- Email: [email protected]
Automatic cardiac rhythm interpretation during resuscitation
Resuscitation 102 (2016) 44–50
Contents lists available at ScienceDirect
Resuscitation journal homepage: www.elsevier.com/locate/resuscitation
Clinical paper
Automatic cardiac rhythm interpretation during resuscitation夽 Ali Bahrami Rad a,b,∗ , Kjersti Engan a , Aggelos K. Katsaggelos b , Jan Terje Kvaløy c , Lars Wik d , Jo Kramer-Johansen d , Unai Irusta e , Trygve Eftestøl a a
Department of Electrical Engineering and Computer Science, University of Stavanger, 4036 Stavanger, Norway Department of Electrical Engineering and Computer Science, Northwestern University, Evanston, IL 60208, USA Department of Mathematics and Natural Sciences, University of Stavanger, 4036 Stavanger, Norway d Norwegian National Advisory Unit on Prehospital Emergency Medicine (NAKOS) and Department of Anaesthesiology, Oslo University Hospital and University of Oslo, Pb 4956 Nydalen, 0424 Oslo, Norway e Communications Engineering Department, University of the Basque Country UPV/EHU, Alameda Urquijo S/N, 48013 Bilbao, Spain b c
a r t i c l e
i n f o
Article history: Received 3 September 2015 Received in revised form 27 December 2015 Accepted 15 January 2016 Keywords: Cardiopulmonary resuscitation Cardiac rhythm interpretation Feature extraction Feature selection Classification
a b s t r a c t Aim: Resuscitation guidelines recommend different treatments depending on the patient’s cardiac rhythm. Rhythm interpretation is a key tool to retrospectively evaluate and improve the quality of treatment. Manual rhythm annotation is time consuming and an obstacle for handling large resuscitation datasets efficiently. The objective of this study was to develop a system for automatic rhythm interpretation by using signal processing and machine learning algorithms. Methods: Data from 302 out of hospital cardiac arrest patients were used. In total 1669 3-second artifact free ECG segments with clinical rhythm annotations were extracted. The proposed algorithms combine 32 features obtained from both wavelet- and time-domain representations of the ECG, followed by a feature selection procedure based on the wrapper method in a nested cross-validation architecture. Linear and quadratic discriminant analyses (LDA and QDA) were used to automatically classify the segments into one of five rhythm types: ventricular tachycardia (VT), ventricular fibrillation (VF), pulseless electrical activity (PEA), asystole (AS), and pulse generating rhythms (PR). Results: The overall accuracy for the best algorithm was 68%. VT, VF, and AS are recognized with sensitivities of 71%, 75%, and 79%, respectively. Sensitivities for PEA and PR were 55% and 56%, respectively, which reflects the difficulty of identifying pulse using only the ECG. Conclusions: An ECG based automatic rhythm interpreter for resuscitation has been demonstrated. The interpreter handles VT, VF and AS well, while PEA and PR discrimination poses a more difficult problem. © 2016 Elsevier Ireland Ltd. All rights reserved.
Introduction Systematic review in order to identify specific factors during resuscitation episodes is a key to identify different quality aspects of therapy. Identifying such factors enables comparison of different therapeutic approaches, e.g., chest compression strategies, effect of interruptions, shock success, etc.1–6 The process seeks to determine the relationship between therapy and the patient’s response captured from the time series data (TSD) recorded by external defibrillators. Thoracic impedance and accelerometer signals are examples of TSD from which information about chest compressions
夽 A Spanish translated version of the summary of this article appears as Appendix in the final online version at http://dx.doi.org/10.1016/j.resuscitation.2016.01.015. ∗ Corresponding author at: Department of Electrical Engineering and Computer Science, University of Stavanger, 4036 Stavanger, Norway. E-mail address: [email protected] (A.B. Rad). http://dx.doi.org/10.1016/j.resuscitation.2016.01.015 0300-9572/© 2016 Elsevier Ireland Ltd. All rights reserved.
and ventilation can be extracted to compute quality indicators while the electrocardiogram (ECG) and end tidal CO2 (ETCO2 ) provide information about the patient’s response.1–7 Manual rhythm annotation is a time consuming task and an obstacle for handling large datasets efficiently. Eftestøl et al.8 demonstrated recently how rhythm state and chest compression sequence annotations could be used as a basis from which higher-level review parameters can be derived automatically. In addition, Ayala et al.9 demonstrated automatic techniques to determine chest compressions, and Irusta et al.10 proposed an algorithms for cardiopulmonary resuscitation (CPR) artifact removal. In this work, we propose an algorithm for automatic rhythm interpretation. In a final stage, all these subsystems will be linked together to produce a fully automatic system for resuscitation episode review. For a detailed account of the patient’s state/response, rhythm interpretation requires a more detailed rhythm classification than that provided by shock advice algorithms.11,12 It generally involves rhythm classification into five categories: ventricular tachycardia
A.B. Rad et al. / Resuscitation 102 (2016) 44–50
45
Fig. 1. Denoising process of two noisy cardiac rhythms VF and AS. In all plots the y axes are the amplitudes of the ECG in mV which are shown in the different ranges for better representation. In the top figure a noisy VF rhythm (baseline wander or low-frequency noise) is denoised using multiresolution analysis. rA8 is the reconstructed signal from approximation coefficients of level 8. rD1 to rD8 are the reconstructed signal from detail coefficients of levels 1–8. Low-frequency noise is represented in rA8. In the bottom figure a noisy AS rhythm (with both low-frequency and high-frequency noise) is denoised. High-frequency noise is represented in rD3, rD2 and rD1. Again, low-frequency noise is represented in rA8.
(VT), ventricular fibrillation (VF), pulseless electrical activity (PEA), pulse generating rhythms (PR) and asystole (AS). In the present study, we build on earlier attempts,13,14 developing a generalizable algorithm with a specific focus on feature extraction, feature selection, and model assessment. Materials and methods ECG dataset We extracted ECG data from 302 out-of-hospital cardiac arrest (OHCA) patient records. The original study was done to measure CPR quality in three geographic locations Akershus (Norway),
Stockholm (Sweden), and London (UK) between March 2002 and September 2004.2,15 The surface ECG was recorded by modified Heartstart 4000 defibrillators with enhanced monitoring capabilities. The sampling rate was 500 Hz with a resolution of 1.031 V per least significant bit, which is equivalent or superior to that of current state of the art defibrillators. For the original studies, recordings were annotated by expert reviewers using the five rhythm types (VT, VF, PEA, PR, AS), and chest compression intervals were annotated using the compression depth available from a CPR assist-pad. ECG segments were automatically extracted based on these annotations with the following criteria: 3-second duration, a single rhythm, and no chest compression artifacts.
46
A.B. Rad et al. / Resuscitation 102 (2016) 44–50
Table 1 Results of the proposed methods (LDA + SFS, LDA + SBS, QDA + SFS, and QDA + SBS). Proposed algorithm LDA + SFS Sen (%) Spe (%) PPV (%) Sen (%) Spe (%) PPV (%) Sen (%) Spe (%) PPV (%) Sen (%) Spe (%) PPV (%) Sen (%) Spe (%) PPV (%)
75.4 (4.1) 86.8 (5.3) AS 66.2 (10.2) 52.1 (5.2) 88.5 (3.4) PEA 57.0 (9.9) 52.7 (13.2) 94.4 (2.4) PR 63.8 (13.7) 72.9 (5.8) 91.1 (4.6) VF 72.9 (8.2) 65.9 (29.0) 95.4 (2.1) VT 61.1 (25.5) 65.7 (4.6) tAcc (%) Number of features 13.2 (2.8)
LDA + SBS
QDA + SFS
QDA + SBS
78.8 (9.7) 87.4 (6.3) 69.1 (8.7) 55.1 (9.4) 89.6 (4.4) 61.0 (11.4) 56.1 (12.6) 95.1 (2.6) 68.6 (7.7) 74.7 (8.6) 91.1 (4.6) 73.0 (12.7) 71.2 (29.0) 95.5 (2.0) 61.1 (29.5) 67.9 (6.3) 27.1 (2.2)
79.8 (7.5) 87.6 (3.4) 68.1 (7.6) 46.2 (11.0) 91.4 (2.7) 60.7 (10.8) 55.0 (9.7) 92.4 (3.9) 59.4 (13.4) 73.5 (13.5) 93.1 (4.2) 78.2 (10.6) 74.0 (22.0) 93.1 (3.2) 57.8 (20.4) 66.4 (4.1) 7.0 (1.5)
78.7 (5.9) 89.9 (2.6) 71.9 (8.0) 45.6 (8.6) 92.1 (2.3) 62.3 (10.4) 64.7 (10.5) 90.7 (4.1) 57.4 (12.4) 72.2 (8.1) 92.0 (2.6) 74.1 (7.3) 67.3 (18.2) 93.1 (1.7) 53.9 (22.2) 66.4 (4.3) 26.4 (2.4)
The numbers are the unweighted average of the results of 10-fold external CV loop (see Appendix A); the numbers in parentheses are their standard deviations.
Proposed method The following is an abridged description of the methods (see Appendix A for a detailed technical description). ECG data was denoised and filtered using discrete wavelet transforms (DWT) to remove low-frequency interferences such as baseline wander and high-frequency noise, as shown in Fig. 1. The DWT and time-domain representations of the ECG were used to extract 32 features. These features were fed into linear and quadratic discriminant analysis (LDA and QDA) classifiers. Algorithm performance was assessed and features were selected using a patient-wise nested crossvalidation (CV) scheme. Two suboptimal methods were tested for feature selection for each classifier; sequential forward selection (SFS) and sequential backward selection (SBS). The combination of classifiers and feature selection methods resulted in four algorithms. Evaluation of the performance The performance of the algorithms was evaluated using sensitivity (Sen), specificity (Spe), and positive predictive value (PPV). These metrics are usually defined for binary classification, so the definitions were adapted to our five class problem using multi-class confusion matrices.16 In summary, once all segments are labeled into one of the classes by the algorithm the procedure reduces to treating each rhythm type against the rest as a binary decision. Then, the true positive rate (Sen), the true negative rate (Spe), and the positive predictive value (PPV) are determined for each class. In addition, total accuracy (tAcc), the proportion of correct decisions for all classes, was used as a global measure of performance. In calculation of tAcc, the sum of the entries on the main diagonal of the confusion matrix (correct decisions) is divided by the sum of the all entries (all decisions). Results A total of 1669 segments were extracted from the ECG dataset, 411 AS (n = 225 patients), 369 PEA (n = 205), 264 PR (n = 111), 397 VF (n = 162), and 228 VT (n = 29). The results for the different classifiers are shown in Table 1 for all the rhythm classes. In addition, tAcc is reported as an overall measure of the performance of each method, together with the average numbers of selected features.
A detailed analysis of the classification is shown in Table 2, which shows the confusion matrix for the algorithm with the highest tAcc (LDA + SBS). This confusion matrix provides all the information needed to calculate correct/incorrect classification rates for each rhythm type in detail. To make it clearer, beside each number in the confusion matrix there is another number in parenthesis which shows the classification/misclassification rate. To calculate those numbers the entries of each row of the confusion matrix is normalized. Thus, the percentage of correct and incorrect classification for each rhythm can be easily found. For example, the first row of this table shows 78.8% of AS rhythms are correctly classified as AS, but 12.9%, 0.7%, 7.1%, and 0.5% are incorrectly classified as PEA, PR, VF, and VT. As shown in Table 2 most classification errors correspond to decisions between PEA/PR, VF/AS and PEA/AS rhythms. Fig. 2 shows one example of the ECG for each entry of the confusion matrix of Table 2, where the rows show the labels assigned by the expert reviewers and the columns show the classification by the proposed algorithm. Discussion We have designed and demonstrated an automatic ECG based interpreter of resuscitation rhythms into five categories. These algorithms go beyond the binary shock/no-shock decision in defibrillators, and set the ground for the development of automatic methods for ECG interpretation with the level of detail required by an automated review system. Comparative evaluations of automatic interpretation of different cardiac rhythms The total accuracy of the algorithms introduced in this work is in the order of 70% which could be sufficient as a pre-screening tool in a semi-automatic rhythm interpreter. However, for quality assurance a better accuracy is required, and that is part of our ongoing research following the approaches/solutions proposed in this section. To use the current version of the algorithm, once it has labeled the data, an expert reviewer would have to go through the episodes in the final stage of the quality assurance of the annotations. But, his/her workload would be substantially reduced by the automatic annotations of our system, which would have to be only sporadically corrected. In the following, we discuss the performance of the algorithm in detail. The sensitivities reveal that the performance of all rhythm classes is in the same range (70% or higher) except for PEA and PR. The results underline the intrinsic difficulty of discriminating PEA from PR using only the ECG. Both PEA and PR are organized (ORG) cardiac rhythms, differentiated by the presence of a palpable pulse. Such a distinction is difficult even for experts,17 and involves the combined analysis of the ECG and impedance in automatic PEA/PR discrimination algorithms.18,19 The results from Table 2 show that when PEA and PR are grouped into a single ORG class the sensitivity for ORG rhythms rises to 75%. Future developments may then include the combined analysis of the ECG, impedance data, and/or ETCO2 to discriminate PEA from PR;18,19 or, in the context of a fully automated review system, to combine the analyses of the algorithms with clinical ROSC annotations made on site.8 Inspection of the data in Table 2, Figs. 2 and 3 reveals the main sources of misclassification. Errors are due to borderline rhythms such as: VF with low amplitude and dominant frequency (AS/VF discrimination), bradycardic rhythms (AS/PEA) and tachycardias of either ventricular or supraventricular origin (VT/PEA and VT/PR). Fig. 2 shows examples of these cases. There were also errors caused by transitional states, as shown by the example in Fig. 3(a), or by
A.B. Rad et al. / Resuscitation 102 (2016) 44–50
47
Table 2 The confusion matrix of the LDA + SBS classification algorithm; the numbers in parentheses show the classification/misclassification rate. LDA + SBS algorithm label
Expert reviewers’ label
AS PEA PR VF VT
AS
PEA
PR
VF
VT
324 (78.8%) 70 (19.0%) 14 (5.3%) 61 (15.4%) 4 (1.8%)
53 (12.9%) 206 (55.8%) 62 (23.5%) 11 (2.8%) 8 (3.5%)
3 (0.7%) 58 (15.7%) 147 (55.7%) 4 (1.0%) 3 (1.3%)
29 (7.1%) 24 (6.5%) 15 (5.7%) 295 (74.3%) 46 (20.2%)
2 (0.5%) 11 (3.0%) 26 (9.8%) 26 (6.5%) 167 (73.2%)
Fig. 2. Different examples of classified rhythms using the LDA + SBS algorithm. The rows show the labels assigned by the expert reviewers, and the columns show the classification by the proposed algorithm. The x and y axes of each plot are time (in s) and amplitudes (in mV).
rhythms analyzed immediately after stopping chest compressions, as shown in Fig. 3(b). A possible way to address these misclassifications would be to classify consecutive windows by using two layers of classifiers.20 The first layer classifies ECG in the signal level, as introduced in the present study, and generates the state level labels. The second level classifies a sequence of the state level labels to the final labels. Some mismatches are due to a few inconsistencies in the original annotations, both in the rhythm labels and in the chest compression intervals. For instance the rhythm labeled as VT and classified as AS in Fig. 2, also shown in Fig. 3(c), is an example of a mislabeled rhythm, and the rhythm labeled as AS and classified as VT in Fig. 2 corresponds to an asystole during chest compressions, as shown in Fig. 3(d). Although these instances are very few in our data, their presence confirms the importance of quality control in the review process of resuscitation data.
Finally, we emphasize that at this stage our goal was to classify rhythms for retrospective analysis purposes, and it is not intended to be used for guiding therapy. It is obvious that if we include further considerations like resuscitation action, we can optimize our algorithms for that specific application. For example, in this work the costs of misclassification are the same for all rhythms, however we can easily change the cost function for a specific rhythm in our algorithm. Does ECG carry any information which is useful for PEA/PR discrimination? It is a common belief that the presence of pulse cannot be identified in the ECG. Mathematically speaking, it means that in the absence of pulse information for PEA/PR discrimination we cannot have any better results than the random guess. However, we
48
A.B. Rad et al. / Resuscitation 102 (2016) 44–50
Fig. 3. Demonstration of some problematic segments in the classification process. The gray arrows indicate the labels assigned by experts (C-PR stands for PR rhythm during chest compression). The red dashed circles show the considered segments and their position in Fig. 2. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
showed that using signal processing and machine learning methods it is possible to classify PEA/PR from only ECG signal to some degree. Table 2 shows that 55.8% of PEA rhythms were correctly classified, but only 15.7% were incorrectly classified as PR. These two numbers should be almost equal if ECG does not carry any information useful for PEA/PR discrimination. The same argument can be used for PR rhythms where 55.7% of them were correctly classified, but 23.5% were incorrectly classified as PEA. This indicates that the ECG signal carries some information or patterns, which could be hidden from human eyes, and can be utilized for PEA/PR discrimination.
Comparative evaluations of the performance of the present and previous methods All the results with respect to the total accuracy (tAcc) have decreased in relation to our prior works. There could be different reasons for this. Firstly, in the new settings we kept track of patients to be sure that the data of each patient was used either only in the training or only in the testing datasets. In the previous works, the training and testing datasets were disjoint, but patients were not audited. Therefore, data from the same patient could appear in both training and testing sets, and thus there could be a data leakage of patients from the test data into the training data. Data leakage, which is among the top 10 machine learning and data mining mistakes,21 can result in unrealistically good performance of machine learning algorithms.22 Secondly, since in the previous studies there was no cross-validation step, the models could be biased on the data. Thirdly, in our previous studies the data was highly imbalanced. For imbalanced datasets tAcc is a biased estimate of performance.16 For example, while tAcc was 79% for Klocal hyperplane distance nearest-neighbor (HKNN) classifier and 76% for K-nearest neighbor (KNN) classifier, the Sen of VT rhythms were 10% and 0%, respectively.14 In this study we used an almost
balanced dataset, so while tAcc was 68% for LDA + SBS, Sen of VT improved to 71%. Comparative evaluations of the performance of the proposed classifiers The best result in the sense of overall accuracy belongs to LDA + SBS (Table 1). This may be surprising since the QDA classifier has greater flexibility than the LDA classifier. However, since the number of estimated parameters is larger in QDA than in LDA, in general QDA classifiers require larger training datasets.23 In addition, it seems that QDA is more sensitive to the violation of the basic assumptions like Gaussian distribution of the class conditional densities.24 Toward automated review International collaboration with treatment recommendations relies heavily on comparison between randomized controlled trials and evaluation of registry data.25,26 Still, unresolved issues remain regarding relative importance of the different aspects of treatment (chest compressions and pauses in chest compressions, ventilations, defibrillation attempts, medications) and their relationship in the treatment recommendations. These issues have been identified as knowledge gaps in resuscitation science by the International Liaison Committee on Resuscitation.27 Eftestøl et al.8 previously demonstrated how a large number of parameters registered during the review of resuscitation episodes could be automatically replicated from a minimal set of basic parameters describing the resuscitation episode. These basic parameters included start and end times of chest compressions, the electronic defibrillator event report, and annotations of the rhythm transition times. As automatic chest compression detection is possible,9 and the electronic event reports are commonly
A.B. Rad et al. / Resuscitation 102 (2016) 44–50
available, solving the problem of automatic rhythm interpretation will make it possible to do the review in a fully automatic fashion. The review process for quality assurance can be made more efficient as it will be possible to determine important indicators (e.g., time intervals) and quality parameters (e.g., the ratio of hands off intervals during treatable rhythms). This will be important for quality assurance and benchmarking within and between different sites as the same automatic review systems can be applied in different locations. It will also enable outcomes of clinical trials to be controlled for varying quality of basic treatment, and monitor and develop strategies for implementation of new treatment protocols and improvements in teaching/education.
49
Classification Features were fed into LDA and QDA classifiers, and in both cases the class conditional densities (CCD) were modeled with multivariate Gaussian distributions but with equal covariance matrices (homoscedasticity) in LDA and unequal covariance matrices (heteroscedasticity) in QDA. Mean and covariance matrices were estimated using maximum likelihood, and the pooled covariance matrix32 was computed for the LDA classifier. Bayes rule was applied for classification after modeling CCDs33 using uniform class priors. Model assessment and feature selection
Conclusion In this study, we have introduced a system for automatic rhythm interpretation of five cardiac rhythms (AS, PEA, PR, VF, VT) during resuscitation, exploring different feature extraction, feature selection, and classification methods in a systematic manner. Many previous studies of important treatment factors during resuscitation have relied on tedious manual interpretation of these five rhythms, and an automatic interpretation system would facilitate the handling of large resuscitation databases. With an overall accuracy of 68%, the current system has potential for improvement. Specific areas for improvement have been identified and are the focus of our ongoing work.
Conflict of interest statement The authors have no conflicts of interest.
Appendix A. Preprocessing and filtering ECG data was denoised and filtered in the wavelet domain where the ECG is represented as a summation of smooth and detailed parts in different scales.28,29 ECG signals were decomposed using an 8-level DWT with a daubechies-4 wavelet.30 Detail coefficients of levels 4–8 were used to reconstruct the denoised ECG, as shown in Fig. 1. In the 8-level DWT decomposition, the level 8 approximation coefficient is roughly related to frequencies below 1 Hz, responsible for low-frequency noise like baseline wander. Highfrequency noise (above 30 Hz) is suppressed by eliminating the detail coefficients in levels 1–3.
Feature extraction ECG features from both wavelet- and time-domain representations were used. In the wavelet domain, three statistical descriptors of the levels 4–8 detail coefficients were extracted: interquartile ranges (features 1–5), variances (6–10), and first quartiles (11–15). Morphological and time-domain features were extracted from the denoised ECG. Features 16–19 were the coefficients of the order 4 autoregressive (AR) model of the ECG, estimated using Burg’s method.31 The variance of the white noise input to the AR model was feature 20. The number of peaks of the autocorrelation above a threshold was feature 21. Then 8 statistical descriptors of the ECG, and its first and second derivatives were obtained (22–29).14 The last three features (30–32) were positive area (the sum of the positive signal values), negative area (the sum of the negative signal values), and amplitude range (the difference between the maximum and the minimum signal values).
A nested CV architecture34,35 was used, with a 10-fold patientwise external CV for an unbiased estimate of the performance (model assessment), and 5-fold internal CV for feature selection. The wrapper approach, in which the decision is based on the performance of the classifier, was adopted for feature selection.36,37 The total error (tErr = 1 − tAcc) was used to determine when to stop in the feature selection process. Two suboptimal feature selection approaches were explored, SBS38 where starting from the full set features are sequentially removed, and SFS39 where starting from an empty set features are sequentially added. After selecting the best feature subset in the internal CV loop, the resulting features were used to estimate the performance of the method. Moreover, when training the classifiers, the outliers of the training data were removed by three-sigma rule40 while the test data was untouched. References 1. Sunde K, Eftestøl T, Askenberg C, Steen PA. Quality assessment of defibrillation and advanced life support using data from the medical control module of the defibrillator. Resuscitation 1999;41:237–47. 2. Wik L, Kramer-Johansen J, Myklebust H, et al. Quality of cardiopulmonary resuscitation during out-of-hospital cardiac arrest. JAMA 2005;293:299–304. 3. Abella BS, Alvarado JP, Myklebust H, et al. Quality of cardiopulmonary resuscitation during in-hospital cardiac arrest. JAMA 2005;293:305–10. 4. Skogvoll E, Eftestøl T, Gundersen K, et al. Dynamics and state transitions during resuscitation in out-of-hospital cardiac arrest. Resuscitation 2008;78: 30–7. 5. Kvaløy JT, Skogvoll E, Eftestøl T, et al. Which factors influence spontaneous state transitions during resuscitation? Resuscitation 2009;80:863–9. 6. Nordseth T, Bergum D, Edelson DP, et al. Clinical state transitions during advanced life support (ALS) in in-hospital cardiac arrest. Resuscitation 2013;84:1238–44. 7. Qvigstad E, Kramer-Johansen J, Tømte Ø, et al. Clinical pilot study of different hand positions during manual chest compressions monitored with capnography. Resuscitation 2013;84:1203–7. 8. Eftestøl T, Sherman LD. Towards the automated analysis and database development of defibrillator data from cardiac arrest. BioMed Res Int 2014:276965:2014. 9. Ayala U, Eftestøl T, Alonso E, et al. Automatic detection of chest compressions for the assessment of CPR-quality parameters. Resuscitation 2014;85: 957–63. 10. Irusta U, Ruiz J, Ruiz de Gauna S, Eftestøl T, Kramer-Johansen J. A least mean-square filter for the estimation of the cardiopulmonary resuscitation artifact based on the frequency of the compressions. IEEE Trans Biomed Eng 2009;56:1052–62. 11. Irusta U, Ruiz J, Aramendi E, de Gauna SR, Ayala U, Alonso E. A high-temporal resolution algorithm to discriminate shockable from nonshockable rhythms in adults and children. Resuscitation 2012;83:1090–7. 12. Didon JP, Krasteva V, Ménétré S, Stoyanov T, Jekova I. Shock advisory system with minimal delay triggering after end of chest compressions: accuracy and gained hands-off time. Resuscitation 2011;82:S8–15. 13. Rad AB, Eftestøl T, Kvaløy JT, Ayala U, Kramer-Johansen J, Engan K. Probabilistic classification approaches for cardiac arrest rhythm interpretation during resuscitation. Comput Cardiol (CinC) 2013;40:125–8. 14. Rad AB, Eftestøl T, Kvaløy JT, Ayala U, Kramer-Johansen J, Engan K. Nearestmanifold classification approach for cardiac arrest rhythm interpretation during resuscitation. In: IEEE international conference on acoustics, speech and signal processing (ICASSP). 2014. p. 3621–5. 15. Kramer-Johansena J, Myklebust H, Wik L, et al. Quality of out-of-hospital cardiopulmonary resuscitation with real time automated feedback: a prospective interventional study. Resuscitation 2006;71:283–92.
50
A.B. Rad et al. / Resuscitation 102 (2016) 44–50
16. Mar T, Zaunseder S, Martínez JP, Llamedo M, Poll R. Optimization of ECG classification by means of feature selection. IEEE Trans Biomed Eng 2011;58:2168–77. 17. Eberle B, Dick WF, Schneider T, Wisser G, Doetsch S, Tzanova I. Checking the carotid pulse check: diagnostic accuracy of first responders in patients with and without a pulse. Resuscitation 1996;33:107–16. 18. Risdal M, Aase SO, Kramer-Johansen J, Eftestøl T. Automatic identification of return of spontaneous circulation during cardiopulmonary resuscitation. IEEE Trans Biomed Eng 2008;55:60–8. 19. Ruiz J, Ayala U, Ruiz de Gauna S, et al. Feasibility of automated rhythm assessment in chest compression pauses during cardiopulmonary resuscitation. Resuscitation 2013;84:1223–8. 20. Kwok H, Coult J, Drton M, Rea TD, Sherman L. Adaptive rhythm sequencing: a method for dynamic rhythm classification during CPR. Resuscitation 2015;91:26–31. 21. Nisbet R, Elder J, Miner G. Handbook of statistical analysis and data mining applications. Academic Press; 2009. 22. Kaufman S, Rosset S, Perlich C, Stitelman O. Leakage in data mining: formulation, detection, and avoidance. ACM Trans Knowl Discov Data 2012;6:15. 23. Wahl PW, Kronmal RA. Discriminant functions when covariances are unequal and sample sizes are moderate. Biometrics 1977;33:479–84. 24. Friedman JH. Regularized discriminant analysis. JASA 1989;84:165–75. 25. Nolan J, Baskett P, editors. European Resuscitation Council guidelines for resuscitation 2005. Resuscitation. 2005. p. 67. 26. Proceedings of the 2005 international consensus on cardiopulmonary resuscitation and emergency cardiovascular care science with treatment recommendations. Resuscitation 2005;67:157–341. 27. Gazmuri RJ, Nolan JP, Nadkarni VM, et al. Scientific knowledge gaps and clinical research priorities for cardiopulmonary resuscitation and emergency cardiovascular care identified during the 2005 International Consensus Conference
28. 29. 30. 31.
32. 33. 34. 35.
36. 37.
38. 39. 40.
on ECC and CPR Science with Treatment Recommendations. Resuscitation 2007;75:400–11. Mallat SG. Multiresolution approximations and wavelet orthonormal bases of L2 (R). Trans Am Math Soc 1989;315:69–87. Meyer Y, Salinger DH. Wavelets and operators. Cambridge: Cambridge University Press; 1995. Daubechies I. Ten lectures on wavelets. Philadelphia: SIAM; 1992. Burg JP. A new analysis technique for time series data. NATO Advanced Study Institute of Signal Processing with emphasis on Underwater Acoustics; 1968. p. 1. Legendre P, Legendre L. Numerical ecology. 2nd English ed. Amsterdam: Elsevier; 1998. Theodoridis S, Koutroumbas K. Pattern recognition. 4th ed. Academic Press; 2008. Varma S, Simon R. Bias in error estimation when using cross-validation for model selection. BMC Bioinf 2006;7:91. Krstajic D, Buturovic LJ, Leahy DE, Thomas S. Cross-validation pitfalls when selecting and assessing regression and classification models. J Cheminf 2014;6:10. Kohavi R, John GH. Wrappers for feature subset selection. Artif Intell 1997;97:273–324. John GH, Kohavi R, Pfleger K. Irrelevant features and the subset selection problem. In: Machine learning: proceedings of the eleventh international conference. 1994. p. 121–9. Marill T, Green DM. On the effectiveness of receptors in recognition systems. IEEE Trans Inf Theory 1963;9:11–7. Whitney AW. A direct method of nonparametric measurement selection. IEEE Trans Comput 1971;20:1100–3. Maronna RA, Martin RD, Yohai VJ. Robust statistics. John Wiley & Sons; 2006.
Contents lists available at ScienceDirect
Resuscitation journal homepage: www.elsevier.com/locate/resuscitation
Clinical paper
Automatic cardiac rhythm interpretation during resuscitation夽 Ali Bahrami Rad a,b,∗ , Kjersti Engan a , Aggelos K. Katsaggelos b , Jan Terje Kvaløy c , Lars Wik d , Jo Kramer-Johansen d , Unai Irusta e , Trygve Eftestøl a a
Department of Electrical Engineering and Computer Science, University of Stavanger, 4036 Stavanger, Norway Department of Electrical Engineering and Computer Science, Northwestern University, Evanston, IL 60208, USA Department of Mathematics and Natural Sciences, University of Stavanger, 4036 Stavanger, Norway d Norwegian National Advisory Unit on Prehospital Emergency Medicine (NAKOS) and Department of Anaesthesiology, Oslo University Hospital and University of Oslo, Pb 4956 Nydalen, 0424 Oslo, Norway e Communications Engineering Department, University of the Basque Country UPV/EHU, Alameda Urquijo S/N, 48013 Bilbao, Spain b c
a r t i c l e
i n f o
Article history: Received 3 September 2015 Received in revised form 27 December 2015 Accepted 15 January 2016 Keywords: Cardiopulmonary resuscitation Cardiac rhythm interpretation Feature extraction Feature selection Classification
a b s t r a c t Aim: Resuscitation guidelines recommend different treatments depending on the patient’s cardiac rhythm. Rhythm interpretation is a key tool to retrospectively evaluate and improve the quality of treatment. Manual rhythm annotation is time consuming and an obstacle for handling large resuscitation datasets efficiently. The objective of this study was to develop a system for automatic rhythm interpretation by using signal processing and machine learning algorithms. Methods: Data from 302 out of hospital cardiac arrest patients were used. In total 1669 3-second artifact free ECG segments with clinical rhythm annotations were extracted. The proposed algorithms combine 32 features obtained from both wavelet- and time-domain representations of the ECG, followed by a feature selection procedure based on the wrapper method in a nested cross-validation architecture. Linear and quadratic discriminant analyses (LDA and QDA) were used to automatically classify the segments into one of five rhythm types: ventricular tachycardia (VT), ventricular fibrillation (VF), pulseless electrical activity (PEA), asystole (AS), and pulse generating rhythms (PR). Results: The overall accuracy for the best algorithm was 68%. VT, VF, and AS are recognized with sensitivities of 71%, 75%, and 79%, respectively. Sensitivities for PEA and PR were 55% and 56%, respectively, which reflects the difficulty of identifying pulse using only the ECG. Conclusions: An ECG based automatic rhythm interpreter for resuscitation has been demonstrated. The interpreter handles VT, VF and AS well, while PEA and PR discrimination poses a more difficult problem. © 2016 Elsevier Ireland Ltd. All rights reserved.
Introduction Systematic review in order to identify specific factors during resuscitation episodes is a key to identify different quality aspects of therapy. Identifying such factors enables comparison of different therapeutic approaches, e.g., chest compression strategies, effect of interruptions, shock success, etc.1–6 The process seeks to determine the relationship between therapy and the patient’s response captured from the time series data (TSD) recorded by external defibrillators. Thoracic impedance and accelerometer signals are examples of TSD from which information about chest compressions
夽 A Spanish translated version of the summary of this article appears as Appendix in the final online version at http://dx.doi.org/10.1016/j.resuscitation.2016.01.015. ∗ Corresponding author at: Department of Electrical Engineering and Computer Science, University of Stavanger, 4036 Stavanger, Norway. E-mail address: [email protected] (A.B. Rad). http://dx.doi.org/10.1016/j.resuscitation.2016.01.015 0300-9572/© 2016 Elsevier Ireland Ltd. All rights reserved.
and ventilation can be extracted to compute quality indicators while the electrocardiogram (ECG) and end tidal CO2 (ETCO2 ) provide information about the patient’s response.1–7 Manual rhythm annotation is a time consuming task and an obstacle for handling large datasets efficiently. Eftestøl et al.8 demonstrated recently how rhythm state and chest compression sequence annotations could be used as a basis from which higher-level review parameters can be derived automatically. In addition, Ayala et al.9 demonstrated automatic techniques to determine chest compressions, and Irusta et al.10 proposed an algorithms for cardiopulmonary resuscitation (CPR) artifact removal. In this work, we propose an algorithm for automatic rhythm interpretation. In a final stage, all these subsystems will be linked together to produce a fully automatic system for resuscitation episode review. For a detailed account of the patient’s state/response, rhythm interpretation requires a more detailed rhythm classification than that provided by shock advice algorithms.11,12 It generally involves rhythm classification into five categories: ventricular tachycardia
A.B. Rad et al. / Resuscitation 102 (2016) 44–50
45
Fig. 1. Denoising process of two noisy cardiac rhythms VF and AS. In all plots the y axes are the amplitudes of the ECG in mV which are shown in the different ranges for better representation. In the top figure a noisy VF rhythm (baseline wander or low-frequency noise) is denoised using multiresolution analysis. rA8 is the reconstructed signal from approximation coefficients of level 8. rD1 to rD8 are the reconstructed signal from detail coefficients of levels 1–8. Low-frequency noise is represented in rA8. In the bottom figure a noisy AS rhythm (with both low-frequency and high-frequency noise) is denoised. High-frequency noise is represented in rD3, rD2 and rD1. Again, low-frequency noise is represented in rA8.
(VT), ventricular fibrillation (VF), pulseless electrical activity (PEA), pulse generating rhythms (PR) and asystole (AS). In the present study, we build on earlier attempts,13,14 developing a generalizable algorithm with a specific focus on feature extraction, feature selection, and model assessment. Materials and methods ECG dataset We extracted ECG data from 302 out-of-hospital cardiac arrest (OHCA) patient records. The original study was done to measure CPR quality in three geographic locations Akershus (Norway),
Stockholm (Sweden), and London (UK) between March 2002 and September 2004.2,15 The surface ECG was recorded by modified Heartstart 4000 defibrillators with enhanced monitoring capabilities. The sampling rate was 500 Hz with a resolution of 1.031 V per least significant bit, which is equivalent or superior to that of current state of the art defibrillators. For the original studies, recordings were annotated by expert reviewers using the five rhythm types (VT, VF, PEA, PR, AS), and chest compression intervals were annotated using the compression depth available from a CPR assist-pad. ECG segments were automatically extracted based on these annotations with the following criteria: 3-second duration, a single rhythm, and no chest compression artifacts.
46
A.B. Rad et al. / Resuscitation 102 (2016) 44–50
Table 1 Results of the proposed methods (LDA + SFS, LDA + SBS, QDA + SFS, and QDA + SBS). Proposed algorithm LDA + SFS Sen (%) Spe (%) PPV (%) Sen (%) Spe (%) PPV (%) Sen (%) Spe (%) PPV (%) Sen (%) Spe (%) PPV (%) Sen (%) Spe (%) PPV (%)
75.4 (4.1) 86.8 (5.3) AS 66.2 (10.2) 52.1 (5.2) 88.5 (3.4) PEA 57.0 (9.9) 52.7 (13.2) 94.4 (2.4) PR 63.8 (13.7) 72.9 (5.8) 91.1 (4.6) VF 72.9 (8.2) 65.9 (29.0) 95.4 (2.1) VT 61.1 (25.5) 65.7 (4.6) tAcc (%) Number of features 13.2 (2.8)
LDA + SBS
QDA + SFS
QDA + SBS
78.8 (9.7) 87.4 (6.3) 69.1 (8.7) 55.1 (9.4) 89.6 (4.4) 61.0 (11.4) 56.1 (12.6) 95.1 (2.6) 68.6 (7.7) 74.7 (8.6) 91.1 (4.6) 73.0 (12.7) 71.2 (29.0) 95.5 (2.0) 61.1 (29.5) 67.9 (6.3) 27.1 (2.2)
79.8 (7.5) 87.6 (3.4) 68.1 (7.6) 46.2 (11.0) 91.4 (2.7) 60.7 (10.8) 55.0 (9.7) 92.4 (3.9) 59.4 (13.4) 73.5 (13.5) 93.1 (4.2) 78.2 (10.6) 74.0 (22.0) 93.1 (3.2) 57.8 (20.4) 66.4 (4.1) 7.0 (1.5)
78.7 (5.9) 89.9 (2.6) 71.9 (8.0) 45.6 (8.6) 92.1 (2.3) 62.3 (10.4) 64.7 (10.5) 90.7 (4.1) 57.4 (12.4) 72.2 (8.1) 92.0 (2.6) 74.1 (7.3) 67.3 (18.2) 93.1 (1.7) 53.9 (22.2) 66.4 (4.3) 26.4 (2.4)
The numbers are the unweighted average of the results of 10-fold external CV loop (see Appendix A); the numbers in parentheses are their standard deviations.
Proposed method The following is an abridged description of the methods (see Appendix A for a detailed technical description). ECG data was denoised and filtered using discrete wavelet transforms (DWT) to remove low-frequency interferences such as baseline wander and high-frequency noise, as shown in Fig. 1. The DWT and time-domain representations of the ECG were used to extract 32 features. These features were fed into linear and quadratic discriminant analysis (LDA and QDA) classifiers. Algorithm performance was assessed and features were selected using a patient-wise nested crossvalidation (CV) scheme. Two suboptimal methods were tested for feature selection for each classifier; sequential forward selection (SFS) and sequential backward selection (SBS). The combination of classifiers and feature selection methods resulted in four algorithms. Evaluation of the performance The performance of the algorithms was evaluated using sensitivity (Sen), specificity (Spe), and positive predictive value (PPV). These metrics are usually defined for binary classification, so the definitions were adapted to our five class problem using multi-class confusion matrices.16 In summary, once all segments are labeled into one of the classes by the algorithm the procedure reduces to treating each rhythm type against the rest as a binary decision. Then, the true positive rate (Sen), the true negative rate (Spe), and the positive predictive value (PPV) are determined for each class. In addition, total accuracy (tAcc), the proportion of correct decisions for all classes, was used as a global measure of performance. In calculation of tAcc, the sum of the entries on the main diagonal of the confusion matrix (correct decisions) is divided by the sum of the all entries (all decisions). Results A total of 1669 segments were extracted from the ECG dataset, 411 AS (n = 225 patients), 369 PEA (n = 205), 264 PR (n = 111), 397 VF (n = 162), and 228 VT (n = 29). The results for the different classifiers are shown in Table 1 for all the rhythm classes. In addition, tAcc is reported as an overall measure of the performance of each method, together with the average numbers of selected features.
A detailed analysis of the classification is shown in Table 2, which shows the confusion matrix for the algorithm with the highest tAcc (LDA + SBS). This confusion matrix provides all the information needed to calculate correct/incorrect classification rates for each rhythm type in detail. To make it clearer, beside each number in the confusion matrix there is another number in parenthesis which shows the classification/misclassification rate. To calculate those numbers the entries of each row of the confusion matrix is normalized. Thus, the percentage of correct and incorrect classification for each rhythm can be easily found. For example, the first row of this table shows 78.8% of AS rhythms are correctly classified as AS, but 12.9%, 0.7%, 7.1%, and 0.5% are incorrectly classified as PEA, PR, VF, and VT. As shown in Table 2 most classification errors correspond to decisions between PEA/PR, VF/AS and PEA/AS rhythms. Fig. 2 shows one example of the ECG for each entry of the confusion matrix of Table 2, where the rows show the labels assigned by the expert reviewers and the columns show the classification by the proposed algorithm. Discussion We have designed and demonstrated an automatic ECG based interpreter of resuscitation rhythms into five categories. These algorithms go beyond the binary shock/no-shock decision in defibrillators, and set the ground for the development of automatic methods for ECG interpretation with the level of detail required by an automated review system. Comparative evaluations of automatic interpretation of different cardiac rhythms The total accuracy of the algorithms introduced in this work is in the order of 70% which could be sufficient as a pre-screening tool in a semi-automatic rhythm interpreter. However, for quality assurance a better accuracy is required, and that is part of our ongoing research following the approaches/solutions proposed in this section. To use the current version of the algorithm, once it has labeled the data, an expert reviewer would have to go through the episodes in the final stage of the quality assurance of the annotations. But, his/her workload would be substantially reduced by the automatic annotations of our system, which would have to be only sporadically corrected. In the following, we discuss the performance of the algorithm in detail. The sensitivities reveal that the performance of all rhythm classes is in the same range (70% or higher) except for PEA and PR. The results underline the intrinsic difficulty of discriminating PEA from PR using only the ECG. Both PEA and PR are organized (ORG) cardiac rhythms, differentiated by the presence of a palpable pulse. Such a distinction is difficult even for experts,17 and involves the combined analysis of the ECG and impedance in automatic PEA/PR discrimination algorithms.18,19 The results from Table 2 show that when PEA and PR are grouped into a single ORG class the sensitivity for ORG rhythms rises to 75%. Future developments may then include the combined analysis of the ECG, impedance data, and/or ETCO2 to discriminate PEA from PR;18,19 or, in the context of a fully automated review system, to combine the analyses of the algorithms with clinical ROSC annotations made on site.8 Inspection of the data in Table 2, Figs. 2 and 3 reveals the main sources of misclassification. Errors are due to borderline rhythms such as: VF with low amplitude and dominant frequency (AS/VF discrimination), bradycardic rhythms (AS/PEA) and tachycardias of either ventricular or supraventricular origin (VT/PEA and VT/PR). Fig. 2 shows examples of these cases. There were also errors caused by transitional states, as shown by the example in Fig. 3(a), or by
A.B. Rad et al. / Resuscitation 102 (2016) 44–50
47
Table 2 The confusion matrix of the LDA + SBS classification algorithm; the numbers in parentheses show the classification/misclassification rate. LDA + SBS algorithm label
Expert reviewers’ label
AS PEA PR VF VT
AS
PEA
PR
VF
VT
324 (78.8%) 70 (19.0%) 14 (5.3%) 61 (15.4%) 4 (1.8%)
53 (12.9%) 206 (55.8%) 62 (23.5%) 11 (2.8%) 8 (3.5%)
3 (0.7%) 58 (15.7%) 147 (55.7%) 4 (1.0%) 3 (1.3%)
29 (7.1%) 24 (6.5%) 15 (5.7%) 295 (74.3%) 46 (20.2%)
2 (0.5%) 11 (3.0%) 26 (9.8%) 26 (6.5%) 167 (73.2%)
Fig. 2. Different examples of classified rhythms using the LDA + SBS algorithm. The rows show the labels assigned by the expert reviewers, and the columns show the classification by the proposed algorithm. The x and y axes of each plot are time (in s) and amplitudes (in mV).
rhythms analyzed immediately after stopping chest compressions, as shown in Fig. 3(b). A possible way to address these misclassifications would be to classify consecutive windows by using two layers of classifiers.20 The first layer classifies ECG in the signal level, as introduced in the present study, and generates the state level labels. The second level classifies a sequence of the state level labels to the final labels. Some mismatches are due to a few inconsistencies in the original annotations, both in the rhythm labels and in the chest compression intervals. For instance the rhythm labeled as VT and classified as AS in Fig. 2, also shown in Fig. 3(c), is an example of a mislabeled rhythm, and the rhythm labeled as AS and classified as VT in Fig. 2 corresponds to an asystole during chest compressions, as shown in Fig. 3(d). Although these instances are very few in our data, their presence confirms the importance of quality control in the review process of resuscitation data.
Finally, we emphasize that at this stage our goal was to classify rhythms for retrospective analysis purposes, and it is not intended to be used for guiding therapy. It is obvious that if we include further considerations like resuscitation action, we can optimize our algorithms for that specific application. For example, in this work the costs of misclassification are the same for all rhythms, however we can easily change the cost function for a specific rhythm in our algorithm. Does ECG carry any information which is useful for PEA/PR discrimination? It is a common belief that the presence of pulse cannot be identified in the ECG. Mathematically speaking, it means that in the absence of pulse information for PEA/PR discrimination we cannot have any better results than the random guess. However, we
48
A.B. Rad et al. / Resuscitation 102 (2016) 44–50
Fig. 3. Demonstration of some problematic segments in the classification process. The gray arrows indicate the labels assigned by experts (C-PR stands for PR rhythm during chest compression). The red dashed circles show the considered segments and their position in Fig. 2. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
showed that using signal processing and machine learning methods it is possible to classify PEA/PR from only ECG signal to some degree. Table 2 shows that 55.8% of PEA rhythms were correctly classified, but only 15.7% were incorrectly classified as PR. These two numbers should be almost equal if ECG does not carry any information useful for PEA/PR discrimination. The same argument can be used for PR rhythms where 55.7% of them were correctly classified, but 23.5% were incorrectly classified as PEA. This indicates that the ECG signal carries some information or patterns, which could be hidden from human eyes, and can be utilized for PEA/PR discrimination.
Comparative evaluations of the performance of the present and previous methods All the results with respect to the total accuracy (tAcc) have decreased in relation to our prior works. There could be different reasons for this. Firstly, in the new settings we kept track of patients to be sure that the data of each patient was used either only in the training or only in the testing datasets. In the previous works, the training and testing datasets were disjoint, but patients were not audited. Therefore, data from the same patient could appear in both training and testing sets, and thus there could be a data leakage of patients from the test data into the training data. Data leakage, which is among the top 10 machine learning and data mining mistakes,21 can result in unrealistically good performance of machine learning algorithms.22 Secondly, since in the previous studies there was no cross-validation step, the models could be biased on the data. Thirdly, in our previous studies the data was highly imbalanced. For imbalanced datasets tAcc is a biased estimate of performance.16 For example, while tAcc was 79% for Klocal hyperplane distance nearest-neighbor (HKNN) classifier and 76% for K-nearest neighbor (KNN) classifier, the Sen of VT rhythms were 10% and 0%, respectively.14 In this study we used an almost
balanced dataset, so while tAcc was 68% for LDA + SBS, Sen of VT improved to 71%. Comparative evaluations of the performance of the proposed classifiers The best result in the sense of overall accuracy belongs to LDA + SBS (Table 1). This may be surprising since the QDA classifier has greater flexibility than the LDA classifier. However, since the number of estimated parameters is larger in QDA than in LDA, in general QDA classifiers require larger training datasets.23 In addition, it seems that QDA is more sensitive to the violation of the basic assumptions like Gaussian distribution of the class conditional densities.24 Toward automated review International collaboration with treatment recommendations relies heavily on comparison between randomized controlled trials and evaluation of registry data.25,26 Still, unresolved issues remain regarding relative importance of the different aspects of treatment (chest compressions and pauses in chest compressions, ventilations, defibrillation attempts, medications) and their relationship in the treatment recommendations. These issues have been identified as knowledge gaps in resuscitation science by the International Liaison Committee on Resuscitation.27 Eftestøl et al.8 previously demonstrated how a large number of parameters registered during the review of resuscitation episodes could be automatically replicated from a minimal set of basic parameters describing the resuscitation episode. These basic parameters included start and end times of chest compressions, the electronic defibrillator event report, and annotations of the rhythm transition times. As automatic chest compression detection is possible,9 and the electronic event reports are commonly
A.B. Rad et al. / Resuscitation 102 (2016) 44–50
available, solving the problem of automatic rhythm interpretation will make it possible to do the review in a fully automatic fashion. The review process for quality assurance can be made more efficient as it will be possible to determine important indicators (e.g., time intervals) and quality parameters (e.g., the ratio of hands off intervals during treatable rhythms). This will be important for quality assurance and benchmarking within and between different sites as the same automatic review systems can be applied in different locations. It will also enable outcomes of clinical trials to be controlled for varying quality of basic treatment, and monitor and develop strategies for implementation of new treatment protocols and improvements in teaching/education.
49
Classification Features were fed into LDA and QDA classifiers, and in both cases the class conditional densities (CCD) were modeled with multivariate Gaussian distributions but with equal covariance matrices (homoscedasticity) in LDA and unequal covariance matrices (heteroscedasticity) in QDA. Mean and covariance matrices were estimated using maximum likelihood, and the pooled covariance matrix32 was computed for the LDA classifier. Bayes rule was applied for classification after modeling CCDs33 using uniform class priors. Model assessment and feature selection
Conclusion In this study, we have introduced a system for automatic rhythm interpretation of five cardiac rhythms (AS, PEA, PR, VF, VT) during resuscitation, exploring different feature extraction, feature selection, and classification methods in a systematic manner. Many previous studies of important treatment factors during resuscitation have relied on tedious manual interpretation of these five rhythms, and an automatic interpretation system would facilitate the handling of large resuscitation databases. With an overall accuracy of 68%, the current system has potential for improvement. Specific areas for improvement have been identified and are the focus of our ongoing work.
Conflict of interest statement The authors have no conflicts of interest.
Appendix A. Preprocessing and filtering ECG data was denoised and filtered in the wavelet domain where the ECG is represented as a summation of smooth and detailed parts in different scales.28,29 ECG signals were decomposed using an 8-level DWT with a daubechies-4 wavelet.30 Detail coefficients of levels 4–8 were used to reconstruct the denoised ECG, as shown in Fig. 1. In the 8-level DWT decomposition, the level 8 approximation coefficient is roughly related to frequencies below 1 Hz, responsible for low-frequency noise like baseline wander. Highfrequency noise (above 30 Hz) is suppressed by eliminating the detail coefficients in levels 1–3.
Feature extraction ECG features from both wavelet- and time-domain representations were used. In the wavelet domain, three statistical descriptors of the levels 4–8 detail coefficients were extracted: interquartile ranges (features 1–5), variances (6–10), and first quartiles (11–15). Morphological and time-domain features were extracted from the denoised ECG. Features 16–19 were the coefficients of the order 4 autoregressive (AR) model of the ECG, estimated using Burg’s method.31 The variance of the white noise input to the AR model was feature 20. The number of peaks of the autocorrelation above a threshold was feature 21. Then 8 statistical descriptors of the ECG, and its first and second derivatives were obtained (22–29).14 The last three features (30–32) were positive area (the sum of the positive signal values), negative area (the sum of the negative signal values), and amplitude range (the difference between the maximum and the minimum signal values).
A nested CV architecture34,35 was used, with a 10-fold patientwise external CV for an unbiased estimate of the performance (model assessment), and 5-fold internal CV for feature selection. The wrapper approach, in which the decision is based on the performance of the classifier, was adopted for feature selection.36,37 The total error (tErr = 1 − tAcc) was used to determine when to stop in the feature selection process. Two suboptimal feature selection approaches were explored, SBS38 where starting from the full set features are sequentially removed, and SFS39 where starting from an empty set features are sequentially added. After selecting the best feature subset in the internal CV loop, the resulting features were used to estimate the performance of the method. Moreover, when training the classifiers, the outliers of the training data were removed by three-sigma rule40 while the test data was untouched. References 1. Sunde K, Eftestøl T, Askenberg C, Steen PA. Quality assessment of defibrillation and advanced life support using data from the medical control module of the defibrillator. Resuscitation 1999;41:237–47. 2. Wik L, Kramer-Johansen J, Myklebust H, et al. Quality of cardiopulmonary resuscitation during out-of-hospital cardiac arrest. JAMA 2005;293:299–304. 3. Abella BS, Alvarado JP, Myklebust H, et al. Quality of cardiopulmonary resuscitation during in-hospital cardiac arrest. JAMA 2005;293:305–10. 4. Skogvoll E, Eftestøl T, Gundersen K, et al. Dynamics and state transitions during resuscitation in out-of-hospital cardiac arrest. Resuscitation 2008;78: 30–7. 5. Kvaløy JT, Skogvoll E, Eftestøl T, et al. Which factors influence spontaneous state transitions during resuscitation? Resuscitation 2009;80:863–9. 6. Nordseth T, Bergum D, Edelson DP, et al. Clinical state transitions during advanced life support (ALS) in in-hospital cardiac arrest. Resuscitation 2013;84:1238–44. 7. Qvigstad E, Kramer-Johansen J, Tømte Ø, et al. Clinical pilot study of different hand positions during manual chest compressions monitored with capnography. Resuscitation 2013;84:1203–7. 8. Eftestøl T, Sherman LD. Towards the automated analysis and database development of defibrillator data from cardiac arrest. BioMed Res Int 2014:276965:2014. 9. Ayala U, Eftestøl T, Alonso E, et al. Automatic detection of chest compressions for the assessment of CPR-quality parameters. Resuscitation 2014;85: 957–63. 10. Irusta U, Ruiz J, Ruiz de Gauna S, Eftestøl T, Kramer-Johansen J. A least mean-square filter for the estimation of the cardiopulmonary resuscitation artifact based on the frequency of the compressions. IEEE Trans Biomed Eng 2009;56:1052–62. 11. Irusta U, Ruiz J, Aramendi E, de Gauna SR, Ayala U, Alonso E. A high-temporal resolution algorithm to discriminate shockable from nonshockable rhythms in adults and children. Resuscitation 2012;83:1090–7. 12. Didon JP, Krasteva V, Ménétré S, Stoyanov T, Jekova I. Shock advisory system with minimal delay triggering after end of chest compressions: accuracy and gained hands-off time. Resuscitation 2011;82:S8–15. 13. Rad AB, Eftestøl T, Kvaløy JT, Ayala U, Kramer-Johansen J, Engan K. Probabilistic classification approaches for cardiac arrest rhythm interpretation during resuscitation. Comput Cardiol (CinC) 2013;40:125–8. 14. Rad AB, Eftestøl T, Kvaløy JT, Ayala U, Kramer-Johansen J, Engan K. Nearestmanifold classification approach for cardiac arrest rhythm interpretation during resuscitation. In: IEEE international conference on acoustics, speech and signal processing (ICASSP). 2014. p. 3621–5. 15. Kramer-Johansena J, Myklebust H, Wik L, et al. Quality of out-of-hospital cardiopulmonary resuscitation with real time automated feedback: a prospective interventional study. Resuscitation 2006;71:283–92.
50
A.B. Rad et al. / Resuscitation 102 (2016) 44–50
16. Mar T, Zaunseder S, Martínez JP, Llamedo M, Poll R. Optimization of ECG classification by means of feature selection. IEEE Trans Biomed Eng 2011;58:2168–77. 17. Eberle B, Dick WF, Schneider T, Wisser G, Doetsch S, Tzanova I. Checking the carotid pulse check: diagnostic accuracy of first responders in patients with and without a pulse. Resuscitation 1996;33:107–16. 18. Risdal M, Aase SO, Kramer-Johansen J, Eftestøl T. Automatic identification of return of spontaneous circulation during cardiopulmonary resuscitation. IEEE Trans Biomed Eng 2008;55:60–8. 19. Ruiz J, Ayala U, Ruiz de Gauna S, et al. Feasibility of automated rhythm assessment in chest compression pauses during cardiopulmonary resuscitation. Resuscitation 2013;84:1223–8. 20. Kwok H, Coult J, Drton M, Rea TD, Sherman L. Adaptive rhythm sequencing: a method for dynamic rhythm classification during CPR. Resuscitation 2015;91:26–31. 21. Nisbet R, Elder J, Miner G. Handbook of statistical analysis and data mining applications. Academic Press; 2009. 22. Kaufman S, Rosset S, Perlich C, Stitelman O. Leakage in data mining: formulation, detection, and avoidance. ACM Trans Knowl Discov Data 2012;6:15. 23. Wahl PW, Kronmal RA. Discriminant functions when covariances are unequal and sample sizes are moderate. Biometrics 1977;33:479–84. 24. Friedman JH. Regularized discriminant analysis. JASA 1989;84:165–75. 25. Nolan J, Baskett P, editors. European Resuscitation Council guidelines for resuscitation 2005. Resuscitation. 2005. p. 67. 26. Proceedings of the 2005 international consensus on cardiopulmonary resuscitation and emergency cardiovascular care science with treatment recommendations. Resuscitation 2005;67:157–341. 27. Gazmuri RJ, Nolan JP, Nadkarni VM, et al. Scientific knowledge gaps and clinical research priorities for cardiopulmonary resuscitation and emergency cardiovascular care identified during the 2005 International Consensus Conference
28. 29. 30. 31.
32. 33. 34. 35.
36. 37.
38. 39. 40.
on ECC and CPR Science with Treatment Recommendations. Resuscitation 2007;75:400–11. Mallat SG. Multiresolution approximations and wavelet orthonormal bases of L2 (R). Trans Am Math Soc 1989;315:69–87. Meyer Y, Salinger DH. Wavelets and operators. Cambridge: Cambridge University Press; 1995. Daubechies I. Ten lectures on wavelets. Philadelphia: SIAM; 1992. Burg JP. A new analysis technique for time series data. NATO Advanced Study Institute of Signal Processing with emphasis on Underwater Acoustics; 1968. p. 1. Legendre P, Legendre L. Numerical ecology. 2nd English ed. Amsterdam: Elsevier; 1998. Theodoridis S, Koutroumbas K. Pattern recognition. 4th ed. Academic Press; 2008. Varma S, Simon R. Bias in error estimation when using cross-validation for model selection. BMC Bioinf 2006;7:91. Krstajic D, Buturovic LJ, Leahy DE, Thomas S. Cross-validation pitfalls when selecting and assessing regression and classification models. J Cheminf 2014;6:10. Kohavi R, John GH. Wrappers for feature subset selection. Artif Intell 1997;97:273–324. John GH, Kohavi R, Pfleger K. Irrelevant features and the subset selection problem. In: Machine learning: proceedings of the eleventh international conference. 1994. p. 121–9. Marill T, Green DM. On the effectiveness of receptors in recognition systems. IEEE Trans Inf Theory 1963;9:11–7. Whitney AW. A direct method of nonparametric measurement selection. IEEE Trans Comput 1971;20:1100–3. Maronna RA, Martin RD, Yohai VJ. Robust statistics. John Wiley & Sons; 2006.