Risk Prediction for Cardiovascular Disease using ECG Data in the

China Kadoorie Biobank

Yanting Shen1 , Yang Yang1,2 , Sarah Parish3 , Zhengming Chen3 , Robert Clarke3 , David A. Clifton1

Abstract We set out to use machine learning techniques knowledge of accurate reference values of the risk factors.
to analyse ECG data to improve risk evaluation of cardio- This approach has grown popular in risk evaluation and
vascular disease in a very large cohort study of the Chinese diagnostics for chronic diseases [4]. For example, in CVD,
population. We performed this investigation by (i) detecting
abnormality using 3 one-class classification methods, and (ii) Knuiman et al. have predicted coronary mortality in the
predicting probabilities of normality, arrhythmia, ischemia, Busselton cohort using logistic regression[5]. The electrocar-
and hypertrophy using a multiclass approach. diogram (ECG), being an important measurement of cardiac
For one-class classification, we considered 5 possible defini- function that is relatively easy to obtain, is surprisingly
tions for normality and used 10 automatically-extracted ECG rarely used for risk prediction. The aim of this paper is to
features along with 4 blood pressure features. The one-class
approach was able to identify abnormality with area-under- address the need for risk metrics that include ECG-derived
curve (AUC) 0.83, and with 75.6% accuracy. features by analysing CVD risks associated with abnormal
For four-class classification, we used 86 features in total, ECG, using the China Kadoorie Biobank dataset. This pa-
with 72 additional features extracted from the ECG. Accu- per describes two risk evaluation tasks: (i) abnormality
racy for this four-class classifier reached 75.1%. The methods detection and (ii) prediction of probabilities of normal,
demonstrated proof-of-principle that cardiac abnormality can
be detected using machine learning in a large cohort study. arrhythmia, ischemia, and hypertrophy. Since abnormality is
relatively rare in this database, novelty detection (the aim of
I. I NTRODUCTION which is to classify an under-sampled abnormal class) is an
Cardiovascular diseases (CVD) are the leading causes appropriate approach to address the first task. To address the
of mortality worldwide and in China [1]. There are large second task, we build models of normality, arrhythmia,
geographical and economic variations of CVD mortality in ischemia, and hypertrophy using a multiclass approach.
China [2], suggesting appropriate measures are needed for
II. DATASET D ESCRIPTION
prevention and effective treatment of the disease. The World
Health Organisation advises people at high CVD risk to The China Kadoorie Biobank (CKB) is a prospective
access early detection and treatment for prevention of CVD cohort study of over 520,000 adults from 10 areas in China
[1]. Identifying risks in individuals in the population could during 2004-2008 [6]. Data were collected using question-
help to provide advice to people to improve their lifestyle and naires and anthropometric and physiological measurements
help clinicians to discover appropriate treatments for specific were recorded at baseline and all participants provided a
conditions to reduce mortality and healthcare expenditure. blood sample. Information on cause of death rates was col-
Traditional CVD risk factors include smoking, hyper- lected from health insurance data and mortality and disease
cholesterolaemia, hypertension, diabetes, and obesity [2] registries. After five years, approximately 25,000 surviving
among many others. Traditional risk factors do not fully ex- participants were resurveyed with further questionnaires,
plain the risk of CVD in populations. Personalised medicine measurements, and blood collection. We have institutional
requires integrating all risk factors to which a person is ethics approval to use the data. Public access to the CKB data
exposed and then predicting risks for specific diseases, so can be found at http://www.ckbiobank.org/site/Data+Access.
to optimise preventative measures for individuals. Examples The data available for our study include:
of research for this purpose include [3], in which 5-year
mortality rate was predicted on 20 risk factors from 498,103 A. ECG time series
participants in UK Biobank using proportional hazard models Standard 12-lead ECG (10-s duration, 500Hz) was
and Harrells C-index. With an increasing number of risk fac- recorded on 24,369 participants using a Mortara ELIx50
tors being identified, and especially with abundant genetic- device in 2013-2014. Also available is a typical cycle from
and lifestyle- data now available, it can be expected that such each lead for each participant, which was generated by the
an approach will face difficulty as the healthy range of the device using a proprietary algorithm.
newly-identified factors is difficult to quantify.
Machine learning has the advantage of estimating the B. ECG Features
associations between risk factors and diseases without prior The Mortara device provides 10 main features (age,
average RR interval, P wave duration, the time point of
1 Department of Engineering Science, University of Oxford, UK
2 QRS offset, PR interval, QRS duration, QT duration, P axis,
Department of Mechanical Engineering, Shanghai Jiao Tong University,
200240, China QRS axis, and T axis) which were automatically extracted
3 Nuffield Department of Population Health, University of Oxford, UK from the typical cycles for each participant. A schematic

978-1-4577-0220-4/16/$31.00 2016 IEEE 2419

 TABLE I
4 C LASS INCLUSION CRITERIA AND CLASS SIZES

Class Number (%) Inclusion criteria

Normal 10803 (43.2) Normal ECG
Arrhythmia 2162 (8.6) Abnormal rhythm
Atrial fibrillation
Early repolarisation
Preexcitation
Premature ectopic beats
Ectopia
Blocks
Uncertain rhythm
Ischemia 1868 (7.5) Explicitly stated ischemia
Hypertrophy 3761 (15.0) Hypertrophy or enlargement
Unclassified 6425 (25.7) None of the above
Fig. 1. Schematic representation of key ECG features for Lead I

considered by us are (C1) sinus rhythm, (C2) normal ECG,

representation of a selection of these features is shown in (C3) abnormal ECG=0, (C4) sinus rhythm AND abnormal
Figure 1. ECG=0, and (C5) sinus rhythm OR abnormal ECG=0. These
will be referred to as the normal criteria C1-5.
C. Blood pressure data 2) Labels for four-class classification: For 4-class classi-
Systolic blood pressure and diastolic blood pressure were fication of normal, arrhythmia, ischemia and hyper-
recorded twice on each individual using standard methods. trophy, it is important to construct class models as precisely
as possible. We also require criteria for the 4-class setting,
D. Textual labels and here we consider schemes according to [7] which are
The Mortara device automatically provides up to 10 shown in Table I.
textual labels from 236 possibilities for each participant, such III. F EATURE E XTRACTION
as sinus rhythm, normal ECG, and atrial fibrillation. These
labels were produced by application of the Minnesota coding Six additional features were extracted from each of the
system which is a heuristic scheme defined. Sinus rhythm and 12-lead typical cycles. Amplitudes of P, Q, R, S, and T
normal ECG are the most commonly-observed textual labels waves relative to the baseline, and ST level (which was
in our dataset, representing 83.2% and 44.1% of all records, approximated as the level of QRS offset) constitute 72
respectively. additional features which were added to the 14 features in the
The outputs of the Mortara device included a label for one-class classification for use in the four-class classification.
abnormality for the waveform (abnormal ECG), which was The baseline was approximated as the average level of the
coded as 1 for abnormality and 0 for normality. segment between P offset and Q onset, because this segment
1) Labels for one-class novelty detection: Sinus rhythm, is used to define ST deviation in clinic. We note that those 72
normal ECG, and abnormal ECG=0 were the only three features derived from the ECG time-series were in addition
labels available in our dataset that were used as potential to the 10 features automatically produced by the Mortara
criteria for defining normality. There were 19,925, 10,550, device, and the 4 blood pressure measurements, described
and 18,397 records labelled sinus rhythm, normal ECG, and earlier.
abnormal ECG=0, respectively. Sinus rhythm and abnormal The positions of the onsets and offsets of these waves are
ECG=0 had 15,628 records in overlap. Normal ECG, which provided by the Mortara device (part of the Mortara set of
was a subset of the union of sinus rhythm and abnormal features described above).
ECG=0, included 10,550 records. There are only 1,675 IV. A NALYSIS
complete records without any of the three labels for defining
normality. A. One-class classification
Since the labels were produced according to the heuristic We present three methods to predict the posterior prob-
Minnesota criteria, their relationships to diseases require ability of a feature vector belonging to the abnormal class,
future study. In addition, these criteria have a large number and hence predict their cardiovascular disease risk.
of rules in disagreement due to contradictions and incon- 1) Cross-Validation and Partitioning of Training and Test
sistencies in the coding system. As a result, it may not Sets: The training and test sets were generated using a 5-
be convincing to treat any of these labels alone as being fold cross-validation by permuting the entire dataset and
the gold standard for use in defining normality when assigning a different 20% for each fold of cross-validation
training a classification algorithm. This report compares as the test set. Results shown later are the mean values over
the performance of three algorithms against five reasonable this 5-fold cross-validation. All training and test sets were
combinations of the three labels. The five combinations normalised and standardised according to the training sets.

2420
 2) Balancing of the Test Sets: To make a fair comparison equation 4. Similarly the posterior was thresholded at 0.5 for
between the normal criteria C1-C5 which have different classification.
class ratios (i.e. balance between normal and abnormal data), 5) Discriminative Support Vector Machine for one-class
we use the accuracy and AUC in balanced sets for model classification: To compare the results of KDE, we also used
evaluation. SVM. The coefficients C and (when using a Gaussian
We therefore created a balanced test set (a subset of kernel) control the flexibility of the separation boundary,
the unbalanced test set), containing all abnormal test data and were optimised by a grid search via 5 fold cross-
and the same number of normal data. The training set validation on the training set. The classification score was
remained unbalanced. The balanced test sets under criteria mapped to probabilities and thus the training set posterior
C1-5 contain 1,824, 5,688, 2,452, 3,586, and 688 data points P (C|y) was learned.
respectively.
3) Generative Kernel Density Estimator: We adapted the B. Four-class classification
model described in [8]. In brief, the normal probability 1) Constructing the training and test sets: We obtained
density function was learned from the training set by placing balanced training and test sets by taking all data from the
a multivariate Gaussian distribution on each 14-dimensional smallest class (Table I) and the same number of data points
data point. For ease of computation, we performed k-means from each of the other classes were randomly selected to
clustering to summarise the normal data with 500 cluster construct the training and test sets for 5-fold cross-validation.
centres in the 14-dimensional space. Only the most normal For example, the four-class balanced training-and-test set
(i.e., those labelled Normal ECG) were used in clustering. contains 1868 4 = 7472 datapoints. To illustrate the
The data likelihood is calculated via: distinctiveness of each class, three-class and two-class clas-
sifications of any combinations of the normal, ischemia,
N
1 X | xxi |2 arrhythmia, and hypertrophy were also performed for
p(x) = D e 2 2 (1)
N (2) D
2 comparison using the same approach to balance classes.
i=1
2) Training the 4-class model with Support Vector Regres-
A novelty score, y, is then calculated using equation 2. sion: A P (C|x) was estimated for each of the classes using
y(x) = log p(x) (2) support vector regression in a one-vs-all approach; i.e. the
regressor i was learned in a training set with only the class
i labelled 1 and other classes were labelled 0. The class
P (y|C)P (C) probability P (C|x) was calculated from the predicted value
P (C|y) = (3) of the regressor i according to Equation 5. Finally the data
P (y)
point was classified to the class with the highest probability:
We propose treating this novelty score as a univari-
|1yi |
ate summary of the 14-dimensional data, which may then e i

subsequently used by probabilistic models to predict the Pi (Ci |x) = P4 , i = 1, 2, 3, 4 (5)

2i j=1 Pj (Cj |x)
probability of test data belonging to the abnormal class. First,
we estimated the likelihood P (y|C) in the training set, by
performing a kernel density estimation of the class-specific C. Results and Discussion
pdf. For the unbalanced test set, the prior P (C) was set to
1) One-class classification: In the balanced hold-out test
the class ratio of the training set; for the balanced set, the
set, the discriminative SVM achieved high accuracy, 71.6%
class prior equals 0.5. The posterior P (C|y) is
to 75.6%, for all 5 candidate criteria for defining normality,
Ptrain (C|y)Ptrain (y)Ptest (C) while the generative KDE had comparable result (74.8%)
Ptest (C|y) = (4)
Ptrain (C)Ptest (y) under the criteria C5 (Figure II).
The discriminative KDE had similar AUC values as the
The threshold for posterior Ptest (C|y) was set at 0.5 for generative KDE, but lower accuracy, which implies a better
classification. optimisation of this method may be needed.
4) Discriminative KDE: Alternatively, we added the nov- For both KDE and SVM C5 was the best-performing cri-
elty score into a discriminative framework by solving the terion, suggesting C5 may be the most appropriate criterion
inference problem P (C|y) directly. The posterior was es- among the 5 studied to be a gold standard for training
timated in the training set by binning y and calculating algorithms for one-class classification.
the frequency of observing a particular class in a bin. For 2) Four-class classification: The four-class classification
example, for normal class C0 , the posterior in each bin was results using support vector regression are shown in Figure
calculated as the proportion of data points belonging to C0 2. The 72 new features improved the results in all cases,
in the bin. Ideally the bin size should be determined by most markedly in classifications involving ischemia and
cross-validation [9]. In this paper, the bin size was set to hypertrophy. This agrees with our expectation since the 10
y = 1. The posterior of a set with a different prior, in Mortara features do not contain information concerning the
this paper the balanced test set, was calculated according to amplitudes of the peaks, while ischemia and hypertrophy

2421
 TABLE II
AUC AND ACCURACY OF PREDICTING THE 5 NORMAL CRITERIA BY GENERATIVE KDE, DISCRIMINATIVE KDE, AND DISCRIMINATIVE SVM IN THE
BALANCED SETS . R ESULTS ARE PRESENTED AS THE MEAN STANDARD DEVIATION IN 5- FOLD CROSS - VALIDATION .

SVM Discriminative Generative KDE Discriminative KDE

AUC Accuracy % AUC Accuracy % AUC Accuracy %
C1 0.79 71.7 0.73 67.2 0.73 59.6
C2 0.79 71.6 0.73 63.5 0.73 61.1
C3 0.82 77.1 0.72 65.1 0.72 59.3
C4 0.80 73.7 0.75 64.5 0.75 61.4
C5 0.83 75.6 0.81 74.8 0.80 60.6

the borderline data that are otherwise difficult to classify.

We can further improve the classification accuracy by
extracting more features, such as heart-rate variability and T-
wave alternans. Also the original 10s signal may lend more
information than the typical cycle as the former contains
more time-dependent information than the latter. The length
of the signal is a major limitation to our feature extraction,
because many informative features such as ST-level need
longer (>60s) signals to be evaluated accurately. Future
work will link our analysis of ECG data to electronic health
records for each participant, obtained from the Chinese
medical insurance system, disease and mortality registries.
ACKNOWLEDGMENT
Fig. 2. Classification accuracies. Numbers on the edges are classification The authors gratefully acknowledges the support of the
accuracies between the two classes on the nodes; numbers in the centres of RCUK Digital Economy Programme grant EP/G036861/1
the triangles are classification accuracies of the 3 classes on the nodes of (Oxford Centre for Doctoral Training in Healthcare Inno-
the triangle. The red and black numbers are results with and without the
72 features derived from the ECG, over and above the basic set of the 10 vation), China Scholarship Council (CSC), the K.C. Wong
features provided by the Mortara device and the 4 blood pressure features, Fellowship, the EPSRC, and the Royal Academy of Engi-
respectively. neering. The China Kadoorie Biobank study was supported
by grants from the MRC, Wellcome Trust, and the British
Heart Foundation.
were highly correlated with amplitude abnormalities, espe-
cially ST-levels, and R-S amplitudes. R EFERENCES
It is encouraging that the classification accuracies with [1] WHO, Cardiovascular diseases factsheet, 2015. [Online]. Available:
the new features are 30 to 50 percentage points higher http://www.who.int/mediacentre/factsheets/fs317/en/
[2] H. B. Hubert, M. Feinleib, P. M. McNamara, and W. P. Castelli,
than those that would be obtained by chance, suggesting Obesity as an independent risk factor for cardiovascular disease: a
machine learning methods can achieve high agreement with 26-year follow-up of participants in the Framingham heart study.
clinical knowledge, without resorting to a complex rule- Circulation, vol. 67, no. 5, pp. 968977, 1983.
[3] A. Ganna and E. Ingelsson, 5 year mortality predictors in 498 103 uk
based system. However it is worth noting that all labels biobank participants: a prospective population-based study, The Lancet,
here were provided by the Mortara device, and our models vol. 386, no. 9993, pp. 533540, 2015.
need further validation by comparing with labels from human [4] L. Mena and J. A. Gonzalez, Machine learning for imbalanced
datasets: application in medical diagnostics. in FLAIRS Conference,
experts in ECG resurvey. 2006, pp. 574579.
[5] M. W. Knuiman and H. Vu, Prediction of coronary heart disease mor-
V. C ONCLUSIONS AND F UTURE W ORK tality in Busselton, Western Australia: an evaluation of the Framingham
national health epidemiologic follow up study, and WHO ERICA risk
The first task of this research was to detect abnormal- scores. Journal of Epidemiology and Community Health, vol. 51, no. 5,
ity, by exploring different one-class novelty detection algo- pp. 515519, 1997.
rithms under various criteria of normality. The algorithms [6] L. Yang, L. Li, S. Lewington, Y. Guo, P. Sherliker, Z. Bian, R. Collins,
R. Peto, Y. Liu, R. Yang et al., Outdoor temperature, blood pressure,
favoured criteria C5, which was the least stringent of all and cardiovascular disease mortality among 23 000 individuals with di-
possibilities, and is thus appropriate for detection of outliers. agnosed cardiovascular diseases from china, European Heart Journal,
In view of the relatively good performance of the discrimina- vol. 36, no. 19, pp. 11781185, 2015.
[7] P. Ramrakha and J. Hill, Oxford handbook of cardiology. Oxford
tive SVM, in the second task of this study we modelled the University Press, 2012.
normal, arrhythmia, ischemia, and hypertrophy using multi- [8] L. Tarassenko, D. Clifton, P. Bannister, S. King, and D. King, Novelty
class support vector regression, aiming to produce accurate detection, in Encyclopedia of structural health monitoring. Wiley,
London, 2009.
models for prediction of the unclassified data points. The [9] L. Clifton, D. A. Clifton, Y. Zhang, P. Watkinson, L. Tarassenko, and
encouraging results suggest the multiclass models may be H. Yin, Probabilistic novelty detection with support vector machines,
appropriate to predict the probability of class membership of Reliability, IEEE Transactions on, vol. 63, no. 2, pp. 455467, 2014.

2422