Resuscitation 91 (2015) 26–31
Contents lists available at ScienceDirect
Resuscitation journal homepage: www.elsevier.com/locate/resuscitation
Clinical Paper
Adaptive rhythm sequencing: A method for dynamic rhythm classification during CPR夽 Heemun Kwok a,b,∗ , Jason Coult a,c , Mathias Drton d , Thomas D. Rea a,b,e , Lawrence Sherman a,b,c a
Center for Progress in Resuscitation, University of Washington, Seattle, WA, United States Department of Medicine, University of Washington School of Medicine, Seattle, WA, United States c Department of Bioengineering, University of Washington, Seattle, WA, United States d Department of Statistics, University of Washington, Seattle, WA, United States e King County Emergency Medical Services Division, Seattle King County Department of Public Health, Seattle, WA, United States b
a r t i c l e
i n f o
Article history: Received 27 November 2014 Received in revised form 31 January 2015 Accepted 18 February 2015 Keywords: Cardiac arrest Resuscitation Cardiac rhythm Hidden Markov model
a b s t r a c t Objective: The accuracy of methods that classify the cardiac rhythm despite CPR artifact could potentially be improved by utilizing continuous ECG data. Our objective is to compare three approaches which use identical ECG features and differ only in their degree of temporal integration: (1) static classification, which analyzes 4-s ECG frames in isolation; (2) “best-of-three averaging,” which takes the average of three consecutive static classifications successively; and (3) “adaptive rhythm sequencing,” which uses hidden Markov models to model ECG segments as rhythm sequences. Methods: Defibrillator recordings from 95 out-of-hospital cardiac arrests were divided into training and test sets. Each method classified the rhythm as asystole, organized rhythm or shockable rhythm throughout the recordings. Classifications were compared to the gold standard of physician review. The primary outcome was accuracy during CPR, which was estimated using a generalized linear mixed-effects model. Results: In the training set, accuracies during CPR were 0.89 (95% CI 0.85, 0.92), 0.92 (95% CI 0.89, 0.94) and 0.97 (95% CI 0.95, 0.98) for the static, best-of-three averaging and adaptive rhythm sequencing methods, respectively. The corresponding results in the test set were 0.92 (95% CI 0.86, 0.96), 0.94 (95% CI 0.89, 0.97), and 0.97 (95% CI 0.94, 0.99). Of the dynamic methods, only adaptive rhythm sequencing was significantly more accurate than static classification in the training (p < 0.001) and test (p = 0.03) sets. Conclusion: In a continuous monitoring setting, adaptive rhythm sequencing was significantly more accurate than static rhythm classification during CPR. © 2015 Elsevier Ireland Ltd. All rights reserved.
1. Introduction Out-of-hospital cardiac arrest (OHCA) due to ventricular fibrillation affects hundreds of thousands of individuals worldwide each year. High-quality CPR and early defibrillation are two interventions which can improve survival,1,2 but it is often necessary to interrupt CPR repeatedly in order to reliably assess the cardiac rhythm. These pauses decrease coronary perfusion pressure, reduce the benefit of CPR, and are associated with decreased survival.3,4 An automated cardiac rhythm classification method
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.2015.02.031. ∗ Corresponding author at: Box 359702, 325 Ninth AV, Seattle, WA 98104, United States. E-mail address: [email protected] (H. Kwok). http://dx.doi.org/10.1016/j.resuscitation.2015.02.031 0300-9572/© 2015 Elsevier Ireland Ltd. All rights reserved.
which accurately identifies the rhythm in the presence of artifact could potentially improve survival. Although a variety of methods have been developed to “read-through” CPR artifact, no published approach is externally validated to achieve American Heart Association accuracy goals.5–8 Moreover, an inaccurate method could have an adverse effect on survival. False positive classification (i.e., rhythm classified as shockable when the true rhythm is non-shockable) could lead to an unnecessary pause in chest compressions or an unnecessary shock, and false negative classification (i.e., rhythm classified as non-shockable when the true rhythm is shockable) could delay defibrillation. In one simulation study, a method with a sensitivity of 94% and a specificity of 81% would produce unnecessary CPR interruptions in 42% of non-shockable rhythm episodes.9 Unlike classification methods that require a pause in CPR for rhythm analysis and therefore analyze short ECG segments (typically 4 or 5 s), methods that read through CPR could be employed
H. Kwok et al. / Resuscitation 91 (2015) 26–31
continuously over longer time periods. Additionally, these methods would optimally provide specific rhythm determination—not merely whether the rhythm is shockable—to guide resuscitation. Repeated measurements have the potential to increase accuracy over a single, isolated measurement, because the rhythm at any point is directly related to the preceding rhythm. However, simply averaging repeated measurements over an entire ECG segment disregards the possibility that the rhythm can change at any second and would delay recognition of rhythm transitions. Therefore, it is unclear how to utilize accumulating ECG data in a continuous monitoring environment to increase accuracy over a static approach which analyzes short ECG segments in isolation. In this study, we evaluate two methods which extend classification algorithms during CPR to a dynamic setting where the past rhythm influences the current rhythm. The “best-of-three averaging” method applies a static method to consecutive, nonoverlapping frames, where the frame is defined as the unit of ECG analyzed by a static method. If two consecutive classifications agree, that classification is assigned to both frames; if not, a third classification resolves the discrepancy. The second method is termed “adaptive rhythm sequencing.” A sequence may be defined as a set of related events that follow each other in a particular order. Adaptive rhythm sequencing models a continuous ECG segment as a rhythm sequence using hidden Markov models. Hidden Markov models are a statistical tool used to analyze sequential processes in a variety of fields, such as automated speech recognition and DNA sequencing.10–12 A hidden Markov model integrates ECG data from all frames within a segment and determines the probability of any possible rhythm sequence. Dynamic classification methods may be used for real-time rhythm analysis in a clinical setting or post hoc analysis in research and quality improvement settings. The objective of this study is to compare the accuracy of rhythm classification among static, best-of-three averaging, and adaptive rhythm sequencing methods when applied to post hoc rhythm analysis. The three approaches employ identical ECG features to assess the underlying ECG through CPR artifact and differ only in their degree of temporal integration. We hypothesized that the adaptive rhythm sequencing method would have greater accuracy than the static and best-of-three averaging methods.
27
Each ECG recording was divided into separate ECG segments by defibrillation attempts. ECG segments were then subdivided into consecutive, non-overlapping, 4-s frames prior to classification by each method. 2.2. ECG features Each classification method may be implemented in conjunction with any quantitative ECG features. Because our primary interest is relative performance and not evaluation of any particular features, all methods utilized an identical set of three ECG features under development in our laboratory, and these features were calculated for each frame. These features are based upon the amplitude and cross-correlation with a series of wavelets designed to detect specific ECG morphologies characteristic of the three rhythm classes.13 2.3. Static classification We derived a quadratic discriminant analysis model to classify each frame as asystole, an organized rhythm, or a shockable rhythm, using the training data.14 This model assumes that each rhythm has a characteristic distribution of ECG features but ignores temporal correlation between frames. The same model was employed regardless of whether or not CPR was in process.
A Signal framet-1
Signal framet
Signal framet+ 1
EC G featurest-1
EC G featurest
EC G featurest+ 1
Rhythm statet-1
Rhythm statet
Rhythm statet+ 1
2. Methods 2.1. Study design, setting, and defibrillator database This study was approved by the Institutional Review Board for Human Subjects Research at the University of Washington and the Department of Public Health—Seattle and King County (#47354). The study cohort was a convenience sample of 95 OHCA cases who received at least one shock by a Philips MRX defibrillator as part of attempted resuscitation from a metropolitan, two-tier EMS system in 2006–2007. Cases were randomly divided into a training set (N = 47) and test set (N = 48). Two physicians reviewed the ECG, impedance, and, when available, accelerometry signals of the complete defibrillator recordings during attempted resuscitation. The cardiac rhythm was classified in a continuous fashion, including periods with and without CPR, as asystole, an organized rhythm, or a shockable rhythm (ventricular fibrillation or ventricular tachycardia), according to standard rhythm definitions.7 Rhythm classifications during CPR were confirmed by examining the ECG within adjacent sections without CPR. A third reviewer was consulted for cases of disagreement, and the rhythm was classified as indeterminate if consensus was not achieved. The presence of chest compressions was determined from the impedance and accelerometry signals.
B
0.02 0.97
Asystole
Organized
0.99
0.003 0.01
0.002
0.001
0.01
Shockable
0.998 Fig. 1. (A) Directed graph representing the hidden Markov model. The arrows reflect probabilistic, causal relationships.23 While the ECG features depend upon the rhythm states, Bayes’ rule may be used to make inferences regarding the hidden rhythm states from the ECG features. (B) State diagram of a hidden Markov model showing possible rhythm states and transitions between states. Transitions between states are governed by transition probabilities, which are shown adjacent to the arrows.
28
H. Kwok et al. / Resuscitation 91 (2015) 26–31
A
B
C
Fig. 2. (A) ECG segment of ventricular fibrillation during CPR. This 24 s segment is divided into six, 4-s frames. The rhythm is obscured by chest compression artifact. (B) Static classifications were obtained by analyzing each frame in isolation. Slice areas represent the probability of each rhythm based on a quadratic discriminant analysis model. The static method misclassifies the third and fourth frames as an organized rhythm. (C) Adaptive rhythm sequencing uses a hidden Markov model to calculate the posterior probability of any sequence given the ECG data. The five sequences shown have the highest posterior probabilities of the 729 possible sequences (see Table 3 for further details).
2.4. Best-of-three averaging classification The best-of-three averaging method builds upon the static classifications obtained by the quadratic discriminant analysis model. If two consecutive classifications agreed, that classification was assigned to those frames. If the first two classifications disagreed, a third classification determined the final classification applied to all three frames. This process was repeated until the end of the continuous ECG segment. 2.5. Adaptive rhythm sequencing classification With a hidden Markov model, the observed ECG is modeled using a sequence of “hidden” rhythm states, one state for each frame. The unobserved, hidden states are reflected by their ECG features, since each rhythm has a characteristic emission distribution, or pattern, of ECG features (Fig. 1A). The model accounts for temporal dependence among states by assuming that the sequence forms a Markov chain. In a Markov chain, future states in a sequence depend only upon the present state and not past states, and the process of transitioning from one state to the next is governed by a collection of transition probabilities (Fig. 1B).10,15 The final components of the hidden Markov model are the initial probabilities for the first state within an ECG segment. All model parameters (emission distributions, transition probabilities and initial probabilities) were estimated from the physician-classified training data. For any sequence of states, its prior probability quantifies how likely it is from a temporal perspective without (“prior to”)
considering the observed ECG features; prior probability is calculated from the initial and transition probabilities. The emission likelihood reflects agreement between the states and the ECG features and is calculated from the emission distributions. The posterior probability is the sequence probability that incorporates both sources of information: the temporal dependence between successive states and the ECG features. It is calculated from the prior probability and the emission likelihood using Bayes’ rule. The adaptive rhythm sequencing classification is defined as the sequence with the maximum posterior probability and is identified by a dynamic programming algorithm known as the Viterbi algorithm.10 To illustrate with a concrete example, consider the 24-s segment of ventricular fibrillation during CPR in Fig. 2A. This segment consists of six, 4-s frames. With three possible rhythm states for each frame, there are 36 or 729 possible rhythm sequences. The true sequence, {shockable, shockable, shockable, shockable, shockable, shockable}, is one possible sequence. Its prior probability is 0.33, its emission likelihood is 9.7 × 10−6 , and using Bayes rule, its posterior probability is 0.98. A technical description of this method is provided in the Supplementary Material. 2.6. Data analysis Each ECG segment was classified according to the three methods, and the classifications were then compared to the gold standard of physician review. The primary outcome was accuracy,
H. Kwok et al. / Resuscitation 91 (2015) 26–31
29
Table 1 Characteristics of the training and test sets. Variable Clinical characteristics Age (y), median (IQR) Male Presumed cardiac etiology Location Home Public Witnessed Bystander CPR EMS response time (min), median (IQR) Initial rhythm Shockable PEA Asystole Return of spontaneous circulation Survival to hospital admission Survival to hospital discharge Defibrillator recording characteristics Duration (min), median (IQR) Distribution of rhythms Asystole Organized rhythm Ventricular fibrillation Ventricular tachycardia Indeterminate rhythm CPR fraction, median (IQR) Total number of sequences Number of defibrillation attempts, median (IQR) Number of spontaneous rhythm transitions, median (IQR)
Training set, N = 47
Test set, N = 48
Combined, N = 95
62 (52, 73) 79% 87%
62 (52, 70) 73% 85%
62 (52, 72) 76% 85%
68% 32% 83% 60% 5 (4, 6)
73% 27% 69% 60% 5 (4, 6)
70% 29% 76% 60% 5 (4, 6)
87% 11% 2% 57% 55% 36%
85% 10% 4% 68% 65% 38%
86% 11% 3% 63% 60% 37%
10.5 (4.9, 17)
9.5 (4.9, 23)
9.9 (4.9, 20.2)
15% 46% 36% 3% 3% 0.70 (0.63, 0.77) 183 2 (2, 3) 2 (1, 4)
15% 43% 36% 6% 4% 0.69 (0.60, 0.77) 193 2 (1,4) 2 (1,4)
15% 45% 36% 4% 3% 0.69 (0.61, 0.77) 376 2 (1, 3) 2 (1, 4)
Abbreviations: EMS, emergency medical services; IQR, interquartile range.
Table 2 Accuracy of static, best-of-three-averaging and adaptive rhythm sequencing classification methods during CPR. Method
Static Best-of-three averaging Adaptive rhythm sequencing
Training set
Test set
Accuracy
p-Value
Accuracy
p-Value
0.89 (0.85, 0.92) 0.92 (0.89, 0.94) 0.97 (0.95, 0.98)
Referent 0.18
Contents lists available at ScienceDirect
Resuscitation journal homepage: www.elsevier.com/locate/resuscitation
Clinical Paper
Adaptive rhythm sequencing: A method for dynamic rhythm classification during CPR夽 Heemun Kwok a,b,∗ , Jason Coult a,c , Mathias Drton d , Thomas D. Rea a,b,e , Lawrence Sherman a,b,c a
Center for Progress in Resuscitation, University of Washington, Seattle, WA, United States Department of Medicine, University of Washington School of Medicine, Seattle, WA, United States c Department of Bioengineering, University of Washington, Seattle, WA, United States d Department of Statistics, University of Washington, Seattle, WA, United States e King County Emergency Medical Services Division, Seattle King County Department of Public Health, Seattle, WA, United States b
a r t i c l e
i n f o
Article history: Received 27 November 2014 Received in revised form 31 January 2015 Accepted 18 February 2015 Keywords: Cardiac arrest Resuscitation Cardiac rhythm Hidden Markov model
a b s t r a c t Objective: The accuracy of methods that classify the cardiac rhythm despite CPR artifact could potentially be improved by utilizing continuous ECG data. Our objective is to compare three approaches which use identical ECG features and differ only in their degree of temporal integration: (1) static classification, which analyzes 4-s ECG frames in isolation; (2) “best-of-three averaging,” which takes the average of three consecutive static classifications successively; and (3) “adaptive rhythm sequencing,” which uses hidden Markov models to model ECG segments as rhythm sequences. Methods: Defibrillator recordings from 95 out-of-hospital cardiac arrests were divided into training and test sets. Each method classified the rhythm as asystole, organized rhythm or shockable rhythm throughout the recordings. Classifications were compared to the gold standard of physician review. The primary outcome was accuracy during CPR, which was estimated using a generalized linear mixed-effects model. Results: In the training set, accuracies during CPR were 0.89 (95% CI 0.85, 0.92), 0.92 (95% CI 0.89, 0.94) and 0.97 (95% CI 0.95, 0.98) for the static, best-of-three averaging and adaptive rhythm sequencing methods, respectively. The corresponding results in the test set were 0.92 (95% CI 0.86, 0.96), 0.94 (95% CI 0.89, 0.97), and 0.97 (95% CI 0.94, 0.99). Of the dynamic methods, only adaptive rhythm sequencing was significantly more accurate than static classification in the training (p < 0.001) and test (p = 0.03) sets. Conclusion: In a continuous monitoring setting, adaptive rhythm sequencing was significantly more accurate than static rhythm classification during CPR. © 2015 Elsevier Ireland Ltd. All rights reserved.
1. Introduction Out-of-hospital cardiac arrest (OHCA) due to ventricular fibrillation affects hundreds of thousands of individuals worldwide each year. High-quality CPR and early defibrillation are two interventions which can improve survival,1,2 but it is often necessary to interrupt CPR repeatedly in order to reliably assess the cardiac rhythm. These pauses decrease coronary perfusion pressure, reduce the benefit of CPR, and are associated with decreased survival.3,4 An automated cardiac rhythm classification method
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.2015.02.031. ∗ Corresponding author at: Box 359702, 325 Ninth AV, Seattle, WA 98104, United States. E-mail address: [email protected] (H. Kwok). http://dx.doi.org/10.1016/j.resuscitation.2015.02.031 0300-9572/© 2015 Elsevier Ireland Ltd. All rights reserved.
which accurately identifies the rhythm in the presence of artifact could potentially improve survival. Although a variety of methods have been developed to “read-through” CPR artifact, no published approach is externally validated to achieve American Heart Association accuracy goals.5–8 Moreover, an inaccurate method could have an adverse effect on survival. False positive classification (i.e., rhythm classified as shockable when the true rhythm is non-shockable) could lead to an unnecessary pause in chest compressions or an unnecessary shock, and false negative classification (i.e., rhythm classified as non-shockable when the true rhythm is shockable) could delay defibrillation. In one simulation study, a method with a sensitivity of 94% and a specificity of 81% would produce unnecessary CPR interruptions in 42% of non-shockable rhythm episodes.9 Unlike classification methods that require a pause in CPR for rhythm analysis and therefore analyze short ECG segments (typically 4 or 5 s), methods that read through CPR could be employed
H. Kwok et al. / Resuscitation 91 (2015) 26–31
continuously over longer time periods. Additionally, these methods would optimally provide specific rhythm determination—not merely whether the rhythm is shockable—to guide resuscitation. Repeated measurements have the potential to increase accuracy over a single, isolated measurement, because the rhythm at any point is directly related to the preceding rhythm. However, simply averaging repeated measurements over an entire ECG segment disregards the possibility that the rhythm can change at any second and would delay recognition of rhythm transitions. Therefore, it is unclear how to utilize accumulating ECG data in a continuous monitoring environment to increase accuracy over a static approach which analyzes short ECG segments in isolation. In this study, we evaluate two methods which extend classification algorithms during CPR to a dynamic setting where the past rhythm influences the current rhythm. The “best-of-three averaging” method applies a static method to consecutive, nonoverlapping frames, where the frame is defined as the unit of ECG analyzed by a static method. If two consecutive classifications agree, that classification is assigned to both frames; if not, a third classification resolves the discrepancy. The second method is termed “adaptive rhythm sequencing.” A sequence may be defined as a set of related events that follow each other in a particular order. Adaptive rhythm sequencing models a continuous ECG segment as a rhythm sequence using hidden Markov models. Hidden Markov models are a statistical tool used to analyze sequential processes in a variety of fields, such as automated speech recognition and DNA sequencing.10–12 A hidden Markov model integrates ECG data from all frames within a segment and determines the probability of any possible rhythm sequence. Dynamic classification methods may be used for real-time rhythm analysis in a clinical setting or post hoc analysis in research and quality improvement settings. The objective of this study is to compare the accuracy of rhythm classification among static, best-of-three averaging, and adaptive rhythm sequencing methods when applied to post hoc rhythm analysis. The three approaches employ identical ECG features to assess the underlying ECG through CPR artifact and differ only in their degree of temporal integration. We hypothesized that the adaptive rhythm sequencing method would have greater accuracy than the static and best-of-three averaging methods.
27
Each ECG recording was divided into separate ECG segments by defibrillation attempts. ECG segments were then subdivided into consecutive, non-overlapping, 4-s frames prior to classification by each method. 2.2. ECG features Each classification method may be implemented in conjunction with any quantitative ECG features. Because our primary interest is relative performance and not evaluation of any particular features, all methods utilized an identical set of three ECG features under development in our laboratory, and these features were calculated for each frame. These features are based upon the amplitude and cross-correlation with a series of wavelets designed to detect specific ECG morphologies characteristic of the three rhythm classes.13 2.3. Static classification We derived a quadratic discriminant analysis model to classify each frame as asystole, an organized rhythm, or a shockable rhythm, using the training data.14 This model assumes that each rhythm has a characteristic distribution of ECG features but ignores temporal correlation between frames. The same model was employed regardless of whether or not CPR was in process.
A Signal framet-1
Signal framet
Signal framet+ 1
EC G featurest-1
EC G featurest
EC G featurest+ 1
Rhythm statet-1
Rhythm statet
Rhythm statet+ 1
2. Methods 2.1. Study design, setting, and defibrillator database This study was approved by the Institutional Review Board for Human Subjects Research at the University of Washington and the Department of Public Health—Seattle and King County (#47354). The study cohort was a convenience sample of 95 OHCA cases who received at least one shock by a Philips MRX defibrillator as part of attempted resuscitation from a metropolitan, two-tier EMS system in 2006–2007. Cases were randomly divided into a training set (N = 47) and test set (N = 48). Two physicians reviewed the ECG, impedance, and, when available, accelerometry signals of the complete defibrillator recordings during attempted resuscitation. The cardiac rhythm was classified in a continuous fashion, including periods with and without CPR, as asystole, an organized rhythm, or a shockable rhythm (ventricular fibrillation or ventricular tachycardia), according to standard rhythm definitions.7 Rhythm classifications during CPR were confirmed by examining the ECG within adjacent sections without CPR. A third reviewer was consulted for cases of disagreement, and the rhythm was classified as indeterminate if consensus was not achieved. The presence of chest compressions was determined from the impedance and accelerometry signals.
B
0.02 0.97
Asystole
Organized
0.99
0.003 0.01
0.002
0.001
0.01
Shockable
0.998 Fig. 1. (A) Directed graph representing the hidden Markov model. The arrows reflect probabilistic, causal relationships.23 While the ECG features depend upon the rhythm states, Bayes’ rule may be used to make inferences regarding the hidden rhythm states from the ECG features. (B) State diagram of a hidden Markov model showing possible rhythm states and transitions between states. Transitions between states are governed by transition probabilities, which are shown adjacent to the arrows.
28
H. Kwok et al. / Resuscitation 91 (2015) 26–31
A
B
C
Fig. 2. (A) ECG segment of ventricular fibrillation during CPR. This 24 s segment is divided into six, 4-s frames. The rhythm is obscured by chest compression artifact. (B) Static classifications were obtained by analyzing each frame in isolation. Slice areas represent the probability of each rhythm based on a quadratic discriminant analysis model. The static method misclassifies the third and fourth frames as an organized rhythm. (C) Adaptive rhythm sequencing uses a hidden Markov model to calculate the posterior probability of any sequence given the ECG data. The five sequences shown have the highest posterior probabilities of the 729 possible sequences (see Table 3 for further details).
2.4. Best-of-three averaging classification The best-of-three averaging method builds upon the static classifications obtained by the quadratic discriminant analysis model. If two consecutive classifications agreed, that classification was assigned to those frames. If the first two classifications disagreed, a third classification determined the final classification applied to all three frames. This process was repeated until the end of the continuous ECG segment. 2.5. Adaptive rhythm sequencing classification With a hidden Markov model, the observed ECG is modeled using a sequence of “hidden” rhythm states, one state for each frame. The unobserved, hidden states are reflected by their ECG features, since each rhythm has a characteristic emission distribution, or pattern, of ECG features (Fig. 1A). The model accounts for temporal dependence among states by assuming that the sequence forms a Markov chain. In a Markov chain, future states in a sequence depend only upon the present state and not past states, and the process of transitioning from one state to the next is governed by a collection of transition probabilities (Fig. 1B).10,15 The final components of the hidden Markov model are the initial probabilities for the first state within an ECG segment. All model parameters (emission distributions, transition probabilities and initial probabilities) were estimated from the physician-classified training data. For any sequence of states, its prior probability quantifies how likely it is from a temporal perspective without (“prior to”)
considering the observed ECG features; prior probability is calculated from the initial and transition probabilities. The emission likelihood reflects agreement between the states and the ECG features and is calculated from the emission distributions. The posterior probability is the sequence probability that incorporates both sources of information: the temporal dependence between successive states and the ECG features. It is calculated from the prior probability and the emission likelihood using Bayes’ rule. The adaptive rhythm sequencing classification is defined as the sequence with the maximum posterior probability and is identified by a dynamic programming algorithm known as the Viterbi algorithm.10 To illustrate with a concrete example, consider the 24-s segment of ventricular fibrillation during CPR in Fig. 2A. This segment consists of six, 4-s frames. With three possible rhythm states for each frame, there are 36 or 729 possible rhythm sequences. The true sequence, {shockable, shockable, shockable, shockable, shockable, shockable}, is one possible sequence. Its prior probability is 0.33, its emission likelihood is 9.7 × 10−6 , and using Bayes rule, its posterior probability is 0.98. A technical description of this method is provided in the Supplementary Material. 2.6. Data analysis Each ECG segment was classified according to the three methods, and the classifications were then compared to the gold standard of physician review. The primary outcome was accuracy,
H. Kwok et al. / Resuscitation 91 (2015) 26–31
29
Table 1 Characteristics of the training and test sets. Variable Clinical characteristics Age (y), median (IQR) Male Presumed cardiac etiology Location Home Public Witnessed Bystander CPR EMS response time (min), median (IQR) Initial rhythm Shockable PEA Asystole Return of spontaneous circulation Survival to hospital admission Survival to hospital discharge Defibrillator recording characteristics Duration (min), median (IQR) Distribution of rhythms Asystole Organized rhythm Ventricular fibrillation Ventricular tachycardia Indeterminate rhythm CPR fraction, median (IQR) Total number of sequences Number of defibrillation attempts, median (IQR) Number of spontaneous rhythm transitions, median (IQR)
Training set, N = 47
Test set, N = 48
Combined, N = 95
62 (52, 73) 79% 87%
62 (52, 70) 73% 85%
62 (52, 72) 76% 85%
68% 32% 83% 60% 5 (4, 6)
73% 27% 69% 60% 5 (4, 6)
70% 29% 76% 60% 5 (4, 6)
87% 11% 2% 57% 55% 36%
85% 10% 4% 68% 65% 38%
86% 11% 3% 63% 60% 37%
10.5 (4.9, 17)
9.5 (4.9, 23)
9.9 (4.9, 20.2)
15% 46% 36% 3% 3% 0.70 (0.63, 0.77) 183 2 (2, 3) 2 (1, 4)
15% 43% 36% 6% 4% 0.69 (0.60, 0.77) 193 2 (1,4) 2 (1,4)
15% 45% 36% 4% 3% 0.69 (0.61, 0.77) 376 2 (1, 3) 2 (1, 4)
Abbreviations: EMS, emergency medical services; IQR, interquartile range.
Table 2 Accuracy of static, best-of-three-averaging and adaptive rhythm sequencing classification methods during CPR. Method
Static Best-of-three averaging Adaptive rhythm sequencing
Training set
Test set
Accuracy
p-Value
Accuracy
p-Value
0.89 (0.85, 0.92) 0.92 (0.89, 0.94) 0.97 (0.95, 0.98)
Referent 0.18
