| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
We estimated the predictive power with respect to defibrillation outcome of ventricular fibrillation (VF) mean frequency (FREQ), mean peak-to-trough amplitude (AMPL), and their combination. We examined VF electrocardiogram signals of 64 pigs from 4 different cardiac arrest models with different durations of untreated VF, different durations of cardiopulmonary resuscitation, and use of different drugs (epinephrine, vasopressin, N-nitro-L-arginine methyl ester, or saline placebo). The frequency domain was restricted to the range from 4.33 to 30 Hz. In the 10-s epoch between 20 and 10 s before the first defibrillation shock, FREQ and AMPL were estimated. We introduced the survival index (SI; 0.68 Hz-1 · FREQ + 12.69 mV-1 · AMPL) by use of multiple logistic regression. Kruskal-Wallis nonparametric one-way analysis was used to analyze the different porcine models for significant difference. The variables FREQ, AMPL, and SI were compared with defibrillation outcome by means of univariate logistic regression and receiver operating characteristic curves. SI increased predictive power compared with AMPL or FREQ alone, resulting in 89% sensitivity and 86% specificity. The probabilities of predicting defibrillation outcome for FREQ, AMPL, and SI were 0.85, 0.89 and 0.90, respectively. FREQ, AMPL, and SI values were not sensitive in regard to the four different cardiac arrest models but were significantly different for vasopressin and epinephrine animals. IMPLICATIONS: We present a retrospective data analysis to evaluate the predictive power of different ventricular fibrillation electrocardiogram variables in pigs with respect to defibrillation outcome. We showed that our combination of variables leads to an improved forecast, which may help to reduce harmful unsuccessful defibrillation attempts.
During cardiopulmonary resuscitation (CPR), the international guidelines recommend defibrillation 1 to 3 min after drug administration (1). Unfortunately, this strategy does not address individual response to CPR efforts, rendering defibrillation attempts in many cases a matter of chance. Thus, a lucky shock may convert ventricular fibrillation (VF) into return of spontaneous circulation (ROSC) with subsequent long-term survival, whereas less fortunate defibrillation attempts may simply cause massive thermal injury to the heart (2,3), which may cause fatal cardiac failure after ROSC in the intensive care unit. To improve defibrillation success, algorithms to analyze VF wave forms during CPR have been developed. However, VF analysis is usually based on only one variable derived from the electrocardiogram (ECG) signal of the fibrillating heart. It has been demonstrated in clinical and animal studies that the amplitude and frequency of the ECG signal during VF were good predictors of defibrillation outcome (4). When the predictive power of VF variables and threshold values for outcome prediction are compared, careful distinction has to be made between human and animal studies, short versus prolonged duration of untreated VF, VF during CPR and untreated VF, drug effects, and different VF analysis strategies. It is surprising that few attempts have been made to combine VF analysis variables to improve the prediction of defibrillation success. Although predictive power in regard to defibrillation success has been improved by combining median and dominant frequency (5), combining amplitude and the number of baseline crossings per second (6), combining mean amplitude and dominant frequency (7), and combining four spectral features (8), both sensitivity (Se) and specificity (Sp) remained disappointingly low. Moreover, because human VF wave forms are more difficult to analyze than those of animals in VF because of a larger effect of chest compression-related artifacts, it seems obvious that a new VF analysis strategy needs to be extremely good to successfully extrapolate laboratory experience into clinical practice. In this study, we retrospectively examined VF data from 64 pigs undergoing CPR. Furthermore, we introduced a combination of mean peak-to-trough amplitude (AMPL) and mean frequency (FREQ), denominated as survival index (SI). Our hypothesis was that SI is a more effective predictor of defibrillation outcome than AMPL or FREQ.
This project was approved by the Austrian Federal Animal Investigational Committee, and the animals were managed in accordance with American Physiological Society and institutional guidelines. This study was performed according to Utstein-style guidelines (9) on healthy, 12- to 16-wk-old swine (Tyrolean domestic pigs) of either sex weighing 30 to 40 kg. The animals were fasted overnight but had free access to water. The pigs were premedicated with azaperone (neuroleptic drug; 4 mg/kg IM) and atropine (0.1 mg/kg IM) 1 h before surgery, and anesthesia was induced with thiopental (7 to 15 mg/kg IV). After intubation during spontaneous respiration, the pigs were ventilated with a volume-controlled ventilator (EV-A; Draeger, Lübeck, Germany) with 100% oxygen at 20 breaths/min and with a tidal volume adjusted to maintain normocapnia. Anesthesia was maintained with propofol (6 to 8 mg · kg-1 · h-1) and a single dose of piritramide (30 mg). We achieved muscle paralysis with 8 mg of pancuronium after intubation and subsequently with repeated doses of 8 mg of pancuronium as needed. Lactated Ringers solution (6 mL · kg-1 · h-1) and a 3% gelatin solution (4 mL · kg-1 · h-1) was administered in the preparation phase before the induction of cardiac arrest and in the postresuscitation phase. A standard lead III ECG was used to monitor cardiac rhythm; depth of anesthesia was judged according to blood pressure, heart rate, and electroencephalography (Neurotrac; Engström, Munich, Germany). If cardiovascular variables or electroencephalography indicated a reduced depth of anesthesia, we increased the propofol dose, and additional piritramide was given. Body temperature was maintained with a heating blanket between 38.0°C and 39.0°C. A 7F catheter was advanced into the descending aorta via femoral cutdown for measurement of arterial blood pressure. Another 7F catheter was placed into the right atrium via femoral cutdown for drug administration. Blood pressures were measured with saline-filled catheters attached to pressure transducers (model 1290A; Hewlett-Packard, Böblingen, Germany), which were calibrated to atmospheric pressure at the level of the right atrium. The data of four different, but similar, partly previously published models of cardiac arrest (1013), including 64 pigs, were analyzed retrospectively (Fig. 1, Table 1). In one model, after 4 min of cardiac arrest, followed by 3 min of basic life support (BLS) CPR, 16 animals were randomly assigned to receive every 5 min either vasopressin (0.4, 0.4, and 0.8 U/kg; n = 11) or epinephrine (45, 45, and 200 µg/kg; n = 5). Another nine animals were randomly allocated after 4 min of cardiac arrest, followed by 8 min of BLS CPR, to receive every 5 min either vasopressin (0.4 and 0.8 U/kg; n = 5) or epinephrine (45 and 200 µg/kg; n = 4). Defibrillation was attempted after 22 min of cardiac arrest (10,11).
In Model 2, BLS CPR was started after 4 min of cardiac arrest, and 11 animals were randomly assigned to receive N -nitro-L-arginine methyl ester (25 mg/kg; n = 6) or saline placebo (n = 5) after 3 and 13 min of BLS CPR, respectively. Defibrillation was attempted after 22 min of cardiac arrest (12). In Model 3, 30 min before the induction of cardiac arrest, 12 pigs received epidural anesthesia with bupivacaine; another 11 pigs received only a saline administration epidurally. After 1 min of cardiac arrest, followed by 3 min of BLS CPR, epidural animals randomly received every 5 min either epinephrine (45, 45, and 200 µg/kg; n = 6) or vasopressin (0.4, 0.4, and 0.8 U/kg; n = 6); likewise, control animals received every 5 min either epinephrine (45, 45, and 200 µg/kg; n = 6) or vasopressin (0.4, 0.4, and 0.8 U/kg; n = 5). Defibrillation was attempted after 19 min of cardiac arrest (13). In Model 4, after 7 min of cardiac arrest, followed by 3 min of BLS CPR, five animals were randomly assigned to receive every 5 min either vasopressin (0.4, 0.4, and 0.8 U/kg; n = 2) or epinephrine (45, 45, and 200 µg/kg; n = 3). Defibrillation was attempted after 25 min of cardiac arrest. Fifteen minutes before cardiac arrest, 5000 U of heparin was administered IV to prevent intracardiac clot formation, a single dose of 15 mg of piritramide and 8 mg of pancuronium was given, and hemodynamic variables, as well as blood gases, were measured. A 50-Hz, 60-V alternating current was then applied via two subcutaneous needle electrodes to induce VF. Cardiopulmonary arrest was defined as the point at which the aortic pressure decreased profoundly to hydrostatic pressure and the ECG showed VF; ventilation was stopped at that point. Closed-chest CPR was performed manually, and mechanical ventilation was resumed with the same setting as before the induction of cardiac arrest. Chest compression was always performed by the same investigator at a rate of 80/min, guided by acoustical audio tones. This investigator was blinded to hemodynamic and end-tidal carbon dioxide monitor tracings.
All drugs were diluted to 10 mL with normal saline and subsequently injected into the right atrium, which was followed by a 20-mL saline flush (investigators were blinded to the drugs). Up to three countershocks were administered with an energy of 3, 4, and 6 J/kg, respectively. If asystole or pulseless electrical activity was present after defibrillation, the experiment was terminated. ROSC was defined as an unassisted pulse with a systolic arterial blood pressure of The VF ECG signal was monitored continuously and recorded on hard disk by a personal computer-based data acquisition system (Dewetron, Graz, Austria; DASYLab GmbH, Mönchengladbach, Germany; and Datalogger, custom-made software). Digitization was performed at a sampling rate of 1000 Hz and with an amplitude resolution of 12 bits (4096 equal steps between minimal and maximal amplitude). The recorded ECG signals were analyzed with the mathematical software package Matlab (The MathWorks Inc., Natick, MA). Computation of the areas under receiver operating characteristic (ROC) curves and P values of their statistical comparison was performed with the software package GraphROC for Windows (Version 2.0). The signals were divided into consecutive 10-s epochs, and each epoch was transformed into the frequency domain by Fourier transformation. For signal analysis, the frequency domain was restricted to the range from 4.33 to 30 Hz, as previously described (14). In the 10-s epoch between 20 and 10 s before the first defibrillation shock, FREQ and AMPL were estimated. In addition, we introduced the SI by defining
This corresponds to the introduction of a new coordinate axis in the FREQ/AMPL plane by means of linear transformation. The coefficients 0.68, SE 0.29 and 12.69, and SE 4.25 were derived from multiple logistic regression by using maximum likelihood estimation with respect to FREQ and AMPL (explanatory variables) and defibrillation outcome, i.e., ROSC or no ROSC after a maximum of three defibrillation attempts (response variable) (15). For the four different models of cardiac arrest, a Kruskal-Wallis nonparametric one-way analysis of variance was performed to look for possible significant differences of FREQ, AMPL, or SI. For the 24 animals receiving epinephrine and the 29 animals receiving vasopressin, a Kruskal-Wallis nonparametric one-way analysis of variance was performed to look for possible significant differences of FREQ, AMPL, or SI. The data of VF (FREQ, AMPL, or SI) of the 10-s epoch between 20 and 10 s before the first defibrillation shock were labeled with 1 in case of ROSC or 0 in case of no ROSC after a maximum of three defibrillation attempts. Threshold values for AMPL, FREQ, and SI were computed by fitting the labeled data, by using maximum likelihood estimation, to the logistic distribution.
where P is the probability of successful defibrillation (15). Variables to be fitted were m, by definition the threshold value of data, i.e., the value of data associated with a 50% probability of successful defibrillation, and s, the steepness of the logistic distribution. Estimates of the SEs for m and s, and the normalized SEs of m, i.e., divided by the respective m, were computed for all variables. Accuracy (Ac), Se, Sp, positive predictive value (PPV), and negative predictive value (NPV) were computed corresponding to selected data and estimated threshold value. ROC curves were computed for FREQ, AMPL, and SI to determine their value as predictors of successful defibrillation. These curves plot the true-positive rate (Se) versus the false-positive rate (1 - Sp) for different threshold values of the respective variable. The area under the ROC curve represents the probability to which the variable can be used to predict defibrillation outcome. In fact, both the area under the ROC curve and the Wilcoxon statistic measure the probability that in a randomly drawn (ROSC or no ROSC) pair the perceived variable values will allow them to be correctly identified (16). For the paired testing of the significance of the difference of areas under two ROC curves, the method of Hanley and McNeil (17) was used. The P value for comparing the area under the ROC curve of FREQ versus AMPL was a two-tailed significance, whereas the P values for SI versus AMPL and SI versus FREQ were one-tailed significances. Optimal threshold values were computed by maximizing the sum of Se and Sp.
The four different models of cardiac arrest did not show significantly different FREQ, AMPL, or SI values (P > 0.5). Figure 2 shows the data distribution of the four models in the FREQ/AMPL plane.
The 24 animals receiving epinephrine and the 29 animals receiving vasopressin showed significantly different FREQ, AMPL, and SI values (P < 0.0001). Figure 3 shows the data distribution of epinephrine and vasopressin animals in the FREQ/AMPL plane.
Figure 4 shows the distribution of all ROSC and no-ROSC data in the FREQ/AMPL plane. FREQ was capable of predicting defibrillation outcome well for values >8.10 Hz (upper threshold) or values <6.80 Hz (lower threshold). AMPL was capable of predicting defibrillation outcome well for values >0.24 mV (upper threshold) or values <0.13 mV (lower threshold). Between these upper and lower thresholds, FREQ and AMPL were poorly predictive.
Figure 5 shows the distribution of ROSC and no-ROSC data, the logistic regression line, and threshold (P50), P5, and P95 for FREQ, AMPL, and SI, respectively. Table 2 shows the estimates for the logistic regression fitting variables m and s together with their SEs; normalized SEs of threshold values; and Ac, Se, Sp, PPV, and NPV regarding FREQ, AMPL, and SI. All values for Ac, Se, Sp, PPV, and NPV of SI were superior to FREQ or AMPL, resulting in 89% Se and 86% Sp. The predictive power of AMPL is comparable to that of FREQ, but the normalized SE of the AMPL threshold is greater compared with FREQ or SI.
The ROC curves (Fig. 6) have been plotted by computing Se and Sp of predicting ROSC/no ROSC with different thresholds for FREQ, AMPL, and SI. The probabilities of predicting defibrillation outcome, i.e., the area under the ROC curve, for FREQ, AMPL, and SI were 0.85 SE 0.05, 0.89 SE 0.05 and 0.90 SE 0.04, respectively. The two-tailed P value for comparison of the areas under the ROC curve for AMPL versus FREQ was 0.51. The one-tailed P values for SI versus FREQ and SI versus AMPL were 0.09 and 0.33, respectively. Optimal thresholds corresponding to ROC curve analysis for FREQ, AMPL, and SI were 7.59 Hz, 0.13 mV, and 6.84, respectively.
This analysis of different animal studies shows that FREQ, AMPL, and SI values, computed shortly before the first defibrillation shock, were not significantly different in regard to the four different cardiac arrest models. This legitimates the pooling of the data. Many human and animal studies have shown that FREQ and AMPL are reliable variables for the prediction of defibrillation outcome (4). In comparison with AMPL, FREQ is independent of electrode contact quality, which makes FREQ more practicable and predictive in humans. In this study, FREQ and AMPL were both capable of predicting defibrillation outcome. Unfortunately, there is a range of FREQ and AMPL between an upper and a lower threshold where prediction of ROSC or no ROSC is poor because of an overlap of ROSC and no-ROSC data. To improve this dilemma of a foggy prediction in some circumstances, we introduce the SI, which corresponds to a new coordinate axis in the FREQ/AMPL plane by means of linear transformation. By using multiple logistic regression analysis, we identified the relevant coefficients of FREQ and AMPL in predicting defibrillation outcome, resulting in the definition of SI. Univariate logistic regression and ROC curve analysis showed that SI had a robust threshold value and predicted defibrillation outcome with better probabilities compared with FREQ or AMPL alone. There are more sophisticated (e.g., nonlinear) methods of combining FREQ and AMPL to a new VF variable, which would probably perform better than SI. However, this would likely lead to over-fitting of the 64 data points. The P5 value of logistic regression for AMPL is negative (Fig. 5). This reflects a limitation of logistic regression when it is applied to data for which only positive values make sense. A way out would be to work with the logarithm of AMPL values. However, for the sake of mathematical comprehensibility, we accepted this inconsistency. In this study, we chose the P50 value of logistic linear regression as the definition of threshold value for a VF variable. One could also choose a different P value. ROC curve analysis rendered slightly different optimal threshold values. Ultimately, an optimal threshold value has to be defined according to clinical relevance. Further analysis has to be performed to investigate whether CPR itself or additional drug administration changes the pattern of VF by means of increased SI and a better probability of ROSC. In fact, FREQ and AMPL served as noninvasive markers to monitor continuing CPR efforts and were sensitive to the administration of vasopressin and epinephrine (18,19). In this study, which analyzes FREQ, AMPL, and SI values shortly before the first defibrillation shock, we showed that animals receiving vasopressin showed significantly higher FREQ, AMPL, and SI values than animals receiving epinephrine.
There are a number of studies that combine various variables of VF rather than comparing these variables. For example, Brown and Dzwonczyk (5) retrospectively analyzed 55 VF patients out of hospital, and they found that the combination of median and dominant frequency improved predictive power when compared with the analysis of median frequency alone. This group reported a Sp of 47% when the threshold value was set to yield a sensitivity of 100%. Furthermore, Monsieurs et al. (6) combined amplitude and the number of baseline crossings per second in human VF into a survival index. With this index, 79% of the survivors and 70% of the nonsurvivors could be classified correctly; adding age increased the correct classification of survivors to 86% and to 73% for the nonsurvivors. In a laboratory study, Noc et al. (7) divided animals into a derivation and a validation group. When mean amplitude and dominant frequency were combined, predictability was improved compared with single use. In the validation group, defibrillation attempts were uniformly unsuccessful if the combination of mean amplitude and dominant frequency did not exceed the threshold values obtained in the derivation study. In the most recent investigation, Eftestol et al. (8) predicted defibrillation outcome by combining four spectral features in VF ECG signals before 868 defibrillation attempts in 156 patients with out-of hospital cardiac arrest. They split the data into training and test sets and used classifier generalization techniques to increase the degree of expected reliability. Different secondary decorrelated feature sets were generated from the four original features by using principal component analysis. Following the advice of the highest performing classifier, which corresponded to the combination of 2 secondary features, 328 (42%) of 781 unsuccessful shocks would have been avoided, whereas 7 (8%) of 87 successful shocks would not have been given. Although it seems obvious that combining VF variables improves predictability of successful defibrillation, it seems obvious as well that this approach still has significant limitations. For example, it has to be noted that a 47% Sp, as observed in Brown and Dzwonczyks study (5), is as good as flipping a coin; furthermore, if 8% of successful shocks are withheld, as in the report of Eftestol et al. (8), efficiency of this strategy is dissatisfying. Eftestol et al. recognized this problem as well and suggested that the low Sp and PPV indicate that other features should be added; this may be extremely important in humans undergoing CPR, because VF analysis in this setting is especially prone to chest compression-related artifacts. In this context, the method of N( In conclusion, FREQ, AMPL, and SI values were not sensitive in regard to the four different cardiac arrest models but were significantly different for vasopressin and epinephrine animals. Further, we showed that a combination of FREQ and AMPL with a new predictive variable, SI, leads to a better forecast of successful defibrillation compared with either FREQ and AMPL alone.
Supported in part by the Austrian National Bank science projects 7280 and 7276, the Austrian Heart Foundation project 98/05, and the Austrian Science Foundation P-14169-MED, all of Vienna, Austria; a research grant from Bruker Medical, Ettlingen, Germany; departmental funds; and a Science Foundation grant of the University of Innsbruck, Austria.
Presented in part at the Austrian Society of Anesthesiology, Resuscitation and Intensive Care Medicine Congress, Vienna, Austria, September, 2001, and the 23rd annual meeting of the European Academy of Anesthesiology, Graz, Austria, August, 2001.
This article has been cited by other articles:
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|