1). Subsequently, a measure of eakness is obtained from S k (f) as the ercentage of ower around the f II (k) with resect to the reference interval Ω R (k). The eakness is defined as: Fig. 1. PPG signals where its more reresentative oints are highlighted: uer image, finger PPG; lower image, forehead PPG. For both PPG signals, the PAV signal is estimated as the amlitude variation between the n Ai and the n Bi : d u PAV(n) = [ (n Ai ) (n Bi )] δ(n n Ai ). (1) i Then, a median-absolute-deviation-based outlier rejection rule was alied, excluding the oints with its distance to the signal median value is higher than the median distance of the rest of the oints. The subsequent series were interolated to 4 Hz by cubic slines. Then, a band-ass filter [0.075, 1] Hz was alied obtaining a signal which is denoted d PAV (n) in this aer. An examle of PAV signal is shown in Figure 2, together with the resiratory signal. Resiration and d PAV Fig. 2. Resiratory signal (blue) and PAV (red) of one subject in the Hz stage. C. Resiratory rate estimation An algorithm based on [5] is alied over d PAV (n) to estimate resiratory rate (F R ) from eaked-conditioned averaged sectra. P k = f II (k)+δ f II (k) δ FR(k 1)+2δ F R(k 1) δ S k (f)df 100, (3) S k (f)df where δ has the exerimental value of as in [5]. Then, a eaked-conditioned average sectra, Sk (f), is obtained by averaging those S k (f) which are eaked enough: L s S k (f) = χ k l S k l (f), (4) l= L s where L s was set to 2 in order to average a maximum of 5 sectra and χ k l is a criterion to consider whether the ower sectrum S k l (f) is eaked enough or not, allowing to take art in the average only to those S k (f) whose P k is above 85%. { 1, Pk 85 χ k = 0, otherwise, (5) Figure 3 dislays two sectra as examles, one with P k < 85% (not eaked enough to take art in the average), and another one with P k > 85% (eaked enough to take art in the average). Finally, resiratory rate is estimated as the maximum of S k (f): F R (k) = arg max f S k (f). (6) D. Performance analysis As mentioned reviously, resiratory rate is estimated every 5 s. The median of all the estimations er stage in every subject is comared with the original rate obtained by the chest-band information. An exerimental margin of error of ±0.03 (±8bm) is given to the estimation to consider that it matches with the gold-standard. If the match haens, a Correct Estimation (CE) is considered. If not, there is a Wrong ISBN EURASIP

3 S k (f) S k (f) Peakness = 89.02% Frequency (Hz) Peakness = 68.75% x III. RESULTS Fig. 4 shows 6 time-frequency mas of the resiratory rate estimation. Each row reresents one different stage: first row, Hz stage; second row, Hz; and third row, Hz. Left column corresonds to the finger PAV signal and the right column corresonds to the forehead PAV signal. As it can be seen, finger PAV signal is useful to estimate the resiratory rate. Nevertheless, when PAV signal is recorded in the forehead, a comonent between and Hz is found in all the stages. This masks the ossible resence of the exected resiratory information. Table I shows the ercentage of the correct estimations at each stage using the PAV signal extracted in the two different locations. The aearance of this comonent causes a huge decrease in the %CE in forehead PAV with resect to the finger one. Only in and Hz have similar values, and this haens because the non resiratory comonent is between this both values. Besides this, general results show a worse caacity to detect the correct rate in the higher frequencies with resect to the lower ones, indeendently of the sensor location. TABLE I PERCENTAGE OF THE CORRECT RESPIRATORY RATE ESTIMATION (%CE) USING PAV IN BOTH LOCATIONS Frequency (Hz) Fig. 3. Differences between sectra which satisfy the eakness condition and those which do not. Red lines illustrate the limits of the integrating interval of the numerator in P k with the solid line marking the f II (k) value. Black dashed lines illustrate the reference interval Ω R(k), with the solid line reresenting the revious resiratory rate estimated F R(k 1). Estimation (WE). The ercentage of correct estimations for each stage is used as a measure of both PAV erformance. %CE = CE 100. (7) CE + WE The algorithm used is based on sectral analysis for resiratory comonent detection. Therefore, knowing how relevant is the resiratory comonent in both PAV signals is an interesting oint. The ower around the frequency given by the resiratory chest-band (F c, with a bandwidth of ±0.05 Hz) with resect to the entire sectra of exected frequencies (from 0.05 to 5 Hz) is comuted. The relative ower in normalized units (P R ) is defined as: P R = f=fc+0.05 f=f c 0.05 f=5 f=0.05 S k (f). (8) S k (f) Location Natural Finger Forehead Finally, Table II shows the inter-subjects mean and standard deviation (std) of the relative ower in normalized units (n.u.) inside each band. A decrease of the relative ower of the forehead is found in comarison with the finger at any stage. TABLE II MEAN ± STD OF THE RELATIVE POWER (N.U.) FOR EACH STAGE WHEN THE RESPIRATORY RATE IS EXTRACTED USING THE PAV SIGNALINBOTH PPG LOCATIONS Location Finger Forehead ± ± ± ± ± ± ± ± ± 9.09 ± 3.50 ± 2.77 ± 3.52 IV. DISCUSSION Several information can be extracted from the use of PAV signal to estimate the resiratory rate. Using this method it is ISBN EURASIP

4 Finger PPG, Hz stage Forehead PPG, Hz stage (a) (b) Finger PPG, Hz stage Forehead PPG, Hz stage (c) (d) Finger PPG, Hz stage Forehead PPG, Hz stage (e) (f) Fig. 4. Time-frequency mas of the resiratory rate estimation using the finger (left) and the forehead (right) PPG signal in different stages: a) and b) in Hz stage; c) and d) in Hz stage; e) and f) in Hz stage. ossible that, sometimes, resiratory rate can not be estimated in every time instant because the five romediated sectrum do not fulfill the eakness conditions. In this work this henomena haens in 10 over the 70 total ossible cases (10 subjects and 7 different stages) in the finger signal, with a median time of 33.03% where the estimation can not be done. In the forehead signal, this haens in 9 cases, with a median time of 36.36%. The main result of this article is the fact that when PAV signal is used with the PPG sensor located in the forehead, a different behavior has been found. A comonent (maybe related to the symathetic comonent) aears hiding comletely the exected resiratory rate. This fact contrasts with the behavior observed when the resiratory rate is calculated with the same signal but recorded in the finger. In this case, the resiratory information can be observed as other works suggest [4], [5], allowing to obtain a roer estimation of the resiratory rate. The non resiratory comonent induced by the symathetic modulation was observed in the finger PAV signal too, but it is only a slightly comonent that barely can be areciated. Attending to the erformance results, it is confirmed that a higher mistake is made in the estimation of the resiratory rate when higher rates are recorded, in comarison with the lower ones, as another studies noticed, like in [5] where the error in the resiratory rate extraction using the PAV was lower when the frequency was behind 5 Hz. Finally, the relative ower in each band shows that resiratory information is more relevant when PPG signal is recorded in the finger instead of in the forehead. This exlains the decrease of the accuracy when comaring the same stage for the two PPG signals. As the main result of this study, the aearance of a non ISBN EURASIP

5 resiratory comonent in forehead PAV signal requires an exhaustive investigation to be done with more subjects and more in detail in order to find out why this comonent aears only in this signal. Besides, as the combined methods to extract resiratory information include PAV signal, a robust method to extract resiratory information from the PPG has to be imlemented, no matter what location has the PPG sensor, to avoid ossible wrong estimations. V. CONCLUSION In this work, finger and forehead PPG signals are used to estimate the resiratory rate by means of PAV and validated using a resiratory chest-band as the gold-standard. Results shows that using the finger PAV signal the resiratory rate can be extracted. However, a owerful comonent between and Hz aears when the forehead PAV signal is used, being imossible to estimate the resiratory comonent. This behavior cause a decrease in the ower inside the band centered in the exected resiratory rate when a comarison is done between finger and forehead location, showing that resiratory information is less relevant in forehead sectral ower distribution than in the finger one. These results suggest that forehead PAV is not useful as a signal to extract the resiratory rate. Therefore, an extensive study has to be made in order to clarify why this comonent aears and to establish a more robust method to extract resiratory information from PPG no matter where the sensor is located. ACKNOWLEDGMENT This work has been artially financed by Ministerio de Economía, Industria y Cometitividad (MINECO) and by fondos FEDER through the roject TEC P and TIN R; and by Centro Universitario de la Defensa (CUD) under the rojects CUD and UZCUD2016- TEC-03. This work has never been realized without the collaboration of the volunteers of the consolidated grou BSICoS (Biomedical Signal Interretation & Comutational Simulation). REFERENCES [1] C.W. Seymour, J.M. Khan, C.R. Cooke, T.R. Watkins, S.R. Heckbert, T.D. Tea, Prediction of critical illness during out-of-hosital emergency care, JAMA, 304, , [2] M. Nitzan, A. Babchenko, B. Khanokh, D. Landaud, The variability of the hotolethysmograhic signal - a otential method for the evaluation of the autonomic nervous system, Physiol. Meas., 19 (1), , [3] J. Allen, Photolethysmograhy and its alication in clinical hysiological measurement, Physiol. Meas., 28(3), R1-39, [4] D.J. Meredith, D. Clifton, P. Charlton, J. Brooks, C.W. Pugh, L.Tarassenko, Photolethysmograhic derivation of resiratory rate: a review of relevant hysiology, J Med Eng Technol, 36 (1), 1-7, [5] J. Lázaro, E. Gil, R. Bailón, A. Mincholé, P. Laguna, Deriving resiration from hotolethysmograhic ulse width, Med Biol Eng Comut, 51 (1), , [6] L. Nilsson, T. Goscinski, S. Kalman, L.G. Lindberg, A. Johansson, Combined hotolethysmograhic monitoring of resiration rate and ulse: a comarison between different measurement sites in sontaneously breathing subjects, Acta Anaesthesiologica Scandinavica, 51, , [7] E. Gil, J.M. Vergara, P. Laguna, Detection of decreases in the amlitude fluctuation of ulse hotolethysmograhy signal as indication of obstructive slee anea syndrome in children,biomed Signal Process Control, [8] J. Lázaro, E. Gil, J. M. Vergara, P. Laguna, Pulse Rate Variability Analysis for Discrimination of Slee-Anea-Related Decreases in the Amlitude Fluctuations of PPG Signal in Children, IEEE J Biomed Health Inform, 18 (1), , ISBN EURASIP