{"title": "Markov Models for Automated ECG Interval Analysis", "book": "Advances in Neural Information Processing Systems", "page_first": 611, "page_last": 618, "abstract": "", "full_text": "Markov Models for Automated ECG Interval\n\nAnalysis\n\nNicholas P. Hughes, Lionel Tarassenko and Stephen J. Roberts\n\nDepartment of Engineering Science\n\nUniversity of Oxford\nOxford, 0X1 3PJ, UK\n\nfnph,lionel,sjrobg@robots.ox.ac.uk\n\nAbstract\n\nWe examine the use of hidden Markov and hidden semi-Markov mod-\nels for automatically segmenting an electrocardiogram waveform into\nits constituent waveform features. An undecimated wavelet transform\nis used to generate an overcomplete representation of the signal that is\nmore appropriate for subsequent modelling. We show that the state dura-\ntions implicit in a standard hidden Markov model are ill-suited to those\nof real ECG features, and we investigate the use of hidden semi-Markov\nmodels for improved state duration modelling.\n\n1\n\nIntroduction\n\nThe development of new drugs by the pharmaceutical industry is a costly and lengthy pro-\ncess, with the time from concept to \ufb01nal product typically lasting ten years. Perhaps the\nmost critical stage of this process is the phase one study, where the drug is administered\nto humans for the \ufb01rst time. During this stage each subject is carefully monitored for any\nunexpected adverse effects which may be brought about by the drug. Of particular interest\nis the electrocardiogram (ECG1) of the patient, which provides detailed information about\nthe state of the patient\u2019s heart.\n\nBy examining the ECG signal in detail it is possible to derive a number of informative\nmeasurements from the characteristic ECG waveform. These can then be used to assess the\nmedical well-being of the patient, and more importantly, detect any potential side effects\nof the drug on the cardiac rhythm. The most important of these measurements is the \u201cQT\ninterval\u201d. In particular, drug-induced prolongation of the QT interval (so called Long QT\nSyndrome) can result in a very fast, abnormal heart rhythm known as torsade de pointes,\nwhich is often followed by sudden cardiac death 2.\nIn practice, QT interval measurements are carried out manually by specially trained ECG\nanalysts. This is an expensive and time consuming process, which is susceptible to mis-\ntakes by the analysts and provides no associated degree of con\ufb01dence (or accuracy) in the\nmeasurements. This problem was recently highlighted in the case of the antihistamine\n\n1The ECG is also referred to as the EKG.\n2This is known as Sudden Arrhythmia Death Syndrome, or SADS.\n\n\f0.06\n\n0.04\n\n0.02\n\n0\n\n\u22120.02\n\n\u22120.04\n\n\u22120.06\n\n\u22120.08\n\nQRS complex \n\nT wave \n\nP wave \n\nBaseline 1 \n\nU wave \n\nBaseline 2 \n\nP\non\n\nP\noff\n\nQ\n\nJ\n\nT\n\noff\n\nU\n\noff\n\nFigure 1: A human ECG waveform.\n\nterfenadine, which had the side-effect of signi\ufb01cantly prolonging the QT interval in a num-\nber of patients. Unfortunately this side-effect was not detected in the clinical trials and\nonly came to light after a large number of people had unexpectedly died whilst taking the\ndrug [8].\n\nIn this paper we consider the problem of automated ECG interval analysis from a machine\nlearning perspective. In particular, we examine the use of hidden Markov models for auto-\nmatically segmenting an ECG signal into its constituent waveform features. A redundant\nwavelet transform is used to provide an informative representation which is both robust to\nnoise and tuned to the morphological characteristics of the waveform features. Finally we\ninvestigate the use of hidden semi-Markov models for explicit state duration modelling.\n\n2 The Electrocardiogram\n\n2.1 The ECG Waveform\n\nEach individual heartbeat is comprised of a number of distinct cardiological stages, which\nin turn give rise to a set of distinct features in the ECG waveform. These features represent\neither depolarization (electrical discharging) or repolarization (electrical recharging) of the\nmuscle cells in particular regions of the heart. Figure 1 shows a human ECG waveform and\nthe associated features. The standard features of the ECG waveform are the P wave, the\nQRS complex and the T wave. Additionally a small U wave (following the T wave) is\noccasionally present.\n\nThe cardiac cycle begins with the P wave (the start and end points of which are referred\nto as Pon and Po(cid:11)), which corresponds to the period of atrial depolarization in the heart.\nThis is followed by the QRS complex, which is generally the most recognisable feature of\nan ECG waveform, and corresponds to the period of ventricular depolarization. The start\nand end points of the QRS complex are referred to as the Q and J points. The T wave\nfollows the QRS complex and corresponds to the period of ventricular repolarization. The\nend point of the T wave is referred to as To(cid:11) and represents the end of the cardiac cycle\n(presuming the absence of a U wave).\n\n\f2.2 ECG Interval Analysis\n\nThe timing between the onset and offset of particular features of the ECG (referred to as an\ninterval) is of great importance since it provides a measure of the state of the heart and can\nindicate the presence of certain cardiological conditions. The two most important intervals\nin the ECG waveform are the QT interval and the PR interval. The QT interval is de\ufb01ned\nas the time from the start of the QRS complex to the end of the T wave, i.e. To(cid:11) (cid:0) Q, and\ncorresponds to the total duration of electrical activity (both depolarization and repolariza-\ntion) in the ventricles. Similarly, the PR interval is de\ufb01ned as the time from the start of the\nP wave to the start of the QRS complex, i.e. Q (cid:0) Pon, and corresponds to the time from\nthe onset of atrial depolarization to the onset of ventricular depolarization.\n\nThe measurement of the QT interval is complicated by the fact that a precise mathematical\nde\ufb01nition of the end of the T wave does not exist. Thus T wave end measurements are\ninherently subjective and the resulting QT interval measurements often suffer from a high\ndegree of inter- and intra-analyst variability. An automated ECG interval analysis system,\nwhich could provide robust and consistent measurements (together with an associated de-\ngree of con\ufb01dence in each measurement), would therefore be of great bene\ufb01t to the medical\ncommunity.\n\n2.3 Previous Work on Automated ECG Interval Analysis\n\nThe vast majority of algorithms for automated QT analysis are based on threshold methods\nwhich attempt to predict the end of the T wave as the point where the T wave crosses a\npredetermined threshold [3]. An exception to this is the work of Koski [4] who trained a\nhidden Markov model on raw ECG data using the Baum-Welch algorithm. However the\nperformance of this model was not assessed against a labelled data set of ECG waveforms.\nMore recently, Graja and Boucher have investigated the use of hidden Markov tree models\nfor segmenting ECG signals encoded with the discrete wavelet transform [2].\n\n3 Data Collection\n\nIn order to develop an automated system for ECG interval analysis, we collected a data\nset of over 100 ECG waveforms (sampled at 500 Hz), together with the corresponding\nwaveform feature boundaries3 as determined by a group of expert ECG analysts. Due to\ntime constraints it was not possible for each expert analyst to label every ECG waveform\nin the data set. Therefore we chose to distribute the waveforms at random amongst the\ndifferent experts (such that each waveform was measured by one expert only).\n\nFor each ECG waveform, the following points were labelled: Pon, Po(cid:11), Q, J and To(cid:11) (if a\nU wave was present the Uo(cid:11) point was also labelled). In addition, the point corresponding\nto the start of the next P wave (i.e. the P wave of the following heart beat), NPon, was also\nlabelled. During the data collection exercise, we found that it was not possible to obtain\nreliable estimates for the Ton and Uon points, and therefore these were taken to be the J\nand To(cid:11) points respectively.\n\n4 A Hidden Markov Model for ECG Interval Analysis\n\nIt is natural to view the ECG signal as the result of a generative process, in which each\nwaveform feature is generated by the corresponding cardiological state of the heart. In\naddition, the ECG state sequence obeys the Markov property, since each state is solely\n\n3We developed a novel software application which enabled an ECG analyst to label the boundaries\n\nof each of the features of an ECG waveform, using a pair of \u201conscreen calipers\u201d.\n\n\fP wave\nBaseline 1\nQRS complex\nT wave\nBaseline 2\nU wave\n\n5.5\n1.7\n1.0\n0.9\n2.3\n0.6\n\n47.2\n80.0\n11.3\n1.8\n32.2\n25.3\n\n0.5\n1.6\n79.0\n1.2\n1.3\n0.6\n\n4.4\n1.3\n4.6\n83.6\n3.5\n3.9\n\n26.5\n9.5\n2.7\n7.3\n31.8\n26.8\n\n15.9\n5.9\n1.4\n5.2\n28.9\n42.8\n\nTable 1: Percentage confusion matrix for an HMM trained on the raw ECG data.\n\ndependent on the previous state. Thus, hidden Markov models (HMMs) would seem ideally\nsuited to the task of segmenting an ECG signal into its constituent waveform features.\n\nUsing the labelled data set of ECG waveforms we trained a hidden Markov model in a su-\npervised manner. The model was comprised of the following states: P wave, QRS complex,\nT wave, U wave, and Baseline. The parameters of the transition matrix aij were computed\nusing the maximum likelihood estimates, given by:\n\n^aij = nij=Xk\n\nnik\n\n(1)\n\nwhere nij is the total number of transitions from state i to state j over all of the label se-\nquences. We estimated the observation (or emission) probability densities bi for each state\ni by \ufb01tting a Gaussian mixture model (GMM) to the set of signal samples corresponding\nto that particular state4. Model selection for the GMM was performed using the minimum\ndescription length framework [1].\n\nIn our initial experiments, we found that the use of a single state to represent all the regions\nof baseline in the ECG waveform resulted in poor performance when the model was used\nto infer the underlying state sequence of new unseen waveforms. In particular, a single\nbaseline state allowed for the possibility of the model returning to the P wave state, follow-\ning a P wave - Baseline sequence. Therefore we decided to partition the Baseline state into\ntwo separate states; one corresponding to the region of baseline between the Po(cid:11) and Q\npoints (which we termed \u201cBaseline 1\u201d), and a second corresponding to the region between\nthe To(cid:11) and NPon points5 (termed \u201cBaseline 2\u201d).\nIn order to fully evaluate the performance of our model, we performed 5-fold cross-\nvalidation on the data set of 100 labelled ECGs. Prior to training and testing, the raw\nECG data was pre-processed to have zero mean and unit energy. This was done in order\nto normalise the dynamic range of the signals and stabilise the baseline sections. Once\nthe model had been trained, the Viterbi algorithm [9] was used to infer the optimal state\nsequence for each of the signals in the test set.\n\nTable 1 shows the resulting confusion matrix (computed from the state assignments on\na sample-point basis). Although reasonable classi\ufb01cation accuracies are obtained for the\nQRS complex and T wave states, the P wave state is almost entirely misclassi\ufb01ed as Base-\nline 1, Baseline 2 or U wave. In order to improve the performance of the model, we require\nan encoding of the ECG that captures the key temporal and spectral characteristics of the\nwaveform features in a more informative representation than that of the raw time series\ndata alone. Thus we now examine the use of wavelet methods for this purpose.\n\n4We also investigated autoregressive observation densities, although these were found to perform\n\npoorly in comparison to GMMs.\n\n5If a U wave was present the Uo(cid:11) point was used instead of To(cid:11).\n\n\fP wave\nBaseline 1\nQRS complex\nT wave\nBaseline 2\nU wave\n\n74.2\n15.8\n\n0\n0\n1.4\n0.1\n\n14.4\n81.5\n2.1\n0\n0\n0.1\n\n0.1\n1.7\n94.4\n1.0\n0\n0.1\n\n0.3\n0.1\n3.5\n96.1\n1.6\n1.7\n\n11.0\n0.9\n0\n2.2\n95.6\n85.6\n\n0\n0\n0\n0.7\n1.4\n12.4\n\nTable 2: Percentage confusion matrix for an HMM trained on the wavelet encoded ECG.\n\n4.1 Wavelet Encoding of ECG\n\nWavelets are a class of functions that possess compact support and form a basis for all\n\ufb01nite energy signals. They are able to capture the non-stationary spectral characteristics\nof a signal by decomposing it over a set of atoms which are localised in both time and\nfrequency. These atoms are generated by scaling and translating a single mother wavelet.\n\nThe most popular wavelet transform algorithm is the discrete wavelet transform (DWT),\nwhich uses the set of dyadic scales (i.e. those based on powers of two) and translates of\nthe mother wavelet to form an orthonormal basis for signal analysis. The DWT is therefore\nmost suited to applications such as data compression where a compact description of a\nsignal is required. An alternative transform is derived by allowing the translation parameter\nto vary continuously, whilst restricting the scale parameter to a dyadic scale (thus, the\nset of time-frequency atoms now forms a frame). This leads to the undecimated wavelet\ntransform6 (UWT), which for a signal s 2 L2(R), is given by:\n\nw(cid:29)((cid:28) ) =\n\n1\n\np(cid:29) Z +1\n\n(cid:0)1\n\ns(t)  (cid:3)(cid:18) t (cid:0) (cid:28)\n\n(cid:29) (cid:19) dt\n\n(cid:29) = 2k; k 2 Z; (cid:28) 2 R\n\n(2)\n\nwhere w(cid:29)((cid:28) ) are the UWT coef\ufb01cients at scale (cid:29) and shift (cid:28) , and   (cid:3) is the complex con-\njugate of the mother wavelet. In practice the UWT can be computed in O(N log N ) using\nfast \ufb01lter bank algorithms [6].\n\nThe UWT is particularly well-suited to ECG interval analysis as it provides a time-\nfrequency description of the ECG signal on a sample-by-sample basis. In addition, the\nUWT coef\ufb01cients are translation-invariant (unlike the DWT coef\ufb01cients), which is impor-\ntant for pattern recognition applications.\n\nIn order to \ufb01nd the most effective wavelet basis for our application, we examined the per-\nformance of HMMs trained on ECG data encoded with wavelets from the Daubechies,\nSymlet, Coi\ufb02et and Biorthogonal wavelet families. In the frequency domain, a wavelet at\na given scale is associated with a bandpass \ufb01lter7 of a particular centre frequency. Thus\nthe optimal wavelet basis will correspond to the set of bandpass \ufb01lters that are tuned to the\nunique spectral characteristics of the ECG.\n\nIn our experiments we found that the Coi\ufb02et wavelet with two vanishing moments resulted\nin the highest overall classi\ufb01cation accuracy. Table 2 shows the results for this wavelet.\nIt is evident that the UWT encoding results in a signi\ufb01cant improvement in classi\ufb01cation\naccuracy (for all but the U wave state), when compared with the results obtained on the raw\nECG data.\n\n6The undecimated wavelet transform is also known as the stationary wavelet transform and the\n\ntranslation-invariant wavelet transform.\n\n7These \ufb01lters satisfy a constant relative bandwidth property, known as \u201cconstant-Q\u201d.\n\n\f0.03\n\n0.02\n\n0.01\n\n0\n\n0\n\nP wave\n\nTrue\nModel\n\n50\n\n100\n\n150\n\n200\n\nState duration (ms)\n\n0.04\n\n0.035\n\n0.03\n\n0.025\n\n0.02\n\n0.015\n\n0.01\n\n0.005\n\n0\n\n0\n\nQRS complex\n\nTrue\nModel\n\nT wave\n\nTrue\nModel\n\n0.014\n\n0.012\n\n0.01\n\n0.008\n\n0.006\n\n0.004\n\n0.002\n\n100\n50\nState duration (ms)\n\n150\n\n0\n\n0\n\n100\n\n200\n\n300\n\n400\n\nState duration (ms)\n\nFigure 2: Histograms of the true state durations and those decoded by the HMM.\n\n4.2 HMM State Durations\n\nA signi\ufb01cant limitation of the standard hidden Markov model is the manner in which it\nmodels state durations. For a given state i with self-transition coef\ufb01cient aii, the probability\ndensity of the state duration d is a geometric distribution, given by:\n\npi(d) = (aii)d(cid:0)1(1 (cid:0) aii)\n\n(3)\n\nFor the waveform features of the ECG signal, this geometric distribution is inappropriate.\nFigure 2 shows histograms of the true state durations and the durations of the states decoded\nby the HMM, for each of the P wave, QRS complex and T wave states. In each case it\nis clear that a signi\ufb01cant number of decoded states have a duration that is much shorter\nthan the minimum state duration observed with real ECG signals. Thus for a given ECG\nwaveform the decoded state sequence may contain many more state transitions than are\nactually present in the signal. The resulting HMM state segmentation is then likely to be\npoor and the resulting QT and PR interval measurements unreliable.\n\nOne solution to this problem is to post-process the decoded state sequences using a median\n\ufb01lter designed to smooth out sequences whose duration is known to be physiologically\nimplausible. A more principled and more effective approach, however, is to model the\nprobability density of the individual state durations explicitly, using a hidden semi-Markov\nmodel.\n\n5 A Hidden Semi-Markov Model for ECG Interval Analysis\n\nA hidden semi-Markov model (HSMM) differs from a standard HMM in that each of the\nself-transition coef\ufb01cients aii are set to zero, and an explicit probability density is speci\ufb01ed\nfor the duration of each state [5]. In this way, the individual state duration densities govern\nthe amount of time the model spends in a given state, and the transition matrix governs\nthe probability of the next state once this time has elapsed. Thus the underlying stochastic\nprocess is now a \u201csemi-Markov\u201d process.\nTo model the durations pi(d) of the various waveform features of the ECG, we used a\nGamma density since this is a positive distribution which is able to capture the inherent\nskewness of the ECG state durations. For each state i, maximum likelihood estimates of\nthe shape and scale parameters were computed directly from the set of labelled ECG signals\n(as part of the cross-validation procedure).\nIn order to infer the most probable state sequence Q = fq1q2 (cid:1)(cid:1)(cid:1) qTg for a given obser-\nvation sequence O = fO1O2 (cid:1)(cid:1)(cid:1) OTg, the standard Viterbi algorithm must be modi\ufb01ed to\n\n\f0.03\n\n0.02\n\n0.01\n\n0\n\n0\n\nP wave\n\nTrue\nModel\n\n50\n\n100\n\n150\n\n200\n\nState duration (ms)\n\n0.04\n\n0.035\n\n0.03\n\n0.025\n\n0.02\n\n0.015\n\n0.01\n\n0.005\n\n0\n\n0\n\nQRS complex\n\nTrue\nModel\n\nT wave\n\nTrue\nModel\n\n0.014\n\n0.012\n\n0.01\n\n0.008\n\n0.006\n\n0.004\n\n0.002\n\n100\n50\nState duration (ms)\n\n150\n\n0\n\n0\n\n100\n\n200\n\n300\n\n400\n\nState duration (ms)\n\nFigure 3: Histograms of the true state durations and those decoded by the HSMM.\n\nhandle the explicit state duration densities of the HSMM. We start by de\ufb01ning the likeli-\nhood of the most probable state sequence that accounts for the \ufb01rst t observations and ends\nin state i:\n\n(cid:14)t(i) = max\n\nq1q2 (cid:1)(cid:1)(cid:1)qt(cid:0)1\n\np(q1q2 (cid:1)(cid:1)(cid:1) qt = i; O1O2 (cid:1)(cid:1)(cid:1) Otj(cid:21))\n\nwhere (cid:21) is the set of parameters governing the HSMM. The recurrence relation for com-\nputing (cid:14)t(i) is then given by:\n\n(4)\n\n(5)\n\n(cid:14)t(i) = max\n\ndi nmax\n\nj (cid:8)(cid:14)t(cid:0)di (j)aji(cid:9) pi(di) (cid:5)t\n\nt0=t(cid:0)di+1bi(Ot0 )o\n\nwhere the outer maximisation is performed over all possible values of the state duration di\nfor state i, and the inner maximisation is over all states j. At each time t and for each state\ni, the two arguments that maximise equation (5) are recorded, and a simple backtracking\nprocedure can then be used to \ufb01nd the most probable state sequence.\n\nThe time complexity of the Viterbi decoding procedure for an HSMM is given by\nO(K 2 T Dmax), where K is the total number of states, and Dmax is the maximum range\nof state durations over all K states, i.e. Dmax = maxi(max(di) (cid:0) min(di)). As noted\nin [7], scaling the computation of (cid:14)t(i) to avoid under\ufb02ow is non-trivial. However, by\nsimply computing log (cid:14)t(i) it is possible to avoid any numerical problems.\nFigure 3 shows histograms of the resulting state durations for an HSMM trained on a\nwavelet encoding of the ECG (using 5-fold cross-validation). Clearly, the durations of\nthe decoded state sequences are very well matched to the true durations of each of the\nECG features. This improvement in duration modelling is re\ufb02ected in the accuracy and\nrobustness of the segmentations produced by the HSMM.\n\nModel\nHMM on raw ECG\nHMM on wavelet encoded ECG\nHSMM on wavelet encoded ECG\n\nPon Q\n31\n157\n11\n12\n13\n3\n\nJ\n27\n20\n7\n\nTo(cid:11)\n139\n46\n12\n\nTable 3: Mean absolute segmentation errors (in milliseconds) for each of the models.\n\nTable 3 shows the mean absolute errors8 for the Pon, Q, J and To(cid:11) points, for each of the\nmodels discussed. On the important task of accurately determining the Q and To(cid:11) points\nfor QT interval measurements, the HSMM signi\ufb01cantly outperforms the HMM.\n\n8The error was taken to be the time difference from the \ufb01rst decoded segment boundary to the\n\ntrue segment boundary (of the same type).\n\n\f6 Discussion\n\nIn this work we have focused on the two core issues in developing an automated system for\nECG interval analysis: the choice of representation for the ECG signal and the choice of\nmodel for the segmentation. We have demonstrated that wavelet methods, and in particular\nthe undecimated wavelet transform, can be used to generate an encoding of the ECG which\nis tuned to the unique spectral characteristics of the ECG waveform features. With this rep-\nresentation the performance of the models on new unseen ECG waveforms is signi\ufb01cantly\nbetter than similar models trained on the raw time series data. We have also shown that the\nrobustness of the segmentation process can be improved through the use of explicit state\nduration modelling with hidden semi-Markov models. With these models the detection ac-\ncuracy of the Q and To(cid:11) points compares favourably with current methods for automated\nQT analysis [3, 2].\n\nA key advantage of probabilistic models over traditional threshold-based methods for ECG\nsegmentation is that they can be used to generate a con\ufb01dence measure for each segmented\nECG signal. This is achieved by considering the log likelihood of the observed signal\ngiven the model, i.e. log p(Oj(cid:21)), which can be computed ef\ufb01ciently for both HMMs and\nHSMMs. Given this con\ufb01dence measure, it should be possible to determine a suitable\nthreshold for rejecting ECG signals which are either too noisy or too corrupted to provide\nreliable estimates of the QT and PR intervals. The robustness with which we can detect\nsuch unreliable QT interval measurements based on this log likelihood score is one of the\nmain focuses of our current research.\n\nAcknowledgements\n\nWe thank Cardio Analytics Ltd for help with data collection and labelling, and Oxford\nBioSignals Ltd for funding this research. NH thanks Iead Rezek for many useful discus-\nsions, and the anonymous reviewers for their helpful comments.\n\nReferences\n\n[1] M. A. T. Figueiredo and A. K. Jain. Unsupervised learning of \ufb01nite mixture models.\n\nTransactions on Pattern Analysis and Machine Intelligence, 24(3):381\u2013396, 2002.\n\nIEEE\n\n[2] S. Graja and J. M. Boucher. Multiscale hidden Markov model applied to ECG segmentation. In\nWISP 2003: IEEE International Symposium on Intelligent Signal Processing, pages 105\u2013109,\nBudapest, Hungary, 2003.\n\n[3] R. Jan\u00b4e, A. Blasi, J. Garc\u00b4ia, and P. Laguna. Evaluation of an automatic threshold based detector\nof waveform limits in Holter ECG with QT database. In Computers in Cardiology, pages 295\u2013\n298. IEEE Press, 1997.\n\n[4] A. Koski. Modelling ECG signals with hidden Markov models. Arti\ufb01cial Intelligence in\n\nMedicine, 8:453\u2013471, 1996.\n\n[5] S. E. Levinson. Continuously variable duration hidden Markov models for automatic speech\n\nrecognition. Computer Speech and Language, 1(1):29\u201345, 1986.\n\n[6] S. Mallat. A Wavelet Tour of Signal Processing. Academic Press, 2nd edition, 1999.\n[7] K. P. Murphy. Hidden semi-Markov models. Technical report, MIT AI Lab, 2002.\n[8] C. M. Pratt and S. Ruberg. The dose-response relationship between Terfenadine (Seldane) and\nthe QTc interval on the scalar electrocardiogram in normals and patients with cardiovascular\ndisease and the QTc interval variability. American Heart Journal, 131(3):472\u2013480, 1996.\n\n[9] L. R. Rabiner. A tutorial on hidden Markov models and selected applications in speech recogni-\n\ntion. Proceedings of the IEEE, 77(2):257\u2013286, 1989.\n\n\f", "award": [], "sourceid": 2347, "authors": [{"given_name": "Nicholas", "family_name": "Hughes", "institution": null}, {"given_name": "Lionel", "family_name": "Tarassenko", "institution": null}, {"given_name": "Stephen", "family_name": "Roberts", "institution": null}]}