International Journal of Bioelectromagnetism
`Vol. 15, No. 1, pp. 26 - 29, 2013
`
`
`
`www.ijbem.org
`
`
`
`
`
`
`
`
`
`
`
`
`
`Atrial Fibrillation Detection using a Smart Phone
`
`Jinseok Leea, Bersain A. Reyesa, Oscar Mathiasb, David D. McManusb, and
`Ki H. Chona
`aDepartment of Biomedical Engineering, Worcester Polytechnic Institute,
`Worcester, Massachusetts, U.S.A.
`bDepartment of Medicine, University of Massachusetts, Worcester, Massachusetts, U.S.A.
`
`Correspondence: Ki H. Chon, Department of BME, WPI, Worcester, MA, USA.
`E-mail: kichon@wpi.edu, phone 508-831-4114, fax 508-831-6646
`Abstract. We hypothesized that an iPhone 4s can be used to detect atrial fibrillation (AF) based on its
`ability to record a pulsatile PPG signal from a fingertip using the built-in camera lens. To investigate
`the capability of the iPhone 4S for AF detection, 25 prospective subjects with AF pre- and post-
`electrical cardioversion were recruited. Using an iPhone 4s, we collected 2-minute pulsatile time series.
`We investigated 3 statistical methods consisting of the Root Mean Square of Successive Differences
`(RMSSD), the Shannon entropy (ShE) and the Sample entropy (SampE), which have been shown to be
`useful tools for AF assessment. The beat-to-beat accuracy for RMSSD, ShE and SampE was found to
`be 0.9844, 0.8494 and 0.9552, respectively. It should be recognized that for clinical applications, the
`most relevant objective is to detect the presence of AF in the data. Using this criterion, we achieved a
`sensitivity of 100% for iPhone data.
`
`Keywords: atrial fibrillation, smartphone, iPhone, sample entropy, Shannon entropy.
`
`1. Introduction
`Atrial fibrillation is the most common sustained arrhythmia. Over 3 million Americans are currently
`diagnosed, and the prevalence of AF is increasing with the aging of the U.S. population (Go et al.,
`2001). Through its association with increased risk for heart failure, stroke and mortality, AF has a
`profound impact on the longevity and quality of life of a growing number of people (Hajjar and
`Kotchen, 2003; Tsang et al., 2003). Although new AF treatment strategies have emerged over the last
`decade, a major challenge facing clinicians and researchers is the paroxysmal, often short-lived, and
`sometimes asymptomatic nature of AF. Our current inability to diagnose AF in minimally symptomatic
`patients with paroxysmal AF has important clinical implications, since even brief episodes of
`asymptomatic AF are associated with increased risk for stroke, heart failure, hospitalization, and death.
`Moreover, the treatment of patients with disabling symptoms from AF, including shortness of breath,
`syncope, and exertion intolerance, is often impeded by delays in diagnosis. Although the population
`burden of known AF is substantial, [(Humphries et al., 2001) studies have shown that more frequent
`monitoring can improve AF detection (Benjamin et al., 1998). There is therefore a pressing need to
`develop methods for accurate AF detection and monitoring in order to improve patient care and reduce
`healthcare costs associated with treating complications from AF. Such a method would have important
`clinical and research applications for AF screening as well as in assessing treatment response (e.g. after
`cardioversion or AF ablation) and need for anticoagulation. For these reasons, the importance of
`developing new AF detection technologies was emphasized by a recent National Institute of Health
`Heart Lung & Blood Institute Expert panel (Benjamin et al., 2009).
`In our work, we developed a smartphone application to measure pulsatile time series and then use
`this data to detect AF real-time. We have recently successfully demonstrated that using a smart phone’s
`camera to image a finger tip pressed to it will yield pulsatile signals that are similar to heart rate
`fluctuations (Scully et al., 2012). In addition, the use of pulsatile signals from smartphones has recently
`attracted the attention of many researchers (Gregoski et al., 2012; Grimaldi et al., 2011; Jonathan and
`Leahy, 2010; Scully et al., 2012). Note that the approach does not require the need for additional
`hardware as the optical video monitoring of the skin with a standard digital camera contains sufficient
`
`
`
`26
`
`APPLE 1011
`
`1
`
`

`

`information related to variability in the heart rate signal, and it consequently provides accurate heart
`rate time series. The only requirement is that the camera’s illumination and optical sensor be within
`finger tip range of each other.
`In this paper, we introduce the feasibility of AF detection on an iPhone 4s. Specifically, we
`developed a comprehensive iPhone application for collection of pulsatile time series followed by real-
`time detection of AF using the following three statistical methods: RMSSD, ShE and SampE. We
`evaluated the AF detection performance with an iPhone 4s on 25 AF subjects undergoing electrical
`cardioversion.
`
`2. Material and Methods
`2.1. AF Databases and Clinical Data Collection
`For the iPhone 4s data collection, 25 patients with AF who presented for electrical cardioversion to
`the University of Massachusetts Medical Center (UMMC) cardiac electrophysiology laboratory were
`recruited by trained study personnel (McManus, Mathias). 20 men and 5 women with an average age of
`57.95 ± 13.64 years were recruited. Data collection was performed before and after electrical
`cardioversion. Our protocol for data collection was approved by the Institutional Review Boards of
`University of Massachusetts Medical Center (UMMC) and Worcester Polytechnic Institute (WPI). The
`camera of an iPhone 4s was placed on either the index or middle finger of study participants for 2
`minutes prior to, and immediately after, cardioversion. Data were recorded with patients in the supine
`position with spontaneous breathing. Fig. 1 shows an iPhone 4s prototype for AF detection.
`
`
`Figure 1. An iPhone 4s prototype for AF detection.
`
`
`
`2.2 Preprocessing
`For the pulsatile signal acquisition, the iPhone 4s videos were recorded, and the signal was obtained
`by averaging 50x50 pixels of the green band for every frame (Grimaldi et al., 2011; Scully et al., 2012).
`The sampling rate for iPhone 4s was 30 frames/sec. However, in rare cases, the sampling rate was
`slightly lower (e.g. ~25 Hz due to internal processing load). Due to the frame rate variability, we
`interpolated the pulsatile signal to 30 Hz using a cubic spline algorithm followed by peak detection.
`The peak detection algorithm incorporated a filter bank with variable cutoff frequencies, spectral
`estimates of the heart rate, rank-order nonlinear filters and decision logic (Aboy et al., 2005).
`
`2.3 Statistical Approach for AF Detection
`RMSSD
`The RMSSD is used to quantify beat-to-beat variability. Since AF exhibits higher variability than
`NSR, the RMSSD is expected to be higher than those of NSR RR time series. As subjects have different
`mean heart rates, we normalize by dividing the RMSSD by the mean value of the RR time series.
`
`Shannon Entropy
`The second component of the AF detection algorithm is Shannon entropy (ShE). The ShE provides a
`quantitative measure of uncertainty for a random variable. For example, a random white noise signal is
`expected to have the highest ShE value due to maximum uncertainty in predicting the patterns of the
`signal.
`
`
`
`
`27
`
`2
`
`

`

`Sample Entropy
`The third component of the AF detection algorithm is the Sample entropy (SampE). The SampE is
`the negative natural logarithm of an estimate of the conditional probability that a subseries that match
`pointwise within a tolerance r also match at the next point, where self-matches are not included in
`calculating the probability. A high value of SampE indicates low similarity in the time series while a
`low value of Sample entropy indicates high similarity. Thus, the SampE is a useful tool to assess
`randomness of RR time series.
`2.4 Performance Evaluation
`The condition for AF detection is based on each threshold value of THRM, THSE and THSA as
`If RMSSD/mean ≥ THRM then it is AF (RMSSD)
`•
`If ShE ≥ THSE then it is AF (ShE)
`•
`If SampE ≥ THSA then it is AF (SampE)
`•
`For each parameter set, we found the number of True Positives (TP , True Negatives (TN , False
`Positives (FP and False Negative (FN) from the MIT-BIH AF and NSR databases. Subsequently, we
`calculated
`the
`sensitivity
`TP/(TP+FN),
`specificity
`TN/(TN+FP)
`and
`accuracy
`(TP+TN)/(TP+TN+FP+FN). For each statistical method, we found the threshold values providing the
`best (largest) area under the ROC curve. In addition, statistical testing using an ANOVA on ranks was
`done to see if there were significant differences among each dataset.
`
`3. Results
`For subjects in AF, we found that the beat-by-beat accuracy for each algorithm was 0.9844, 0.8494
`and 0.9552, for RMSSD, SE and SampE, respectively. TABLE I summarizes overall sensitivity,
`specificity, and accuracy for each algorithm by database. For clinical applications, the relevant
`objective is to detect the presence of AF episodes from a given dataset. With this criterion, the AF and
`NSR detection accuracy was 100% for all 3 methods. Fig. 2 shows statistical value distribution of a)
`RMSSD/mean, b) ShE and c) SampE for AF subjects pre- and post-cardioversion using an iPhone 4s.
`We found statistically significant differences (p<0.01) between iPhone AF vs. iPhone NSR.
`
`0.4
`
`RMSSD/Mean
`
`(a)
`
`Shannon Entropy
`
`1
`
`(b)
`
`3
`
`(c)
`
`Sample Entropy
`
`0.3
`
`0.2
`
`0.1
`
`0.8
`
`0.6
`
`0.4
`
`0.2
`
`2
`
`1
`
`0
`
`0
`
`0
`
`Post-
`Pre-
`Post-
`Pre-
`Post-
`Pre-
`cardioversion
`cardioversion
`cardioversion
`cardioversion
`cardioversion
`cardioversion
`
`Figure 2. statistical value distribution of a) RMSSD/mean, b) ShE and c) SampE for AF subjects pre- and post-
`cardioversion using an iPhone 4s.
`
`Table 1. Beat-by-beat analysis of sensitivity, specificity and accuracy based on each statistical method on 25 af subjects pre- and
`post- electrical cardioversion
`
`
`Rmssd/ mean
`
`Shannon entropy
`
`Sample entropy
`
`Sensitivity
`
`Specificity
`
`Accuracy
`
`Sensitivity
`
`Specificity
`
`Accuracy
`
`Sensitivity
`
`Specificity
`
`Accuracy
`
`0.9763
`
`0.9961
`
`0.9844
`
`0. 7461
`
`1.0000
`
`0.8494
`
`0. 9258
`
`0.9980
`
`0.9552
`
`
`
`
`
`28
`
`AF subjects
`( iphone 4s)
`
`
`
`
`
`3
`
`

`

`4. Discussion
`In this paper, we show that AF can be accurately detected from pulsatile signals in the human fingertip
`using the camera of an iPhone 4s. The computation time including the processing stage was
`approximately 25 ms for each 64-beat segment on the iPhone 4s. Currently, clinical AF monitoring is
`cumbersome and/or expensive. Given the high prevalence of diagnosed paroxysmal and asymptomatic
`AF, as well as the increasing number of individuals at-risk for this potentially life-threatening
`arrhythmia, better and more readily available AF detection technology is needed. Given the ever-
`growing popularity of cell phones and smartphones, a smartphone-based AF detection application
`provides patients and their caregivers with access to an inexpensive and easy-to-use monitor for AF
`outside of the traditional health care establishment. Because the application does not involve a separate
`ECG sensor and instead employs built-in hardware, it is both novel and cost-effective. We believe this
`package will lead to better acceptance and more widespread use than existing out-of-hospital
`arrhythmia monitors. Further data are needed to explore the acceptability and feasibility of smartphone-
`based AF detection applications in older, at-risk populations.
`
`References
`Aboy, M., McNames, J., Thong, T., Tsunami, D., Ellenby, M.S., and Goldstein, B. An automatic beat detection algorithm for
`pressure signals. IEEE Trans Biomed Eng 52:1662-1670, 2005.
`Benjamin, E.J., Chen, P.S., Bild, D.E., Mascette, A.M., Albert, C.M., Alonso, A., Calkins, H., Connolly, S.J., Curtis, A.B.,
`Darbar, D., et al. Prevention of atrial fibrillation: report from a national heart, lung, and blood institute workshop. Circulation
`119:606-618, 2009
`Benjamin, E.J., Wolf, P.A., D'Agostino, R.B., Silbershatz, H., Kannel, W.B., and Levy, D. Impact of atrial fibrillation on the risk
`of death: the Framingham Heart Study. Circulation 98:946-952, 1998.
`Go, A.S., Hylek, E.M., Phillips, K.A., Chang, Y., Henault, L.E., Selby, J.V., and Singer, D.E. Prevalence of diagnosed atrial
`fibrillation in adults: national implications for rhythm management and stroke prevention: the AnTicoagulation and Risk Factors
`in Atrial Fibrillation (ATRIA) Study. JAMA : the journal of the American Medical Association 285:2370-2375, 2001
`Gregoski, M.J., Mueller, M., Vertegel, A., Shaporev, A., Jackson, B.B., Frenzel, R.M., Sprehn, S.M., and Treiber, F.A.
`Development and validation of a smartphone heart rate acquisition application for health promotion and wellness telehealth
`applications. Int J Telemed Appl, 696324, 2012.
`Grimaldi, D., Kurylyak, Y., Lamonaca, F., and Nastro, A. Photoplethysmography Detection by Smartphone’s Videocamera. In
`The 6th IEEE International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and
`Applications (Prague, Czech Republic), 2011.
`Hajjar, I., and Kotchen, T.A. Trends in prevalence, awareness, treatment, and control of hypertension in the United States, 1988-
`2000. JAMA 290, 199-206, 2003
`Humphries, K.H., Kerr, C.R., Connolly, S.J., Klein, G., Boone, J.A., Green, M., Sheldon, R., Talajic, M., Dorian, P., and
`Newman, D. New-onset atrial fibrillation: sex differences in presentation, treatment, and outcome. Circulation 103:2365-2370,
`2001.
`Jonathan, E., and Leahy, M. Investigating a smartphone imaging unit for photoplethysmography. Physiol Meas 31:N79-83,
`2010.
`Scully, C.G., Lee, J., Meyer, J., Gorbach, A.M., Granquist-Fraser, D., Mendelson, Y., and Chon, K.H. Physiological parameter
`monitoring from optical recordings with a mobile phone. IEEE Trans Biomed Eng 59:303-306, 2012.
`Tsang, T.S., Petty, G.W., Barnes, M.E., O'Fallon, W.M., Bailey, K.R., Wiebers, D.O., Sicks, J.D., Christianson, T.J., Seward,
`J.B., and Gersh, B.J. The prevalence of atrial fibrillation in incident stroke cases and matched population controls in Rochester,
`Minnesota: changes over three decades. J Am Coll Cardiol 42:93-100, 2003.
`
`
`
`
`
`29
`
`4
`
`