AN ALGORITHM FOR THE DETECTION OF ATRIAL

FIBRILLATION USING THE PULSE OXIMETRIC SIGNAL

Giovanni Calcagnini, Michele Triventi, Federica Censi, Eugenio Mattei, Pietro Bartolini

Dept. of Technology and Health, Italian Institute of Health, Viale Regina Elena 299, Rome, Italy

Francesco Mele

Department of Cardiology, S. Filippo Neri Hospital, Rome, Italy

Keywords: Atrial fibrillation, Pulse oximetry, Rhythm classification.

Abstract: A method for the discrimination of atrial fibrillation and sinus rhythm from the pulse oximetric signal is

presented. The method is based on the analysis of the ventricular rhythm irregularity, quantified by the

Coefficient of Variation and the Shannon Entropy of the ventricular inter-beat intervals. A classifier based

on the Mahalanobis distance is then applied. Sixty patients with an history of recurrent atrial fibrillation

were studied. The method yielded a correct classification of 43 out of 43 patients with sinus rhythm, 14 out

of 14 patients with atrial fibrillation, and 3 out of 4 patients with other arrhythmias.

1 INTRODUCTION

Atrial fibrillation (AF) is the most common cardiac

arrhythmia in western countries. The current

prevalence of nontransient AF in the US is 4% in

the population of 65 to 70 years of age, and of 10%

for people ≥ 80 years of age and is projected to

increase considerably by 2050 (Naccarelli et al,

2009). AF is an independent risk factor for death

and a major cause of stroke (Go et al, 2001). There

are evidences that AF sustains itself through a

complex process that is initiated by high atrial rate,

cytosolic calcium overload, metabolic depletion and

contractile dysfunction. Conversion of AF to sinus

rhythm by antiarrhythmic drugs is relatively

effective when AF duration is short (Kirchhof et al,

2009), whereas when AF duration exceeds two

weeks the efficacy is greatly diminished.

These evidences suggest that early diagnosis is a

key element to prevent the progression of AF and

reduce atrial fibrillation-related complications.

Another significant implication of asymptomatic AF

is related to the need for oral anticoagulation.

Withdrawal of oral anticoagulation after therapeutic

interventions (e.g. electrical or pharmacological

cardioversion, radiofrequency ablation) should be

considered carefully, based on reliable and objective

measures rather than symptoms.

Currently, diagnosis of AF is based on the

analysis of the ECG signal. Due to the poor

correlation between symptoms and AF (Israel 2004;

Rho et al, 2005) the rate of detection of AF episode

are strongly affected by the intensity of monitoring.

Arya et al, reviewed the various ECG-based follow-

up strategies to detect AF recurrencies after

radiofrequency ablation and estimated that

conventional Holter electrocardiogram (ECG)

recordings have a low diagnostic yield for

paroxysmal AF, newer technologies like patient-

operated or telemetric ECG systems, long-term

Holter monitors, or even implanted ECG monitors

carry the promise of allowing an early diagnosis of

silent AF.

Automatic detection of AF is achieved by

analysis of the electrocardiographic signal. The

absence of the P-waves is the main criterium for AF

detection. Alternative methods have been proposed.

These methods are based on the measure of the

irregularity of the ventricular rhythm. Various

measures of such irregularity are known. These

measures quantify the variability of the ventricular

inter-beat intervals (RR interval) obtained from the

ECG signals, using combinations of various

features: standard deviations and probability density

429

Calcagnini G., Triventi M., Censi F., Mattei E., Bartolini P. and Mele F..

AN ALGORITHM FOR THE DETECTION OF ATRIAL FIBRILLATION USING THE PULSE OXIMETRIC SIGNAL .

DOI: 10.5220/0003150804290432

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2011), pages 429-432

ISBN: 978-989-8425-35-5

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

function (Tateno and Glass, 2001), wavelet

transform (Duverney et al, 2002) entropy, Lorenz

plots (Esperer et al, 2008), probability density

function of an embedded time series (Hong-Wei et

al, 2009), turning point ratio, standard deviation and

entropy (Dash et al, 2009), Markov modelling in

combination with P-wave analysis (Babaeizadeh et

al, 2009), Poincarè plots (Park et al 2009).

Some of these parameters does not suit a short-

term detection since they requires a relatively large

number of beats, others require significant

computational effort / memory occupation. In this

study, such methods have been excluded.

PP interval, e.g, the ventricular inter-beat interval

measured from the pulse oximetric wave, has been

proposed as an alternative to RR interval, during

normal sinus rhythm (Lu et al, 2008; Foo et al,

2006). The reliability of ventricular rhythm

estimation from PP intervals during AF is not

known.

2 METHODS AND MATERIAL

2.1 AF Detection Algorithm

The detection of an atrial fibrillation episode is

based on the extraction of quantitative indexes from

the PP and ΔPP time series.

To distinguish between SR and AF we use the

entropy (EN) of the PP interval series and the

coefficient of variation (CV) of the ΔPP intervals.

The entropy is estimated as follows:

∑



iiPP

ppEN

log

(1)

where p is the estimated probability density function

of the PP series.

Since the mean of the ΔPP sequence leads to zero,

we calculated the CV by dividing the standard

deviation of the ΔPP intervals by the mean of the PP

sequence

(2)

To implement an automatic decision criterion,

based on the CV and En, we used the Mahalanobis

distance, which takes into account the covariance

among the variables in calculating distances.

Mahalanobis distance (D

) of a multivariate

vector x from a group of values with mean μ and

covariance matrix S is defined as:

)()()(





xSxxD

(3)

In order to have a parameter to discriminate AF

vs. SR patient, Mahalanobis distance from AF and

SR population was calculated for each patient.

The mean values of CV and EN were calculated for

AF and SR patients, and the two covariance matrices

were obtained:







=

0.0086 0.0076



(4)







=

0.0022 0.0054

0.0054 0.2129



(5)

For the i

patient the Mahalanobis distances from

the two groups are obtained as







() =







−







−

















−







−



 (6)







() =







−







−

















−







−



 (7)

We classified the patient as belonging to the

group for which the Mahalanobis distance is

minimal and it is below a given threshold. In the

case the distances from both groups are greater than

the respective thresholds, the rhythm is classified as

“other arrhythmia”. In this study the squared

thresholds were set at 10, on empirical basis.

2.2 Clinical Validation Protocol

The study was conducted at the Atrial Fibrillation

Unit of S. Filippo Neri Hospital, in Rome. We

studied 60 patients undergoing standard 12-lead

ECG exam for a history/suspect of AF. Heart rhythm

at the time of the examination was classified by an

expert cardiologist as AF, SR or other arrhythmia.

Then, a 5-minute pulse oximetric signal was

acquired from the index of the non-dominant hand

using a MIROXY device (Medical International

Research, Italy). The device firmware was modified

to allow the real-time transmission of the pulse

signal to a PC, using the RS-232 connection.

A single ECG lead was also recorded and

digitized using a National Instrument NI-USB6218

DAQ card, for a further confirmation of the actual

patient rhythm. Patients with pacemaker and/or

defibrillator were excluded from the study.

3 RESULTS

Table 1 shows the characteristics of the patients and

the heart rhythm at the moment of the test, as

classified by the cardiologist from the ECG trace.

BIOSIGNALS 2011 - International Conference on Bio-inspired Systems and Signal Processing

430

Table 1: Characteristics of patients’ population.

Rhythm

N A

ge (mean+/- sd, range) Sex (M/F)

SR 43 65.27 +/- 11.96, 21-87 26/17

AF 13 78.14 +/- 8.29, 67-89 7/6

Other 4 67.75 +/- 10.51, 61-73 4/0

The results of the automated classification from

the ventricular rate irregularity obtained by the pulse

oximetric waveforms are reported in the Tables 2,3,4

for AF, SR and OTHER patients, respectively. The

Tables also report the CV and EN values for each

patient, as well as the distance obtained using the

Mahalanobis metrics. The mean values of CV and

EN of each group are also reported.

Table 2: Classification results of AF group.

Rhythm

/Pt CV

'PP

EN D

SR D

Classification

AF1 0.368 3.697 50.01 1.39 AF

AF2 0.296 3.492 29.48 0.15 AF

AF3 0.254 3.384 20.34 1.62 AF

AF4 0.321 3.496 35.93 0.79 AF

AF5 0.312 3.491 33.53 0.43 AF

AF6 0.365 3.668 49.12 1.20 AF

AF7 0.331 3.543 38.84 0.50 AF

AF8 0.379 3.626 53.92 3.08 AF

AF9 0.350 3.769 44.68 3.71 AF

AF10 0.283 3.566 26.75 2.08 AF

AF11 0.218 3.235 13.78 5.06 AF

AF12 0.282 3.603 26.77 3.66 AF

AF13 0.315 3.503 34.41 0.37 AF

mean 0.313 3.544 35.198 1.851

Figure 4 gives a pictorial representation of the

population EN and CV (circles), as well as of the

classification results (crosses).

Table 3: Classification results of SR group.

Rhythm

/Pt

'PP

EN D

SR D

Classification

SR1 0.030 2.019 0.78 154.39 SR

SR2 0.041 2.528 2.35 54.08

SR3 0.055 1.958 0.04 185.75

SR4 0.062 1.349 2.01 457.45

SR5 0.038 2.320 1.38 85.67

SR6 0.019 1.926 1.19 176.53

SR7 0.052 1.953 0.07 185.54

SR8 0.019 1.835 1.07 206.13

SR9 0.060 2.376 0.80 80.05

SR10 0.021 1.612 1.05 292.12

SR11 0.095 1.159 5.52 610.45

SR12 0.093 1.935 0.72 218.11

SR13 0.061 2.085 0.06 149.66

SR14 0.165 2.346 5.84 129.77

SR15

0.085 1.146 4.83 604.87

Table 3: Classification results of SR group.(cont.)

SR16 0.082 1.866 0.46 236.45

SR17 0.024 1.683 0.86 265.31

SR18 0.068 2.334 0.51 90.34

SR19 0.016 1.091 3.25 550.87

SR20 0.057 2.131 0.16 135.21

SR21 0.044 1.783 0.24 240.97

SR22 0.071 2.035 0.03 169.90

SR23 0.170 2.994 7.04 16.00

SR24 0.023 1.730 0.87 246.39

SR25 0.059 0.875 5.93 749.36

SR26 0.026 1.904 0.78 187.41

SR27 0.060 2.927 4.46 28.99

SR28 0.082 2.184 0.23 131.65

SR29 0.019 1.857 1.08 198.79

SR30 0.109 2.858 3.03 24.64

SR31 0.122 2.378 1.92 100.00

SR32 0.020 1.442 1.50 366.75

SR33 0.011 1.647 1.47 270.47

SR34 0.061 2.071 0.04 153.98

SR35 0.019 1.463 1.47 356.08

SR36 0.084 2.232 0.32 120.02

SR37 0.165 2.294 5.87 144.92

SR38 0.157 2.738 5.04 41.58

SR39 0.092 1.891 0.77 233.82

SR40 0.056 2.952 4.96 30.34

SR41 0.065 1.926 0.03 202.50

SR42 0.028 2.361 2.21 76.53

SR43 0.019 1.409 1.62 382.73

mean 0.063 1.991 1.951 217.269

Table 4: Classification results of OTHER group.

Rhythm

/Pt

'PP

EN D

SR D

Classification

OTHER1 0.167 1.523 11.13 486.25 OTHER

OTHER2 0.309 3.207 32.72 18.45

OTHER

OTHER3 0.232 0.942 34.30 998.65

OTHER

OTHER4 0.148 2.695

4.18

45.92 SR

The proposed method yielded a correct classification

of all the patients with AF (13/13), as well as of all

the patients in SR (43/43). One patient of the

OTHER group, who had a low frequency atrial

flutter, was misclassified as normal sinus rhythm,

because he had a Mahalanobis distance from the SR

group below the threshold (see table 4).

AN ALGORITHM FOR THE DETECTION OF ATRIAL FIBRILLATION USING THE PULSE OXIMETRIC SIGNAL

431

Figure 1: Result of the rhythm classification.

4 CONCLUSIONS

In this work, a AF detection algorithm based on the

pulse oximeter signal is proposed. The algorithm is

based on the measure of the irregularity of the

ventricular rate during AF. The experimental

validation demonstrated both high sensitivity and

high specificity in AF and SR discrimination, so the

algorithm can precisely detect AF episodes from a

pulse oximeter device.

The high sensitivity of the algorithm, the

relatively short data required (5 minutes), and its

implementation on a microcontroller suggest that it

is possible to design an home-care device for the

accurate detection of AF episodes, based on

commercial pulse oximeters.

ACKNOWLEDGEMENTS

This research was funded by the FILAS - Regione

Lazio Grant. Authors wish to thank Engg. Boschetti,

Dieli and Pennacchietti from Medical International

Research, for providing the device for the data

collection and for the assistance in the algorithm

implementation.

REFERENCES

Naccarelli, GV., Varker, H., Lin, J., Schulman, KL., 2009.

Increasing prevalence of atrial fibrillation and flutter

in the United States. Am J Cardiol. Dec

1;104(11):1534-9.

Go, A. S., Hylek, E. M., Phillips, K. A., et al., 2001.

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.

May 9;285(18):2370-5.

Kirchhof, P., 2009. Can we improve outcomes in A. F.

patients by early therapy? BMC Med. Nov 26;7:72.

Israel, C. W., 2004. Is there a role for pacing in the

prevention of atrial tachyarrhythmias? Europace.

6(5):380-3.

Rho, R. W., Page, R. L., 2005. Asymptomatic atrial

fibrillation. Prog Cardiovasc Dis.48(2):79-87.

Arya, A., Piorkowski, C., Sommer, P., Kottkamp,

H., Hindricks, G., 2007. Clinical implications of

various follow up strategies after catheter ablation of

atrial fibrillation. Pacing Clin Electrophysiol.

Apr;30(4):458-62.

Tateno, K., Glass, L., 2001. Automatic detection of atrial

fibrillation using the coefficient of variation and

density histograms of RR and delta RR intervals. Med

Biol Eng Comput. Nov;39(6):664-71.

Duverney, D., Gaspoz, J. M., Pichot, V., et al., 2002. High

accuracy of automatic detection of atrial fibrillation

using wavelet transform of heart rate intervals. Pacing

Clin Electrophysiol. 25(4 Pt 1):457-62.

Esperer, H. D., Esperer, C., Cohen, R. J., 2008. Cardiac

Arrhythmias Imprint Specific Signatures on Lorenz

Plots. Annals of Noninvasive Electrocardiology

13(1):44–60.

Hong-Wei, L., Ying, S., Min, L., Pi-Ding, L., Zheng, Z.,

2009. A probability density function method for

detecting atrial fibrillation using R-R intervals. Med

Eng Phys. 31(1):116-23.

Dash, S., Chon, K. H., Lu, S., Raeder, E. A., 2009.

Automatic real time detection of atrial fibrillation. Ann

Biomed Eng. Sep;37(9):1701-9.

Babaeizadeh, S., Gregg, R. E., Helfenbein, E. D.,

Lindauer, J. M., Zhou, S. H., 2009. Improvements in

atrial fibrillation detection for real-time monitoring. J

Electrocardiol. 42(6):522-6.

Park, J., Lee, S., Jeon, M., 2009. Atrial fibrillation

detection by heart rate variability in Poincare plot.

BioMedical Engineering OnLine 8:38.

Lu, S., Zhao, H., Ju, K., et al., 2008. Can

photoplethysmography variability serve as an

alternative approach to obtain heart rate variability

information? J Clin Monit Comput. 22(1):23-9.

Foo, J. Y., Wilson, S. J., 2006. Detection method to

minimize variability in photoplethysmographic signals

for timing-related measurement. J Med Eng Technol.

30(2):93-6.

0.25 0.3 0.350.05 0.1 0.15 0.2 0.40

O = SR

O = AF

O = OTHER

4.0

2.0

1.0

1.5

3.0

3.5

0.5

BIOSIGNALS 2011 - International Conference on Bio-inspired Systems and Signal Processing

432