1. • To develop a classification method used to
distinguish cardiac arrhythmias from normal sinus
rhythm.
• Fractal properties of IBI can be used to classify
cardiac data and aid in diagnosing heart conditions.
Multifractal Detrended Fluctuation Analysis (MFDFA)
• RR interval time series is segmented
into a moving window at different time scales to
create n partitions.
• Time- and scale-dependent root-mean-square
(RMS) values are found using a linear fit applied at
each window and partition.
• Local scaling exponents (Holder exponents) were
determined by finding the slope of a log-log plot
of the RMS values and scale at each time point.
Symbolic Representation (Bag-Of-Words)
• From the newly acquired Holder time series, each
point is assigned to a letter defined by its location
on the series’ probability density function. [5]
• Frequency of occurrence is recorded for each
permutation of letter vectors (words) of specified
lengths.
• It has been shown that signal complexity
can be used as an indicator for
disease and aging. [1]
• Cardiac inter-beat interval fluctuations (IBI)
signal loses its complexity when the heart is in a
diseased state, such as cardiac arrhythmias. [2]
• One measure of complexity is through
fractal properties.
• Fractals are patterns that exhibit self similarity
over different time scales.
• Simple fractal signals can be described by a single
scaling exponent (monofractal). Complex signals
exhibit a spectrum of exponents (multifractal).
[1] A. Goldberger, L. Amaral, J. Hausdorff, P. Ivanov, C. Peng and H. Stanley, "Fractal dynamics in physiology: Alterations with disease
and aging“, Proceedings of the National Academy of Sciences, vol. 99, no. 1, pp. 2466-2472, 2002.
[2] M. Saeed, "Fractals Analysis of Cardiac Arrhythmias", The Scientific World JOURNAL, vol. 5, pp. 691-701, 2005.
[3] A. Goldberger, "Non-linear dynamics for clinicians: chaos theory, fractals, and complexity at the bedside", The Lancet, vol. 347,
no. 9011, pp. 1312-1314, 1996.
[4] E. A. F. Ilhen, "Introduction to Muiltifractal Detrended Fluctuation Analysis in Matlab“, Frontiers in Physiology, vol. 3, 2012.
[5] J. Lin and Y. Li, "Finding Structural Similarity in Time Series Using Bag-of-Patterns Representation," Springer-Verlag, Berlin, 2009.
Classification of Cardiac Data based on
Multifractal Feature Extraction
Prateek Mathur1,2, Hamidreza Saghir MSc1,3, Tom Chau, PhD, P. Eng1,3,4
1Bloorview Research Institute, Holland Bloorview Kids Rehabilitation Hospital; 2Department of Electrical and Computer Engineering, McMaster University;
3Institute of Biomaterials and Biomedical Engineering, University of Toronto; 4Toronto Rehabilitation Institute
• Special thanks to the Ward Family Summer Student Research
Program, University of Toronto IBBME USRP, and NSERC USRA
for funding and facilitating this summer’s research.
• Thanks to Dr. Tom Chau, Hamidreza Saghir, Ka Lun Tam, and
the rest of the PRISM Lab for their insight and advice.
Background
References
• Results show MFDFA combined with symbolic
representation is a viable method to classify
cardiac data.
• These techniques have potential as alternative
diagnosis methods for cardiac conditions, and can
be used to assess a patient’s risk of cardiac event.
• Next steps involve developing classification
methods for multiple types of arrhythmias and
other cardiac conditions, allowing for more
targeted treatment.
Conclusions/Future Work
Methods
Objectives
• ECG recordings of 48 patients from MIT-BIH
Arrhythmia database and 18 patients from
MIT-BIH Normal Sinus Rhythm database.
[3]
d
c
b
a
• 3-6 letter alphabets were tested and classified
for word lengths of 2-6 letters.
• For each word length tested, a table of word
occurrences of each permutation is fed into a
binary linear classifier.
• Using 10-fold cross validation and word counts as
features, signals are classified as either normal
sinus rhythm or arrhythmia.
• Highest average classification accuracy was
obtained with a 6-letter alphabet (91.5%).
Classification/Results
6-Letter Alphabet
Word Length Accuracy Sensitivity Specificity
2 88.6 % 89.7 % 86.1 %
3 94.6 % 95.8 % 89.2 %
4 92.0 % 95.4 % 86.7 %
5 92.7 % 94.9 % 86.7 %
6 89.7 % 91.1 % 87.0 %
[4]
[4]
Dataset
Hypothesis
Acknowledgements
RR Interval
ECG Time Series
RR Interval RR Interval
Mike Borello / Wikimedia Commons / CC BY-SA 3.0 Retrieved from http://edumps.deviantart.com / CC BY-SA 3.0 User: Chris73 / Wikimedia Commons / CC BY-SA 3.0 BodyParts3D/Anatomography / Wikimedia Commons / CC BY_SA 2.1 jpUser: Prokofiev / Wikimedia Commons / CC BY-SA 3.0