Detection of Beat-to-Beat Intervals from Wrist Photoplethysmography in Patients with Sinus Rhythm and Atrial Fibrillation after Surgery

 Abstract — Wrist p hotoplethysmography (PPG) allows unobtrusive monitoring of the heart rate (HR) . PPG is affected by the capillar y b lood perfusion and the pumping function o f the heart, w hich generally de teriorate with age and due t o presence of cardiac arrhythmia. The perfor m ance of w rist PPG in monitoring beat- to -beat HR in older patients with arrhythmia has not been reported earlier. W e m onitored PPG from wrist in 18 patients recovering from surgery in the post anesthesia care unit, and evaluated the inter-beat interval (IBI) detection accuracy against ECG based R - to -R intervals (RRI). Nine subjects had sinus r h ythm (SR , 68.0y ± 10.2y, 6 m ales) and nine subjects had atrial fib rillation (AF , 71.3y ± 7.8y, 4 m ales) during the recording. For the SR group, 99.44% of the be ats were correctly identified, 2. 39 % ex tra beats w ere detected, a nd the mean absolute error (M AE) was 7. 34 ms. For the AF group, 97.49% of the heartbeats were correctly identified , 2. 26% extra beats w ere detected, and the MAE was 14.31 ms. IBI from t he PPG were hence i n close agreement with the ECG refe renc e in both group s . The results suggest that wrist PPG p rovides a comfortable alternative to ECG and ca n be used for lo n g-term monitoring and screening of AF episodes. I. I NTRODUCTI ON Heart rate variability (HRV) provides signif icant inform ation about a pe rson’s health status. It is used for sleep analysis [1], stress and rec overy analy sis [2 ], and als o in clinical applicat ions such as atrial fibril lation (AF) detection [3, 4, 5 ]. Tradition ally, ECG devices have been used in data collecti on for HRV an alysis . The most common are ch est straps and electr ode pat ches, w hich can provid e hig h accu racy for the estim ation of beat- to - beat in tervals [6, 7, 8] when compared to ambulat ory ECG recorders , but can become uncomf ortable when being worn for longer durations. In addition , dry skin o r poor skin contact often disturb ch est strap based HRV m onitoring. Thus, there is a clea r dem and fo r new technologi es, w hich do not interfe re w ith a pers on’s com fort. Photoplethy smogra phy (PPG) provides an alternative method for H R and HRV monitoring [9]. T he skin is illum inated with a LE D and a photodete ctor measures the intensity of the transm itted or reflecte d light. T his intensity depends on the blood volum e in the skin capillaries and the A. Tarniceri u is with PulseOn SA , Jacquet-Droz 1, 2002, N euchâtel, Switzerl and (corresponding e-mail: adrian.tarnice riu@pulseon.com) . J. Harju and A. Yli-Hankala are with the Tampere Uni ve rsity Hospital , Tampere , Finland. A. Vehkaoja, J. Parak, and I. Korhonen are with BioMediTech Institute and the Faculty of Bi ome dical Sciences and Enginee ring, Tampere vasculatu re deeper in the tissue, which, in turn, vary w ith the pumping actions of the heart. Thus, by analy zing the light intensity , we can determine the heart rate and inter-be at intervals (IBI) . Currently , optical heart rate ( OHR ) devices can provide adequate accuracy for heart rate estim ation dur ing r est, s ports, and daily activities [10, 11]. Previous work [12, 13] show ed that u sing th e rig ht algorith ms, IBI can be estimated from w r ist PPG signals with errors below 1 0 ms, which i s accurate enough for HRV analysis. Ho w ever, these r esu lts were obtained using data from healthy workin g age subjects. Elderly people usually have poorer peripher al p e rfusion, different skin structu re, an d ar rhythm ias or o th er illn esses. A ll these factors affect the PPG signal, decreasing the signal- to - noise rati o. This study evaluates the IBI estim ation ac curacy for a group of post-surgery patients, half of w hich su ffer from AF . The main goal is to evaluate wheth er wrist PPG can be used for IBI monitorin g in clini cal applica tions for elderly subject s with arrhythm ia and especially with AF . If proven operational , the wrist PPG technology would provide tremendous benefits in both clinical and hom e monitoring scenarios: it would provide a com fortable, wearable, unobtru sive measur ement method suitable for long-term monitorin g. B esi des life-sty le, sleep, an d stress analysis , it could be also use d in screen ing o f various cardiac an omalies. II. M AT ERIALS AND M E T HODS A. Su bjects All recordings took place in the post -anes thesia care unit of the Tampere University Hospital. The patients had undergone su rgery immediate ly prior t o the recording and were recovering from the effects o f anesthetics. T hey were laying down in bed d ur ing the whole dura tion of the measurem ent. The average d urat ion of each recording is 1.5 hours. 18 patients were included and classifie d in to two groups: with sinus rhythm (SR) and w ith continuous AF during th e recording . The SR group consis t ed of nine subjects - six male, three feman e, 68.0 ± 10.2 years old, and the AF group consis t ed of nin e su bjects - four male, fiv e female , 71.3 ± 7.8 y ears old. University of Technolo gy , Tampere, Finland, and with PulseOn Oy , Espoo, Finland. R. Delgado-Gonz alo a nd P. Renevey are with CSEM - Centre Suisse d’Electronique et Microtechnique, Jacqu et -Droz 1, 2002 Neuchâte l, Switzerl and. Detection of Beat - to -Beat Intervals fr o m W rist Photoplethysmogra phy in Patients with Sinus Rhy thm and Atrial Fibrillation after Surgery Adrian Tarnicer iu, Jarkko Har ju, Antti Vehkaoja, Jak ub Parak, IEEE Student Member , Ricard Delgado- Gonzalo, Philippe Renev ey, Arvi Yli-H ankala, and Ilkka Korhonen, IEE E Senior Member The study protocol, devices, and documentation were approved by the local ethical revi ew board of Pirkanm aa Hospital Dis trict (R1 7024), th e Finnish Nati o nal Supervisory Authority of Health and Welfare, and the technica l department of the hos pital. The test subjects gave their written consent to partici pate after being informed on the purpose of the study and they had the right to w ithdraw from th e study at any tim e. The experim ental procedu res comply with the princip les of the Helsinki Declarati on of 197 5, as revis ed in 20 00. B. Da ta Acquisition Wrist PPG signals were recorded with the PulseOn OHR tracker (ww w.pulseon.f i), pres ented in Fig . 1 . The device w as worn as instruct ed by the manufactur er, about one fing er width from the w r ist bone and tightened b y the person in charg e of data col lection so that the skin co nt act w as firm but stil l comfortab le for the w hole recordin g. F or th e PPG data, the IBI were prov ided by OHR tracke r direct ly. The ECG sig nals w ere measured w ith the GE Health care Carescape B850 ( w ww.gehealth care.com) patient monitor and recorded w ith the S5 Collect softw ar e. The RR intervals w er e obtained from the ECG s ignal using the Kubios HR V softw are, versi on 2.2 (www .kubios.com). T he ECG waveform s were also visually inspected to ensure that no R- waves are m issed. C. Methods As the recording of wrist PPG and ECG signals d i d not start at the same time, w e firstly synchronized the IBI and RRI time series . This wa s done by compensating for eventual time drifts between the PulseOn and Ca rescape B850 c locks and b y min imizing the mean ab s olute error between the IBI and reference RRI vect ors . For final synchronization , w e split the data in in tervals o f one min ute an d pe rform ed a new synchroni zation for each interv al. T h is w as nec essary to allow beat- to -beat level synchronizati on despite slightly differin g nominal clock r at es of the devices. Ectopic beats [14] were excluded f rom the ev aluation. In the next step, for each one-minu te interval, we determ ined the percentage of correctly detecte d beats, extra beats, and missed beats with respect to the E CG reference. This was do n e with a method similar to th e one used in [13]. For every PPG-detec ted beat, w e check ed how m any reference beats we re detected in the inte rval [ 𝑡 − 0.5𝑙 , 𝑡 + 0.5𝑙 ] , w here t is the time wh en the beat was detected and l is the length of the corresponding IBI. If there was only one reference beat, then it was correctly detected . If the re w ere no corres ponding reference beats, then an extr a beat had been detected. The reference b eats with no corr esponding PPG-detected beat we re considere d miss ing beats. Most extra and m issing beats are ex plained by th e f act th at IB I estimati on is not accurate during motion . An example is given in Fig. 2 : motion , depicted as variati o ns in the 3D accelerati o n signa l, cau ses m ore oscilla tions in the PPG-based IB I signa l. This reduces the a ccuracy of beat es timation [15], usually resulting in shorter IBI, as seen between 15 and 35s. These type of artefacts can occur even if the movement is limited to the fingers or hand, and the forearm is immobile, movem ents that are n ot detected by an accelerom eter located in the w rist device. Figure 1. PulseOn OHR tracker place d on the wr ist Figure 2. Effect of motion, depicted as var i a tions in the 3D acceleration signal, on the e stimation of inter-beats from w rist PPG signals. In addition to ext r a detected and missed beats, we compute the mean error (ME), the mean absolute error (MAE), the mean absolute p e rcentage error (MAPE), and the root mean square error (RMSE) for the IBI-RRI pairs. Three HRV paramete rs, we re also computed : the root mean square of successive differenc es (RMSSD) , the percentage of suc cessive IB I that differ by more than 50 ms (pNN5 0), and the IBI standard deviation (ST D), to evaluate their behavior for SR and AF r hythms. As the purpose of this st udy is to estimat e the beat accurac y during rest, and the mis sing or extra beats are a good indicator for the p r esence of motion, we w ill only consider the one-m inute intervals with no missing or extra beats wh en computing the ME, MAE, MAPE, RMSE, RMSSD, pNN 50, and STD. III. R ESULTS AND D ISCUS SION A. Beat Detectio n Performance The beat dete ction resul ts are summarize d in Table I. For the SR set , 99.44% o f the beats we re correc tly detected while for the AF set, 97.49% of the beats were correctly detected. The level o f extra beats is sim ilar betw een the grou ps (2.39% vs 2.26% for SR and AF, corr espondingly ), w hile the AF group has signif icantly m ore missing beats than the SR gro up (2.51% vs 0.56%). Th e low er beat detection rate in the AF group can be expl ained by different puls e m o rphology caused by arrhy thmias. TABLE I . IBI DETECT ION PERFORMANCE SR set AF set Total beats 52726 55565 Correct beats [%] 99.44 97.49 Extra beats [%] 2.39 2.26 Missing beats [%] 0.56 2.51 B. I BI Estimation Figures 3 and 4 illu strate an exam ple of 50 bea ts extract ed from the PP G sign als as w ell as from the ECG reference, and the error between IBI and RRI for SR and AF cases , respective ly. T he diff erence betw ee n SR and AF scenarios is clearly visible fr om these figur es, the AF case h aving a much higher va riation betw een consecutive IB valu es. The MAE and MAPE are approxim ately two-fol d higher for the AF g ro u p as compared to th e SR group (Table I I). Fig . 5 s hows the Bland-Altm an plot s fo r the RRI and IBI. The most likely explanation f or the higher err or in the AF group is that the fiducial point of the pulse wave detection in PPG is dependent o n the pulse morphology , which is widely variant during AF du e t o n on-optimal heart filling and poor pum ping function. However, the MAE for the AF group is still significan tly low er than the dif ference be tween the consecutive beats, as can be seen in Fig. 4, and each ca s e follow s the general t rend of the RR I values extract ed from th e ECG signals . The estim ation error is slig htly b i ased tow ards lower values ( the ME is - 0.40 and -0.47 ms, respectiv ely), m ost likely due to the r ounding t owards ze ro operati ons of th e used fixed-point algorithm . T his error, lo w er than 1 ms, has a negligibl e effect on HRV ana ly sis. For the AF group, there is no visible correlation in the Bland-Altm an plot between the IBI-RRI difference and the values o f the IBI. For the SR group, it looks like the error dispersi on is higher for beats of ~1000 ms. However, this is just a visual effect of the fa ct t hat there are more beats aroun d this value. Fo r 7 out of 9 sets, the average HR is between 55 and 65 beats per minute, correspon ding to IBI between 923 and 1090 ms; but th e error stan dard deviati on is the sam e for beats of ~1000 m s and for beats w ith d if ferent dur ations. C. Heart Rate Variab ility Parameter Comparison Table III presents three HRV p a rameters calculat ed from IBI in SR and AF groups . The HRV param eter s are system atically higher for the AF group, suggesting that they may be used to differentiate AF from SR [17, 18]. Anothe r insight on the usa bility of PP G-de rived inter-beats f or the detecti on of atrial fibrillati on is provided in Fig. 6 . Here, we plot the stan dard d evia tion of groups of 20 consecutiv e IB values. The differenc e between SR and AF cases is clearly visible, and one c ould easily distinguish between the tw o cardiac conditi ons. T h is can be use as the starting point for designing an atrial fibrilla tion detecti on algorith m. TABLE II . IBI E S TIMATION PERFORM ANCE SR set AF set ME [ms] -0.40 -0.47 MAE [ms] 7.34 14.31 MAPE [%] 0.79 1.58 RMSE [ms] 16.70 23.52 Figure 3. Example of IBI and RRI time serie s in a SR case. The low er panel show s the instantaneous erro r between RRI and I B I Figure 4. Example of IBI and RRI time series in an AF c ase . The lower panel show s the instantaneous err or between RRI and IBI Figure 5. B lan d-Altman plots fo r PPG I B intervals, rel ative to the ECG refere nce. Sinus rhythm and atrial fibrillation case s TABLE III . HRV S TATISTICS SR set AF set RMSSD [ms] 36.01 268.34 pNN50 [%] 8.45 83.09 STD [ms] 51.70 211.48 Figure 6. Sta nd ard deviation e xample for g roups of 20 consecutive inter- beat values fo r AF and SR case s IV. C ONCLUSION This study evaluat ed the accuracy of IBI estimati on from wrist PPG signals for elderly patients after surgery with SR and AF . The MAE values are 7.34 ms for the SR group and 14.31 ms for the AF group. This is accurate enough for both HRV analy sis and t o different iate betw een SR and A F cases. Earlier studies have vali dated the estimation of IBI from PPG signals for healthy subjects during slee p [13]. This study validates th e IBI est imation in a more challenging scenario: the subjects are elderly patients with arrhythm ia , and ha ve undergone su rgery prior to the recording. For c omparison, the MAE value obtaine d in [13] is 6.68 ms wh ich is almost identical to MAE o bserved in th is study fo r SR pati ents. In conclusi on, the presen t study confi rms that IBI estimated fr om wrist PPG signals are in clos e agre ement with RRI obtained from the ECG reference. The estimated value s are highly accurate and can be used for both HRV analysis and clinical applic ations such as AF detection . This pro vi des a promisin g altern ative to current monitoring technologies , and an important step tow ards 24/7 m o nito ring. R EFERENCES [1] T. Mylly mäki, H. Rusko, H. Syväo ja, T. Juuti, M.-L. Kinnunen, a nd H. Kyröläinen, “Effects of exercise intensity and duration on nocturnal heart rate variability and sleep quality ,” European Journal of Applied Physiology , vol. 112, pp. 801-809, 2012. [2] First beat Technolog ies Ltd., “Stress and Recovery Analysis Method Based on 24- hour He art Rate Vari ability”, (w hitepaper), 2014. [3] G.B. Moody and R.G. Mark, “A new method for detecting atrial fibrillation us ing RR in tervals,” in Com puters in Car diology, 1983, pp. 227-230. [4] J. L ee, B. A . Reyes, D. D. McManus, O. Maitas and K . H. Ch o n , "Atrial Fibrillation Detection Using an i P h one 4S," in IEEE Transactions on Biomedical E ngineering , vol. 60, no. 1, pp. 203-206, Jan. 2 013. [5] A. G. Bonomi et al ., "Atrial fibrillation det ec t ion using ph o t o- plethysmo graphy and accel eration data at the wrist," 2016 Computing in Cardiology Conference (Ci nC) , Vancouver, BC, 20 16, pp. 277-280. [6] M. Weippe rt, M. Kumar, S. Kreuzfeld, D. Arndt, A. Rieger, and R. Stoll, "Comparison of three mobile d evice s for measuring R – R intervals and heart rate variability: Polar S810i, Suunto t 6 and an ambulatory ECG sy stem ," European Jour nal of Applied Physiology , vol. 109, pp.779-786, 201 4. [7] L. C. Vanderlei, R. A . Silva, C. M .Pastre, F . M. Azevedo, and F. M. Go doy, “Comparison of the Polar S810i monitor and the ECG fo r the analysis of heart rate variability in the time and frequency domains,” Brazilian Journal of Med ical and Biological Rese arch , vol. 4 1, pp. 854 – 859, Octobe r 2008. [8] J. Kristiansen, M. Korshøj. J. H. Skotte, T. Jespersen, K. Søgaard, K. Mortensen, and A. Holtermann, “Comparison of two systems for longterm heart r ate variability monitoring in fre e-living conditions - a pilot study,” BioMedical En gineering OnLi ne , pp. 10 – 27, April 2 011. [9] M. Bertschi, P. Renevey , J. Solà, M. Lemay, J. Parak and I . Korhone n, "Application of optical heart rate monitoring," in Wearable Sensors. Fundamentals, Impleme ntation and Applications , Elsevier Academic Press, 2014, p. 6 56. [10] J. Parak and I. Korhonen, "Evaluatio n of wearable consumer heart rate monitors based on photopletysmography , " in Proc. 36th Annual International Conference of the IEEE Engineerin g in Medicine and Biology Society , Chicago, USA , 2014, pp. 3670-3673. [11] R. Delgado-G onzalo, J. Parak, A . Tarniceriu, P. Renevey, M. Be rtschi, and I. K orhonen, “Evaluation of Accuracy and Reliability of PulseOn Optical Heart Rate Monitoring Device,” in Proc. 37th Annual International Conference of the IEEE Engineerin g in Medicine and Biology Society (EM BC ), Milano, 2015, p p. 430-433. [12] P. Renevey , J. Sola, P. Theurillat , M. Bertschi, J. Krauss, D. Andr ies, and C . Sar tori, "V alidation of a w rist monitor for accurate estimation of RR intervals during sleep," in Proc. 3 5th Annual International Conference of the IEEE Engineeri ng in Medicine and Biology Society (EMBC), Osaka, 2013 pp.5493-5496. [13] J. Parak, A. T arniceriu, P. Reneve y, M. Bertschi, R. De lgado-Gonzalo, and I. Korhinen, “Evaluation of the Beat - to -Be at Detection Accuracy of PulseOn W earable Optical Heart Rate Monitor,” in Proc. 37th Annual Intern ational Confere nce of the IEE E Engineeri ng in Medici ne and Biology S ociety (EMBC) , Milano, 2016, pp. 80 99-8102. [14] J. Mateo and P. L aguna, ”A nalysis of heart rate v ariability in the presence of e ctopic beats using the h eart timing signal,” IEEE Transactions on B i omed ical Engineerin g, vol . 50, pp.334-3 43, 2003. [15] A. Schäfer and J. Vagedes, “How accurate is pulse r ate varia bility as an estimate of heart rate variability? A review on studies comparing photoplethy smographic technology with an electro cardiogram”, International J ournal of Cardiol ogy , vol. 166, pp. 15 – 29, Jun 2013. [16] U. Ra jen dra Achary a, K. Paul Joseph, N. Kannathal, Ch oo Min Lim, and Jasjit S. Suri, “Heart rate variabil ity: a review.” Medical & Biological Eng ineering & Com puting, vol. 44, pp. 1031-1 051, 2006. [17] B. Logan and J. Healey, "Robust detection of atrial fibrillation for a long term telemo nitoring system," Computers in Cardiolo gy, 2005 , Lyon, 2005, p p. 619-622. [18] J. L ee, B. A . Reye s, D. D. M c Manus, O. Mait as and K . H. Chon, "A tria l Fibrillation Detection Using an i P h one 4S," in IEEE Transactions on Biomedical Engine ering , vol. 60, no. 1, pp. 203-206, Jan. 2013.