Interruptions in cardiopulmonary resuscitation (CPR) compromise defibrillation success. However, CPR must be interrupted to analyze the rhythm because although current methods for rhythm analysis during CPR have high sensitivity for shockable rhythms, the specificity for nonshockable rhythms is still too low. This paper introduces a new approach to rhythm analysis during CPR that combines two strategies: a state-of-the-art CPR artifact suppression filter and a shock advice algorithm (SAA) designed to optimally classify the filtered signal. Emphasis is on designing an algorithm with high specificity. The SAA includes a detector for low electrical activity rhythms to increase the specificity, and a shock/no-shock decision algorithm based on a support vector machine classifier using slope and frequency features. For this study, 1185 shockable and 6482 nonshockable 9-s segments corrupted by CPR artifacts were obtained from 247 patients suffering out-of-hospital cardiac arrest. The segments were split into a training and a test set. For the test set, the sensitivity and specificity for rhythm analysis during CPR were 91.0% and 96.6%, respectively. This new approach shows an important increase in specificity without compromising the sensitivity when compared to previous studies.
Out-of-hospital cardiac arrest (OHCA) is a leading cause of mortality in the industrialized world, with an estimated annual incidence between 28 and 55 cases per 100,000 person-years [
Although different approaches to rhythm analysis during CPR have been explored, for instance, algorithms that directly diagnose the ECG corrupt with CPR artifacts [
Currently rhythm analysis during CPR is not possible [
In this study we explore the possibility of combining adaptive filtering techniques with a SAA designed to optimally classify the rhythm after filtering. The aim is to improve the accuracy of current approaches and in particular to overcome the low specificity. When compared to previous studies, our results showed an increased specificity without compromising the sensitivity, for a comprehensive dataset of OHCA rhythms.
The data for this study were extracted from a large prospective study of OHCA conducted between 2002 and 2004 in three European sites [
For this study specific records containing the ECG and CD signals were automatically extracted from the original episodes. First rhythm transitions were identified using the original annotations, and then for each interval without rhythm transitions at most one record was extracted to avoid bias due to data selection. Records were extracted if the following criteria were met: duration of more than 20-s, ongoing CCs, and the same rhythm annotation before and after CCs. Following the AHA statement the records were grouped into a shockable and a nonshockable category. The amplitude thresholds adopted for coarse VF and ASY are those accepted in the literature on SAAs [
organized rhythms (ORG): all nonshockable rhythms except ASY (PEA and PR), asystole (ASY): rhythms with peak-to-peak amplitudes below 100
All signals were resampled to
Following standard practice in SAA design, the rhythm analysis method was designed to analyze three consecutive 3 s windows, so it gives a diagnosis every 9 s [
The block diagram of the rhythm analysis method is shown in Figure
Block diagram of the new approach to rhythm analysis during CPR in which an adaptive filter (LMS filter based on the CD signal) is used in combination with a SAA designed to optimally classify the filtered signal,
CPR artifacts were suppressed using a state-of-the-art method based on an LMS filter [
Filtering example of a 12-s segment. In the first 3 s there is no artifact and the underlying VF is visible. The filter estimates the artifact,
The SAA consists of a LEA detector followed by the Sh/NSh algorithm. The LEA detector identifies LEA windows as nonshockable; the rest of the windows are further processed by the Sh/NSh algorithm for a definitive diagnosis.
Some nonshockable rhythms (ASY, bradyarrhythmias or idioventricular rhythms) may not present QRS complexes in a 3 s analysis window. In these cases, filtering the CC artifact results in where
Examples of a LEA rhythm (a) and a VF (b) window after filtering the CC artifact,
ASY
VF
LEA rhythms have smaller values of
During resuscitation, ORG rhythms with electrical activity may be very different in terms of rate, QRS width, or QRS morphology. Furthermore, even after CPR artifact suppression, rhythms may present important filtering residuals that may resemble VF. Four features derived from the frequency domain and slope analyses were defined. For rhythms with electrical activity, these features emphasize the differences between nonshockable (with QRS complexes) and shockable (without QRS complexes) rhythms.
QRS complexes were enhanced in
Shockable rhythms will present larger values of
Example of the slope analysis for VF (a) and an ORG (b) window. During VF the slope,
For the frequency analysis, a Hamming window was applied to
Shockable rhythms have larger values of
Example of the frequency domain analysis for VF (a) and an ORG (b) window. VF concentrates most of its power around the fibrillation band (blue). ORG rhythms have a spectrum with many harmonics of the heart rate and thus larger
VF
ORG
The Sh/NSh algorithm classified windows using a SVM with a Gaussian kernel [
The rate and depth characteristics of CPR in our data were analyzed for each 9 s segment. The distributions for rate and depth did not pass the Kolmogorov-Smirnov test for normality and are reported as median and 5–95 percentiles.
For each discrimination feature of the SAA, statistical differences in medians between the targeted classification groups of each subalgorithm were measured using the Mann-Whitney where the true positive rate (TPR) and the true negative rate (TNR) are the capacity of the SVM classifier to detect shockable and ORG windows, respectively. Weights were assigned to each class to resolve the unbalance in the number of instances per class [
The performance of the algorithm was measured in the test set in terms of sensitivity and specificity. Since both 3 s windows and 9 s segments correspond to consecutive analyses within a record, the sensitivities, specificities, and their 90% low one-sided confidence intervals (CI) were adjusted for clustering (longitudinal data) within each record, using a longitudinal logistic model fit by generalized estimating equations (GEE) [
Our data comprise 7667 9 s segments within 1396 records extracted from 247 OHCA patient episodes. The median number of 9 s segments per record was 3 (1–19, range 1–44). Table
Number of segments (patients in parenthesis) and characteristics of the CC rate and depth for the training and test datasets. Values for CC rate and depth are presented as median with 5–95 percentiles in parenthesis.
Rhythm type | Training | Testing | ||||
---|---|---|---|---|---|---|
9-s seg. | Rate (cpm) | Depth (mm) | 9-s seg. | Rate (cpm) | Depth (mm) | |
Shockable |
|
|
|
|
|
|
Nonshockable |
|
|
|
|
|
|
AS | 1173 (66) | 118 (92 |
35 (18 |
1309 (60) | 117 (89 |
34 (21 |
ORG | 1959 (66) | 114 (86 |
37 (23 |
2041 (76) | 116 (79 |
35 (21 |
|
||||||
Total |
|
|
|
|
|
|
Figures
Features of the SAA for the rhythm types used in each training stage of the algorithm. For the LEA detector the figure compares ASY versus Sh (panels (a) and (b)), and for the SVM classifier ORG versus Sh (panels (c)–(f)). The boxes show the median and interquartile ranges (IQR) and the whisker shows the last datum within the ±1.5IQR interval. Significant differences were found for the median value of the features between the targeted groups (
Figures
The optimized SAA was used to classify the 3 s windows in the test set; Table
Final classification for the 3-s windows and 9-s segments of the test set compared to the AHA performance goals. Sensitivities, specificities and low one-sided 90% CIs (in parenthesis) were obtained using GEE to adjust for clustering.
Rhythm type | 3-s window | 9-s segment |
AHA goal [ | ||
---|---|---|---|---|---|
|
Se/Sp |
|
Se/Sp | ||
Shockable |
|
|
|
|
>90 (for VF) |
Nonshockable |
|
|
|
|
>95 |
AS | 3927 | 94.3 (93.1) | 1309 | 96.5 (95.2) | >95 |
ORG | 6123 | 95.6 (94.6) | 2041 | 96.7 (95.8) | >95 |
Figure
Examples of correctly and incorrectly classified 9 s segments. Examples (a, c) are correctly classified despite the presence of large filtering residuals. However, in the VF of panel (b) spiky filtering artifacts cause the erroneous classification. In the ASY of panel (d) filtering residuals are large in the last two windows causing the shock diagnosis.
VF
VF
ORG
ASY
Processing time for the complete algorithm, CPR suppression filter based on the LMS filter followed by the SAA, was on average 8.7 ms per 3 s segment. Processing time was broken down into 5.8 ms for the LMS filter and 2.9 ms for the SAA. For decisions taken by the LEA detector the SAA required only 1.8 ms, and for windows in which the LEA detector and the SVM were used it increased to 4.1 ms. In the worst case scenario processing time for the complete algorithm was under 10 ms.
This study presents the first attempt to combine two approaches for rhythm analysis during CPR: adaptive filters to suppress the CPR artifact and an SAA optimized to analyze the rhythm after filtering. Our objective was to increase the specificity, because the low specificity of current methods has restrained their implementation in current defibrillators. Our results indicate that our new design approach might contribute to a substantial increase of the accuracy of rhythm analysis methods during CPR, with results that marginally meet AHA performance goals.
The design efforts were focused on obtaining a high specificity during CPR to allow CCs to continue uninterrupted until the method gives a shock advice. The positive predictive value (PPV) of the algorithm, that is, the confidence in a shock diagnosis, must be kept high to avoid unnecessary CPR interruptions if the underlying rhythm is nonshockable. Since VF is the positive class, the PPV depends on the sensitivity/specificity of the algorithm and on the prevalence of VF,
To this date most methods for rhythm analysis during CPR have focused on the accurate detection of shockable rhythms, resulting in higher values for sensitivity than for specificity. Table
Comparative assessment in terms of accuracy and the composition of the databases (% of ASY in nonshockable rhythm in parenthesis) between the method proposed in this study and previous methods tested on OHCA rhythms.
Authors | Method | Accuracy | Testing datasets | ||
---|---|---|---|---|---|
Se (%) | Sp (%) | Sh | NSh | ||
Eilevstjønn et al. [ |
MC-RAMP | 96.7 | 79.9 | 92 | 174 (30%) |
Aramendi et al. [ |
LMS filter | 95.4 | 86.3 | 87 | 285 (31%) |
Tan et al. [ |
ART filter | 92.1 | 90.5 | 114 | 4155 (NA) |
Li et al. [ |
Direct analysis | 93.3 | 88.6 | 1256 | 964 (4%) |
Krasteva et al. [ |
Direct analysis | 90.1 | 86.1 | 172 | 721 (46%) |
Proposed method | Filtering + SAA | 91.0 | 96.6 | 622 | 3350 (39%) |
The characteristics of the OHCA data used in these studies may affect the sensitivity/specificity results, and in particular the characteristics of CPR, the selection criteria for VF, and the proportion of ASY among nonshockable rhythms. Rate and depth values of CPR in our data are similar to those reported in the original studies [
Our study shows that combining adaptive filtering with special SAAs that optimally diagnose the filtered ECG may result in an increased overall accuracy. In addition, the computational cost of the algorithm is low, as shown by the processing time analysis. The SAA algorithm computes at most six ECG features, and implementing our SVM in an AED requires only a few kilobytes of memory for the support vectors and the computation of the discriminant function (see equation (
Finally, several studies need to be completed before any method could be safely taken to the field. First, more conclusive results require testing the algorithm on data recorded by equipment different from those used for this study and with CPR delivered according to the latest 2010 CPR guidelines. In addition, retrospective studies based on complete resuscitation episodes should be conducted. In this way, the impact of using the method on CPR administration could be evaluated. This involves, among other things, a statistical evaluation of whether the method avoids unnecessary CPR interruptions in nonshockable rhythms and unnecessary CPR prolongations in shockable rhythms [
This work introduces a new method for rhythm analysis during CPR with a novel design approach aimed at obtaining a high specificity. The method combines an adaptive LMS filter to suppress the CPR artifact with a new shock/no-shock classification method based on the analysis of the filtered ECG. The method resulted in an increased specificity of 96.6% without compromising the sensitivity, with overall performance figures that met AHA requirements.
The authors declare that there is no conflict of interests regarding the publication of this paper.
This work received financial support from Spanish Ministerio de Economía y Competitividad (Projects TEC2012-31144 and TEC2012-31928), from the UPV/EHU (unit UFI11/16), and from the Basque government (Grants BFI-2010-174, BFI-2010-235, and BFI-2011-166). The authors would like to thank Professor Rojo-Álvarez from the University Rey Juan Carlos (Madrid, Spain) for his assistance with SVM classifiers and for his thorough review of the paper.