Вы находитесь на странице: 1из 4

AAMI based ECG Heart-beat Time-series Clustering

using Unsupervised ELM and Decision Tree


Jagadeeswara Rao Annam Bapi Raju Surampudi
University of Hyderabad Indian Institute of Information Technology
Hyderabad, India Hyderabad, India
ajagarao@gmail.com bapiks@yahoo.co.in

Abstract— Early detection of cardiovascular diseases can ventricular arrhythmias. Sudden cardiac death due to such
prevent the premature deaths caused by abnormal heartbeat arrhythmias is an important cause of cardiovascular mortality.
problems. Application of unsupervised classification by Extreme
learning machine is addressed for ElectroCardiogram (ECG) But Arrhythmias occur infrequently in some patients and many
heart-beat time series analysis by a hybrid of extreme learning patients have symptoms occurring weekly or monthly, in
machine and Decision tree using full heart-beat time series by which case conventional ECG of short duration is not useful
alignment of R-peaks of all beats is proposed in this work.
for identifying these abnormalities. As these arrhythmias are
PQRST Time series of heart-beats having converted into equal
length series by alignment of R-peaks of all heart-beats based on infrequent and may occur only during certain emotions or
R-peak of largest length PQRST series in the data and by padding activities like stress or walking, it is useful to record the
zeroes to the smaller length series on either side, was used in this electrical activity of the heart over a period of time, usually 24
experimentation. The main objective of this paper is to identify
the abnormalities in ECG heart beats based on AAMI
or 48 hours. An intermittent recorder is more useful for
Categorization. Because of the large patient specific recording infrequent symptoms. These intermittent recorders
characteristics in ECG heartbeat morphology across individuals, are worn by patients for periods of many weeks to capture
the supervised methods tuned for a specific ECG dataset may infrequently occurring events. The extended wearable ECG
decrease performances in other datasets. In addition, these
techniques require a considerable amount of known and labelled recorder is a Holter monitor called as Ambulatory
heartbeats which are not feasible when having long–term ECG electrocardiography (AECG) which uses a storage system for
monitoring. Experiments were conducted on ECG data of 44 acquiring ECG information that can then be transmitted by
patients obtained from MIT-BIH Arrhythmia database. Results
were compared with existing methods such as weighted support
telephone or mobile or internet to the concerned cardiologist
vector machine (SVM), hierarchical SVM and weighted linear and analysed.
discriminant analysis (LDA). Comparative analysis confirms the
viability and superiority of the proposed approach in terms of A. Motivation
Total classification accuracy (TCA). Proposed system achieved Because of the large variability in ECG heartbeat
Sensitivities of 98.13%, 82.25%, 76.49%, 52.20%, PPV of morphology across individuals, the supervised methods tuned
98.13%, 64.46%, 95.47%, 46.54% for N, S, V, and F classes for a specific ECG dataset may decrease performances in other
respectively and TCA of 95.75%. datasets. In addition, these supervised techniques require a
considerable amount of known and labelled heartbeats which
Keywords— Clustering; embedding;extreme learning machine; are not feasible when having long–term ECG monitoring.
manifold regularization; unsupervised learning; AAMI; ECG;
II. METHODS AND MATERIA LS
Heart beat;
A. Unsupervised-ELM
I. INTRODUCTION Classification capability of the generalized SLFNs with the
Cardiovascular diseases (CVDs) are the leading cause of hidden-layer mapping h(x) satisfies the universal
mortality globally as more people die annually from CVDs approximation condition. Solving the regularized least square
problem in ELM is faster than solving the quadratic
than from any other cause. 80% of these CVD deaths take programming problem in standard SVMs.
place in low and middle income countries. Early detection and
The Unsupervised-ELM (US-ELM) for time series
timely treatment of these diseases can prevent these premature
proposed in this paper is related to the Laplacian Eigenmaps
deaths. ECG as a non-invasive diagnostic method, is cost (LE) [14] and spectral clustering (SC) [14] in that they both
effective among all cardiac-related investigations. use spectral techniques for embedding and clustering. In all
Abnormal electrical activity in the heart is known as these algorithms, an affinity matrix is first built from the input
cardiac Arrhythmia and heart attacks may present as patterns. The SC performs eigendecomposition on the
normalized affinity matrix, and then embeds the original data
into a d-dimensional space using the first d eigenvectors (each IV. EXPERIM ENTA L R ESULTS.
row is normalized to have unit length and represents a point in
The clustering experiments were conducted based on the
the embedded space) corresponding to the d largest eigen
division of MIT-BIH records into two data sets as used by
values. The LE algorithm performs generalized eigen-value-
decomposition on the graph Laplacian, and uses the d Lannoy [10] is shown in Table 2. The AAMI standards are
eigenvectors corresponding to the second through the (d+1) th adopted in this study and our results are compared to those of
smallest eigen values for embedding. When LE and SC are works which used AAMI based inter-patient classification.
used for clustering, then k-means is adopted to cluster the data There are two challenging issues in Heart-beat Classification.
in the embedded space. Similar to LE and SC, the US-ELM are 1) Inter-patient beat variations. 2) Intra-patient beat variations.
also based on the affinity matrix, and it is converted to solving Each class manifests inter-patient beat variations i.e across
a generalized eigen-value-decomposition problem. However, different patients, beats of the same class have variations
the eigenvectors obtained in US-ELM are not used for data because of the patient specific data. In Intra-patient data, each
representation directly, but are used as the parameters of the class may have variations i.e some leads or sensors of ECG
network, i.e., the output weights. Note that once the US-ELM may not capture the variations or discriminations across
model is trained, it can be applied to any presented data in the different classes.
original input space. In this way, US-ELM provide a The proposed method overcomes the issue of Inter-patient
straightforward way for handling new patterns without variations in each class and solve the issue of Intra-patient
recomputing eigenvectors as in LE and SC. variations of different classes by the normalization method
B. Translation transform for alignment of R-peaks effectively. Huang [6] used random projection and support
vector machine (SVM) ensemble to detect VEB then the ratio
 Find O, the position of the R-peak in the beat having
of the RR interval to mean RR imterval was compared to a
largest length PQRST series in the data i.e in Train
Data and Test Data. ( i.e position of the R-peak from predetermined threshold to detect SVEB on the dataset 2 of
the beginning of the beat). MIT-BIH Arrhythmia database [2], but used only 3 classes in
experimentation. Our method uses single ELM where as
 Translate beats into 2-d array by applying the transform Huang [6] used Ensemble of 15 SVMs for VEB detection.
to all beats so that all beats align their Rpositions at a V. CONCLUSIONS.
common point i.e O. Classification accuracies of the proposed approach are
1. F( O-Rpeakposition +1 : O )= Beat(1: Rpeakposition ) compared with those of support vector and linear discriminant
based classification schemes. The True positives NTP, STP,
2. F(O+1 : O+Beat length - R peak position) =
Beat(Rpeak position +1 : Beat length ) (1) VTP and FTP and False positives FPN, FPS, FPV and FPF of
N, S, V and F classes are obtained and is shown in Table III,
III. PROPOSED M ETHOD
Table IV and Table V. And the proposed method shows
The proposed method uses a combination of ELM and superior performance results over other methods and thus its
Decision Tree such that from the predicted output of ELM, viability is established.
V-class and F-class are to be detected. And for the remaining
heartbeats, the decision rule based on the normalised RR Future work would focus on improving classification
previous interval (the ratio of interval between current R peak accuracies of all the classes, while retaining overall superior
to the previous R-peak to the mean RR previous interval is classification accuracy.
compared to a predetermined threshold to detect S-class and TABLE 2. AAMI mapping to MIT-BIH classes
N-class). # MIT-BIH MIT-BIH Class and label AAMI
Use of the RR interval ratio can reduce the overlap between 1 . or N Normal (1)
AAMI SVEB class and AAMI class N heartbeat and thus 2 L Left Bundle Branch Block (2)
increase the SVEB detection rate. 3 R Right Bundle Branch blocks (3) N
4 J Nodal (junctional) escape (11)
TABLE 1. Data sets of MIT-BIH Arrhythmia Database
5 E Atrial Escape (34)
DataSet MIT-BIH Records 6 A Aberrated Atrial premature (4)
7 J Nodal (junctional) premature (7)
101, 106, 108, 109, 112, 114, 115, 8 A Atrial premature (8) S
Train 116, 118, 119, 122, 124, 201, 203, 9 S Supraventricular premature (9)
DataSet 205, 207, 208, 209, 215, 220, 223 10 V Ventricle Premature contraction (5) V
& 230 11 E Ventricular escape (10)
12 F Fusion of ventricular &normal (6) F
100, 103, 105, 111, 113, 117, 121, 13 /or p Paced beat (12)
123, 200, 202, 210, 212, 213, 214, 14 F Fusion of Paced and Normal (38) Q
Test
219, 221, 222, 228, 231, 232, 233 15 Q Unclassifiable beat (13)
DataSet
& 234
Table III. Overall Co mparison of Classificat ion Accuracy in%.
Sensitivities Positive Predictive values TCA%
Method N S V F N S V F %
Proposed
ELMtimeclust 98.13% 82.25% 76.49% 52.20% 98.13% 64.46% 95.47% 46.54% 95.75%
Huang, 2014 [6] 99.2 91.1 93.9 - 95.2 42.2 90.9 - 93.8
Weighted LDA, 95 77 81 - 98 38 87 - 93.0
Llamedo 2011 [3]
MLP, TanisMar, 89.6 83.2 86.8 61.1 99.3 33.5 75.9 16.6 89.0
2011 [5]
SVMensemble, 88.9 79.1 85.5 93.8 98.9 35.9 92.8 13.74 86.6
Zhang, 2014, [12]
Hierar.SVM, Park, 86.3 82.6 80.9 54.9 - - - - 85.6
2008 [4]
WeightedCRF+L1 79.8 92.6 85.2 84.5 - - - - 85.4
Lannoy, 2012 [10]
Weighted LDA, De 86.9 75.9 77.7 89.3 99.2 38.5 81.9 8.6 77.7
Chazal, 2004 [11]

Table IV. Sensitivity and Positive Predictive Values (PPV) of 4 Classes used in experiments
Etclust NFP SFP VFP FFP NTP STP VTP FTP TP TOTAL ACC
100 34 4 3 0 2232 0 0 0 2232 2273 98.20%
103 2 1 0 0 2081 0 0 0 2081 2084 99.86%
105 36 1 0 0 2525 1 0 0 2526 2567 98.40%
111 0 0 0 0 2123 0 0 0 2123 2124 99.95%
113 5 0 0 0 1788 1 0 0 1789 1795 99.67%
117 1 1 0 0 1533 0 0 0 1533 1535 99.87%
121 2 0 2 0 1857 0 0 0 1857 1863 99.68%
123 2 0 0 0 1513 0 0 0 1513 1518 99.67%
200 81 6 9 19 1728 0 756 1 2485 2601 95.54%
202 74 1 4 1 2055 1 0 0 2056 2136 96.25%
210 75 28 74 3 2375 0 95 0 2470 2650 93.21%
212 0 0 0 0 2748 0 0 0 2748 2748 100.00%
213 165 394 17 44 2252 6 174 199 2631 3251 80.93%
214 129 0 0 0 2003 0 129 0 2132 2262 94.25%
219 70 0 2 1 2080 1 0 0 2081 2154 96.61%
221 1 0 0 0 2031 0 395 0 2426 2427 99.96%
222 60 54 0 0 2219 149 0 0 2368 2483 95.37%
228 11 69 1 0 1605 0 353 0 1958 2053 95.37%
231 3 3 1 0 1564 0 0 0 1564 1571 99.55%
232 11 175 0 0 182 1340 0 0 1522 1780 85.51%
233 25 92 4 164 2230 0 561 2 2793 3079 90.71%
234 41 4 0 0 2696 12 0 0 2708 2753 98.37%
49705 828 833 117 232 43420 1511 2463 202 47596 49707 95.75%
Se% 98.13% 82.25% 76.49% 52.20% 95.75%
ppv 98.13% 64.46% 95.47% 46.54% 95.75%
[8] Yeh Y.C., Wang, W.J.,Ch iou, C.W. “A novel fuzzy c-means method
References for classifying heartbeat cases from ECG signals" . Measurment, 43,
pp. 1542-1545, 2010.
[1] AAMI. “Testing and reporting performance results of cardiac
rhythm and st segment measurement algorith ms.ANSI/AAMI [9] . K. Tan, K. Chan, and K. Choi, “Detection of the QRS comp lex, P
EC38:1998” ANSI/AAMI EC13:2002, Arlington, VA". Association wave and T wave in electrocardiogram." Advances in medical signal
for the Advancement of Medical Instrumentation. (2002) and information processing, pages 41-47 (2000)
[2] MIT-BIH A rrhythmia Database http://www.physionet.org/physio- [10] G de Lannoy JD D Francois, Verleysen M, “Feature Relevance
bank/database/mitdb/ Assessment In Automatic Inter- patient Heart Beat Classificat ion." In
Proceedings of the International Conference on Bio-inspired
[3] Llamedo-Soria M, Mart inez JP, “Heartbeat classi_cation using Systems and Signal Processing, Valencia, Spain, 13-20(2010)
feature selection driven by database generalization criteria". IEEE
Trans Bio med Eng, 58:616625, (2011) [11] de Lannoy G, Francois D, Delbeke J, Verleysen M, “Weighted
conditional random fields for supervised inter-patient heartbeat
[4] Park KS, Cho BH, Lee DH, Song SH, Lee JS, Chee YJ, Kim IY, Kim classification." IEEE Trans Bio med Eng,59:241-247(2012)
SI, “Hierarch ical Support Vector Machine Based Heartbeat
Classi_cation Using Higher Order Statistics and Hermite Basis [12] P. de Chazal and R. B. Reilly, ”A patient-adapting heartbeat
Function." Proceedings of 35 Annual Co mputers in Cardiology classi_er using ECG mo rphology and heartbeat interval features"
Conference, IEEE Bo logna (2008) IEEE Transactions on Bio medical Engineering, vol. 53, no. 12, pp.
2535--2543, (2006)
[5] Mar Tanis, Zaunseder S, Mart inez JP, “Optimization of ECG
classi_cation by means of feature selection." IEEE Trans [13] Zhang ZC, Dong J, Luo XQ, Choi KS, Wu XJ: “Heartbeat
Bio medical Engineering,58,2168-2177 (2011) classification using disease-specific featureselection." Co mput Bio l
Med, 46:79-89. (2014)
[6] Huang et al. “A new hierarchical method for inter-patient heartbeat
classification using random project ions and RR intervals." [14] Mohamed Elgendi, M irjam Jonkman, Friso De Boer, “Premature
Bio medical Engineering OnLine 2014 13:90.doi:10.1186/1475-925X- Atrial Co mp lexes Detection using The Fisher Linear Discriminant."
13-90 (2014) 7th IEEE International Conference on Cognit ive Informat ics ( ICCI),
pp. 83-88, Aug.2008
[7] G. Huang, S. Song, J. N. D. Gupta, and C. Wu, “Semi-supervised and
Unsupervised Extreme Learning Machines,” (in press) IEEE
Transactions on Cybernetics, 2014

Вам также может понравиться