Capnography: A support tool for the detection of return of spontaneous circulation in out-of-hospital cardiac arrest

Andoni Elola, Elisabete Aramendi, Unai Irusta, Erik Alonso, Yuanzheng Lu, Mary P. Chang, Pamela Owens, Ahamed H. Idris. Capnography: A support tool for the detection of return of spontaneous circulation in out-of-hospital cardiac arrest. Resuscitation, Volume 142, 2019, Pages 153-161, ISSN 0300-9572, https://doi.org/10.1016/j.resuscitation.2019.03.048. (https://www.sciencedirect.com/science/article/pii/S0300957219301327) Abstract Background Automated detection of return of spontaneous circulation (ROSC) is still an unsolved problem during cardiac arrest. Current guidelines recommend the use of capnography, but most automatic methods are based on the analysis of the ECG and thoracic impedance (TI) signals. This study analysed the added value of EtCO2 for discriminating pulsed (PR) and pulseless (PEA) rhythms and its potential to detect ROSC. Materials and methods A total of 426 out-of-hospital cardiac arrest cases, 117 with ROSC and 309 without ROSC, were analysed. First, EtCO2 values were compared for ROSC and no ROSC cases. Second, 5098 artefact free 3-s long segments were automatically extracted and labelled as PR (3639) or PEA (1459) using the instant of ROSC annotated by the clinician on scene as gold standard. Machine learning classifiers were designed using features obtained from the ECG, TI and the EtCO2 value. Third, the cases were retrospectively analysed using the classifier to discriminate cases with and without ROSC. Results EtCO2 values increased significantly from 41 mmHg 3-min before ROSC to 57 mmHg 1-min after ROSC, and EtCO2 was significantly larger for PR than for PEA, 46 mmHg/20 mmHg (p < 0.05). Adding EtCO2 to the machine learning models increased their area under the curve (AUC) by over 2 percentage points. The combination of ECG, TI and EtCO2 had an AUC for the detection of pulse of 0.92. Finally, the retrospective analysis showed a sensitivity and specificity of 96.6% and 94.5% for the detection of ROSC and no-ROSC cases, respectively. Conclusion Adding EtCO2 improves the performance of automatic algorithms for pulse detection based on ECG and TI. These algorithms can be used to identify pulse on site, and to retrospectively identify cases with ROSC. Keywords: Return of spontaneous circulation (ROSC), ROSC detection, Capnography, End-tidal CO2 (EtCO2), Electrocardiogram (ECG), Thoracic impedance Capnography: A support tool for the Detection of Return of Spontaneous Circulation in Out-of-Hospital Cardiac Arrest Andoni Elola∗,a, Elisabete Aramendia, Unai Irustaa, Erik Alonsoa, Yuanzheng Lub, Mary P. Changc, Pamela Owensc, Ahamed H. Idrisc Affiliation and addresses: aCommunications Engineering Department. University of the Basque Country UPV/EHU. 48013 Bilbao, Spain bEmergency and Disaster Medicine Center. The Seventh Affiliated Hospital, Sun Yat-sen University. Shenzhen, China cDepartment of Emergency Medicine. University of Texas SouthWestern Medical Center (UTSW). Dallas, United States Corresponding author: ∗ Andoni Elola email: [email protected] Tel. : +34946013956 Fax. : +34946014259 Word counts: 3255 Abstract word counts: 281 Abstract Background: Automated detection of return of spontaneous circulation (ROSC) is still an unsolved problem during cardiac arrest. Current guidelines recommend the use of capnography, but most automatic methods are based on the analysis of the ECG and thoracic impedance (TI) signals. This study analysed the added value of EtCO2for discriminating pulsed (PR) and pulseless (PEA) rhythms and its potential to detect ROSC. Materials and methods: A total of 426 out-of-hospital cardiac arrest cases, 117 with ROSC and 309 without ROSC, were analysed. First, EtCO2values were compared for ROSC and no ROSC cases. Second, 5098 artefact free 3-second long segments were automatically extracted and labelled as PR (3639) or PEA (1459) using the instant of ROSC annotated by the clinician on scene as gold standard. Machine learning classifiers were designed using features obtained from the ECG, TI and the EtCO2value. Third, the cases were retrospectively analysed using the classifier to discriminate cases with and without ROSC. Results: EtCO2values increased significantly from 41 mmHg 3-min before ROSC to 57 mmHg 1-min after ROSC, and EtCO2was significantly larger for PR than for PEA, 46 mmHg/20 mmHg (p < 0.05). Adding EtCO2to the machine learning models increased their area under the curve (AUC) by over 2 percentage points. The combination of ECG, TI and EtCO2had an AUC for the detection of pulse of 0.92. Finally, the retrospective analysis showed a sensitivity and specificity of 96.6% and 94.5% for the detection of ROSC and no-ROSC cases, respectively. Conclusion: Adding EtCO2improves the performance of automatic algorithms for pulse detection based on ECG and TI. These algorithms can be used to identify pulse on site, and to retrospectively identify cases with ROSC. Keywords Return of spontaneous circulation (ROSC), ROSC detection, capnography, end-tidal CO2(EtCO2), electrocardiogram (ECG), thoracic impedance 1. Introduction1 The main goal of resuscitative efforts during out-of-hospital cardiac arrest (OHCA) is to achieve2 return of spontaneous circulation (ROSC). Those efforts include high quality cardiopulmonary3 resuscitation (CPR), during which chest compressions should be minimally interrupted for actions4 like rhythm analysis or pulse checks. Current pulse detection methods such as carotid pulse check,5 or checking for signs of life as recommended by the current guidelines, are both time consuming and6 inaccurate [1–5]. There is therefore a need for accurate and automated pulse detection methods [6]7 that can be used by emergency medical personnel as a decision support tool to identify ROSC. Such8 methods would contribute to improve therapy, reduce and shorten pauses in chest compressions,9 and increase survival rates [7, 8].10 Current guidelines support the use of capnography for early detection of ROSC [9]. Higher11 values of end tidal CO2(EtCO2), and sudden increases in EtCO2have been linked to ROSC in12 OHCA [10–13]. Although some medical algorithms exist for the detection of ROSC using EtCO2 13 values [14], the only automatic method based on capnography was recently proposed [15].14 Most automatic methods for the detection of pulse in OHCA rest on the analysis of the ECG and15 the thoracic impedance (TI). The TI signal shows low amplitude fluctuations for every effective16 heartbeat [16], so features characterizing the TI signal have been proposed alone [17–19], or in17 combination with ECG features [20–22] for the detection of pulse. In this context, detection of18 pulse is framed as a classification problem with two types of organized rhythms: pulse-generating19 rhythms (PR) and pulseless electrical activity (PEA).20 The purpose of this study was to evaluate the added value of capnography for the classification21 of PR/PEA during OHCA. First, EtCO2values were automatically detected in order to compare22 the values between patients with and without ROSC, and to analyse how EtCO2changed as23 the patient approached ROSC. Then, the added value of EtCO2for PR/PEA classification was24 evaluated by developing machine learning PR/PEA classifiers.25 2. Materials26 For this study we analysed 1561 OHCA episodes retrospectively, treated by the Dallas27 FortWorth Center for Resuscitation Research (UTSW, Dallas) using the Philips HeartStart MRx28 device between 2012 and 2016. The device files included the ECG and TI recorded through29 the defibrillation pads with sampling frequencies of 250 Hz/200 Hz respectively, and capnography30 recorded through sidestream acquisition with a sampling frequency of 125 Hz. The electronic files31 were linked to clinical annotations and ROSC was defined as palpable pulse in any vessel for any32 length of time. The first ROSC instant annotated by the rescuer on scene was the gold standard;33 based on that instant PR and PEA annotations were made automatically and patients with ROSC34 and without ROSC were classified.35 The following patient inclusion/exclusion criteria were applied. Only episodes with TI, ECG36 and capnography were considered (n=835). Cases where ROSC was suspected but not annotated37 by clinicians on site were excluded, which comprised patients transported to hospital (n=252), or38 episodes with long periods (>2 min) without compressions presenting an organized rhythm with39 EtCO2above 25 mmHg (n=26). Episodes with suspected intermittent ROSC were also excluded,40 these were episodes in which shocks or chest compressions (>2 min) were delivered after the41 annotated onset of ROSC (n=76). For our analysis of the ROSC cases, the capnogram had to42 be available at least 4 minutes before and 1 minute after the onset of ROSC. If not, the case43 was excluded (n=55). The final dataset contained 426 episodes, 117 with ROSC and 309 without44 ROSC.45 Figure 1 shows a 3-minute interval from two cases of the study dataset. In the ROSC case (top46 panel) EtCO2increases at ROSC onset, and after ROSC the heart rate increases and there is pulse47 related activity in the TI. In the no-ROSC case (bottom panel) EtCO2is always below 20 mmHg,48 and although the heart rate changes during PEA there is no pulse related activity in the TI.49 3. Methods50 Three analyses were conducted: EtCO2levels in episodes with and without ROSC, development51 and evaluation of a PR/PEA classifier using ECG/TI segments and EtCO2values, and a case study52 of the use of the classifier to retrospectively identify cases as ROSC/no-ROSC.53 3.1. Analysis of EtCO2levels54 Onset and offset of each ventilation were automatically delineated in the capnogram using55 a method introduced in a previous study [23]. For each ventilation, EtCO2was automatically56 calculated as the maximum CO2value during the alveolar plateau (see Figure 1).57 In ROSC cases, median EtCO2levels were computed every minute (MEtCO2) in a five minute58 interval around ROSC (4-min before to 1-min after). Similarly, for patients without ROSC, the59 MEtCO2values were computed for each one of the last five minutes of the episode. The MEtCO2 60 value for the last minute of the episode corresponds to the last minute before the EOE, i.e. the61 instant when the monitor/defibrillator was disconnected.62 3.2. PR/PEA machine learning classifier63 Following the classical scheme proposed in previous studies [20–22], the detection of ROSC64 implies the discrimination between PR and PEA once an organized rhythm is identified by the65 shock advice algorithm. It is therefore a two class classification problem, for which, first, the dataset66 of PR/PEA segments was defined, and then a classifier was designed using features extracted from67 the ECG, TI and capnography signals.68 3.2.1. PR/PEA segment dataset69 PR and PEA segments of 3.2-s duration were extracted during intervals with no chest70 compression artefacts. Pauses in chest compressions were automatically detected using the71 compression depth signal from the CPR assist pad when available [24], or the TI otherwise [25].72 Segments with large ECG amplitude oscillations (>3.5 mV) were discarded as noisy, and then73 organized rhythms (PEA or PR) were detected during the pauses using an offline version of a74 rhythm analysis algorithm of a commercial automated external defibrillator (AED) [26]. In ROSC75 cases, all segments before ROSC onset were labelled as PEA, and those after ROSC onset as PR76 (see Figure 1, panel a). In no-ROSC cases, all segments were labelled as PEA (see Figure 1, panel77 b). A minimum separation between consecutive segments of 20-s was enforced to foster ECG and78 TI waveform diversity in the segments.79 3.2.2. Machine learning PR/PEA classifier80 Nine PR/PEA classification features were computed from the most recently proposed81 algorithms, six ECG features introduced in [27], and three TI features [17, 22, 28]. These ECG82 and TI features are described in detail in Appendix A. The MEtCO2, the median EtCO2in the83 minute before the analysis window (pause in chest compressions with organized ECG rhythm) was84 also added. The features were combined in a Random Forest (RF) classifier, a machine learning85 algorithm based on the aggregate vote of several independently designed uncorrelated decision trees86 [29]. RF classifiers have shown excellent performance in many classification problems, including87 PR/PEA classification [27], and are robust against annotation errors.88 All patients were weighted equally to train the RF classifier and 300 trees were used. For89 each segment, the RF classifier computes the probability of being PR (ppr), and segments were90 classified as PR for ppr >0.5 and as PEA otherwise. The classifier was trained and tested using91 a patient wise 10-fold cross-validation procedure [30]. For each of the 10 folds, the algorithm was92 optimized using 90% of the cases, and the accuracy results were obtained from the remaining 10%93 (test fold). This procedure guaranteed that the optimization of the classifier and the estimation94 of its accuracy were done on data from separate patients, and that the performance was assessed95 using all available data.96 3.3. Case study: Retrospective identification of patients with ROSC97 Using the PR/PEA classifier, a simple method was developed to automatically identify patients98 with ROSC in a retrospective analysis of a set of OHCA episodes. This method may be used as99 an automated tool for post arrest debriefing or annotation. Complete episodes (until EOE) were100 processed and the case was labelled as ROSC if from any three consecutive segments at least two101 were identified as PR by the classifier.102 Our ground truth was the ROSC instant annotated by clinicians on scene, which discriminated103 the group of patients with ROSC and patients without ROSC, and the detection of episodes with104 ROSC was evaluated using the test sets in the 10-fold cross validation procedure.105 3.4. Statistical analysis106 MEtCO2distributions did not pass the Kolmogorov-Smirnov normality test, and are reported107 as median and interquartile range (IQR). MEtCO2distributions at different times (within108 ROSC cases) or between ROSC/no-ROSC cases were compared using the Mann-Whitney U test.109 Differences were considered significant for p < 0.05.110 PR/PEA classification was evaluated using Receiver Operating Characteristic (ROC) curves,111 and the area under the curve (AUC) was used as measure of performance [31]. The Youden index112 was used to define the optimal point in the ROC curve, which gives equal importance to the113 sensitivity (SE, for PR segments) and specificity (SP, for PEA segments) [32].114 When the classifier was used as a retrospective tool to identify ROSC, SE and SP were defined115 as the proportion of correctly identified ROSC and no-ROSC cases, respectively.116 4. Results117 The mean (standard deviation) durations were 58 (23) min and 38 (11) min for the episodes with118 and without ROSC, respectively. The commercial AED algorithm detected 5098 segments with119 organized rhythms. A total of 3639 PR segments were extracted from episodes with ROSC, and120 1459 PEA segments, 308 from episodes with ROSC and 1151 from episodes without ROSC. Some121 examples of the extracted ECG segments can be found in Figure 1 and Figure 4. The median122 (IQR) ventilation rate per episode was 7.8 (5.7-10.5) min−1.123 The MEtCO2for ROSC cases were statistically significantly larger than for no-ROSC cases at124 all time-stamps (Figure 2). Elevated EtCO2levels were observed in patients with ROSC, with an125 upward trend from 41 mmHg (at 3 min before ROSC) to 57 mmHg close to ROSC onset (see Figure126 2 a).127 Figure 3 shows the ROC curves of the RF classifier for different features sets. The curves in128 panel (a) were calculated using the whole dataset, while the curves in panel (b) were calculated129 excluding the PEA segments extracted from patients with ROSC, that is the pre-ROSC PEA130 segments. The analysis of the ROC curves is shown in Table 1. The ROC curves showed that the131 AUC of the PR/PEA classifier increased as features from different sources were added. Including132 MEtCO2in the classifier increased the AUC for all feature combinations, thanks to the added133 uncorrelated information. Adding MEtCO2to an ECG-only and to an ECG+TI based classifiers134 increased their AUCs in 3 and 2-points, respectively. The best classifier combined all features135 and presented an AUC of 0.92 with a SE and SP of 84% and 86%, respectively (see Table 1).136 MEtCO2alone was also a good classifier (AUC around 0.76), the median MEtCO2values were 46137 (32-64) mmHg for PR and 20 (8-38) mmHg for PEA segments (p < 0.05).138 The accuracy of the classifiers increased when PEAs that transitioned to PR (episodes with139 ROSC) were not included. The accuracy increase was on average 4-points for all classifiers (see140 Table 1). Significant differences were observed between PEAs in ROSC and no-ROSC cases.141 The MEtCO2values of the PEA in the ROSC and no-ROSC cases were 31 (20-44) mmHg and 16142 (7-35) mmHg (p < 0.05), respectively. The probabilities of being PR, ppr, for the classifier with all143 features were also significantly different for these two subgroups of PEA, 0.09 (0.03-0.29) for the144 PEA from no-ROSC cases and 0.33 (0.11-0.58) for those of the ROSC cases (p < 0.05).145 Figure 4 shows the performance of the PR/PEA classifier with three consecutive segments in146 [1] Perkins GD, Stephenson B, Hulme J, Monsieurs KG. Birmingham assessment of breathing study (BABS).289 Resuscitation 2005;64(1):109–113.290 [2] Ruppert M, Reith MW, Widmann JH, et al. Checking for breathing: evaluation of the diagnostic capability291 of emergency medical services personnel, physicians, medical students, and medical laypersons. Annals of292 emergency medicine 1999;34(6):720–729.293 [3] Nyman J, Sihvonen M. Cardiopulmonary resuscitation skills in nurses and nursing students. Resuscitation 2000;294 47(2):179–184.295 [4] Eberle B, Dick W, Schneider T, Wisser G, Doetsch S, Tzanova I. Checking the carotid pulse check: diagnostic296 accuracy of first responders in patients with and without a pulse. Resuscitation 1996;33(2):107–116.297 [5] Ochoa FJ, Ramalle-Gomara E, Carpintero J, Garcıa A, Saralegui I. Competence of health professionals to check298 the carotid pulse. Resuscitation 1998;37(3):173–175.299 [6] Babbs CF. We still need a real-time hemodynamic monitor for CPR. Resuscitation 2013;84(10):1297–1298.300 [7] Christenson J, Andrusiek D, Everson-Stewart S, et al. Chest compression fraction determines survival in patients301 with out-of-hospital ventricular fibrillation. Circulation 2009;120(13):1241–1247.302 [8] Berg RA, Sanders AB, Kern KB, et al. Adverse hemodynamic effects of interrupting chest compressions for303 rescue breathing during cardiopulmonary resuscitation for ventricular fibrillation cardiac arrest. Circulation304 2001;104(20):2465–2470.305 [9] Soar J, Nolan JP, B¨ottiger BW, et al. European resuscitation council guidelines for resuscitation 2015: section306 3. Adult advanced life support. Resuscitation 2015;95:100–147.307 [10] Hartmann SM, Farris RW, Di Gennaro JL, Roberts JS. Systematic review and meta-analysis of end-tidal carbon308 dioxide values associated with return of spontaneous circulation during cardiopulmonary resuscitation. Journal309 of intensive care medicine 2015;30(7):426–435.310 [11] Sheak KR, Wiebe DJ, Leary M, et al. Quantitative relationship between end-tidal carbon dioxide and CPR311 quality during both in-hospital and out-of-hospital cardiac arrest. Resuscitation 2015;89:149–154.312 [12] Pokorn´a M, Neˇcas E, Kratochv´ıl J, Skˇripsk`y R, Andrl´ık M, Franˇek O. A sudden increase in partial pressure313 end-tidal carbon dioxide (PETCO2) at the moment of return of spontaneous circulation. The Journal of314 emergency medicine 2010;38(5):614–621.315 [13] Lui CT, Poon KM, Tsui KL. Abrupt rise of end tidal carbon dioxide level was a specific but non-sensitive316 marker of return of spontaneous circulation in patient with out-of-hospital cardiac arrest. Resuscitation 2016;317 104:53–58.318 [14] Davis DP, Sell RE, Wilkes N, et al. Electrical and mechanical recovery of cardiac function following319 out-of-hospital cardiac arrest. Resuscitation 2013;84(1):25–30.320 [15] Brinkrolf P, Borowski M, Metelmann C, Lukas RP, Pidde-K¨ullenberg L, Bohn A. Predicting ROSC in321 out-of-hospital cardiac arrest using expiratory carbon dioxide concentration: Is trend-detection instead of322 absolute threshold values the key? Resuscitation 2018;122:19–24.323 [16] Pellis T, Bisera J, Tang W, Weil MH. Expanding automatic external defibrillators to include automated detection324 of cardiac, respiratory, and cardiorespiratory arrest. Critical care medicine 2002;30(4):S176–S178.325 [17] Cromie NA, Allen JD, Navarro C, Turner C, Anderson JM, Adgey AAJ. Assessment of the impedance326 cardiogram recorded by an automated external defibrillator during clinical cardiac arrest. Critical care medicine327 2010;38(2):510–517.328 [18] Cromie NA, Allen JD, Turner C, Anderson JM, Adgey AAJ. The impedance cardiogram recorded through two329 electrocardiogram/defibrillator pads as a determinant of cardiac arrest during experimental studies. Critical330 Care Medicine 2008;36(5):1578–1584.331 [19] Losert H, Risdal M, Sterz F, et al. Thoracic-impedance changes measured via defibrillator pads can monitor332 signs of circulation. Resuscitation 2007;73(2):221–228.333 [20] Risdal M, Aase SO, Kramer-Johansen J, Eftesol T. Automatic identification of return of spontaneous circulation334 during cardiopulmonary resuscitation. IEEE Transactions on Biomedical Engineering 2008;55(1):60–68.335 [21] Alonso E, Aramendi E, Daya M, et al. Circulation detection using the electrocardiogram and the thoracic336 impedance acquired by defibrillation pads. Resuscitation 2016;99:56–62.337 [22] Ruiz JM, de Gauna SR, Gonz´alez-Otero DM, et al. Circulation assessment by automated external defibrillators338 during cardiopulmonary resuscitation. Resuscitation 2018;128:158–163.339 [23] Aramendi E, Elola A, Alonso E, et al. Feasibility of the capnogram to monitor ventilation rate during340 cardiopulmonary resuscitation. Resuscitation 2017;110:162–168.341 [24] Ayala U, Eftestøl T, Alonso E, et al. Automatic detection of chest compressions for the assessment of342 CPR-quality parameters. Resuscitation 2014;85(7):957–963.343 [25] Alonso E, Ruiz J, Aramendi E, et al. Reliability and accuracy of the thoracic impedance signal for measuring344 cardiopulmonary resuscitation quality metrics. Resuscitation 2015;88:28–34.345 [26] Irusta U, Ruiz J, Aramendi E, de Gauna SR, Ayala U, Alonso E. A high-temporal resolution algorithm to346 discriminate shockable from nonshockable rhythms in adults and children. Resuscitation 2012;83(9):1090–1097.347 [27] Elola A, Aramendi E, Irusta U, Del Ser J, Alonso E, Daya M. ECG-based pulse detection during cardiac arrest348 using random forest classifier. Medical & biological engineering & computing 2019;57(2):453–462.349 [28] Wei L, Chen G, Yang Z, Yu T, Quan W, Li Y. Detection of spontaneous pulse using the acceleration signals350 acquired from CPR feedback sensor in a porcine model of cardiac arrest. PloS one 2017;12(12):e0189217.351 [29] Breiman L. Random forests. Machine learning 2001;45(1):5–32.352 [30] Stone M. Cross-validatory choice and assessment of statistical predictions. Journal of the royal statistical society353 Series B (Methodological) 1974;:111–147.354 [31] Zou KH, O’Malley AJ, Mauri L. Receiver-operating characteristic analysis for evaluating diagnostic tests and355 predictive models. Circulation 2007;115(5):654–657.356 [32] Ruopp MD, Perkins NJ, Whitcomb BW, Schisterman EF. Youden Index and optimal cut-point estimated from357 observations affected by a lower limit of detection. Biometrical Journal: Journal of Mathematical Methods in358 Biosciences 2008;50(3):419–430.359 [33] Kolar M, Kriˇzmari´c M, Klemen P, Grmec ˇ S. Partial pressure of end-tidal carbon dioxide successful predicts360 cardiopulmonary resuscitation in the field: a prospective observational study. Critical care 2008;12(5):R115.361 [34] Touma O, Davies M. The prognostic value of end tidal carbon dioxide during cardiac arrest: a systematic362 review. Resuscitation 2013;84(11):1470–1479.363 [35] Grmec ˇ S, Klemen P. Does the end-tidal carbon dioxide (EtCO2) concentration have prognostic value during364 out-of-hospital cardiac arrest? European Journal of Emergency Medicine 2001;8(4):263–269.365 [36] Weiser C, Poppe M, Sterz F, et al. Initial electrical frequency predicts survival and neurological outcome in out366 of hospital cardiac arrest patients with pulseless electrical activity. Resuscitation 2018;125:34–38.367 [37] Skjeflo GW, Nordseth T, Loennechen JP, Bergum D, Skogvoll E. ECG changes during resuscitation of patients368 with initial pulseless electrical activity are associated with return of spontaneous circulation. Resuscitation369 2018;127:31–36.370 [38] Elola A, Aramendi E, Irusta U, et al. ECG characteristics of Pulseless Electrical Activity associated with Return371 of Spontaneous Circulation in Out-of-Hospital Cardiac Arrest. Resuscitation 2018;130:e54.372 [39] Prosen G, Kriˇzmari´c M, Zavrˇsnik J, Grmec ˇ S. Impact of modified treatment in echocardiographically confirmed373 pseudo-pulseless electrical activity in out-of-hospital cardiac arrest patients with constant end-tidal carbon374 dioxide pressure during compression pauses. Journal of International Medical Research 2010;38(4):1458–1467.375 [40] Raimondi M, Savastano S, Pamploni G, Molinari S, Degani A, Belliato M. End-tidal carbon dioxide monitoring376 and load band device for mechanical cardio-pulmonary resuscitation: Never trust the numbers, believe at the377 curves. Resuscitation 2016;103:e9–e10.378 [41] Ruiz J, Alonso E, Aramendi E, et al. Reliable extraction of the circulation component in the thoracic impedance379 measured by defibrillation pads. Resuscitation 2013;84(10):1345–1352.380 Figure Legends381 Figure 1 ECG, Thoracic Impedance (TI) and capnography signals for a patient382 with ROSC, panel (a), and without ROSC, panel (b). ROSC onset,383 as annotated by a clinician on site, is represented by a red line in384 the first example. The extracted 3.2-s segments are shaded in grey385 and the ECG and TI (green) are zoomed in. Chest compression386 intervals are depicted above TI signal. In the ROSC case a PEA387 and a PR segments were extracted in the depicted interval, and two388 PEA segments in the no-ROSC case. Ventilations were automatically389 detected in the CO2curve, and the automatically measured EtCO2 390 value is highlighted with red dots. In the ROSC case after pulse391 recovery the ECG presents stable and normal QRS complexes and392 heart rate, and chest compressions are stopped so there is no activity393 in the impedance.394 Figure 2 Median EtCO2(MEtCO2) values and their interquartile ranges for395 cases with ROSC (left) and no-ROSC (right). For ROSC cases the396 interval around ROSC onset is analysed, in the no-ROSC cases the397 5 min before the end of episode (EOE) are shown. MEtCO2was398 calculated as the median EtCO2value of all ventilations in a 1-minute399 interval before the indicated time-stamp.400 Figure 3 ROC curves of the RF classifier for different feature sets. Panel (a)401 shows results for the whole dataset, while panel (b) shows the curves402 after excluding the PEAs from episodes with ROSC. The AUC value403 for each classifier is shown between parentheses.404 Figure 4 Examples of the case study. Panels (a), (b) and (c) show a405 correctly identified patient with ROSC, a correctly identified patient406 without ROSC, and a patient without ROSC incorrectly identified,407 respectively. Each panel depicts the three consecutive PEA/PR408 segments analysed. The text on top of each segment indicates its true409 label followed by the predicted label by the classifier. The capnogram410 corresponds to the minute before the onset of the segment and the411 dashed horizontal line represents the MEtCO2.412 Figure 5 Time evolution of ppr for the PEA segments as the patients approach413 ROSC. Blue dots indicate values for each segment, and the red curve414 is fitted to the median values of ppr every 2 minutes.415 15 -2 0 2 ROSC ECG (mV) -1000 0 1000 TI (m ) 11:08:20 11:09:10 11:10:00 11:10:50 11:11:40 11:12:30 11:13:20 0 50 100 CO2 (mmHg) (a) A case of a patient with ROSC PEA PR -2 0 2 ECG (mV) -1000 0 1000 TI (m ) 13:12:00 13:12:50 13:13:40 13:14:30 13:15:20 13:16:10 13:17:00 0 50 100 CO2 (mmHg) (b) A case of a patient without ROSC PEA PEA Figure 1: ECG, Thoracic Impedance (TI) and capnography signals for a patient with ROSC, panel (a), and without ROSC, panel (b). ROSC onset, as annotated by a clinician on site, is represented by a red line in the first example. The extracted 3.2-s segments are shaded in grey and the ECG and TI (green) are zoomed in. Chest compression intervals are depicted above TI signal. In the ROSC case a PEA and a PR segments were extracted in the depicted interval, and two PEA segments in the no-ROSC case. Ventilations were automatically detected in the CO2curve, and the automatically measured EtCO2value is highlighted with red dots. In the ROSC case after pulse recovery the ECG presents stable and normal QRS complexes and heart rate, and chest compressions are stopped so there is no activity in the impedance. -3 min -2 min -1 min ROSC 1 min time (min) 10 20 30 40 50 60 70 80 MEtCO2 (mmHg) (a) ROSC cases -4 min -3 min -2 min -1 min EOE time (min) 10 20 30 40 50 60 70 80 MEtCO2 (mmHg) (b) No-ROSC cases Figure 2: Median EtCO2(MEtCO2) values and their interquartile ranges for cases with ROSC (left) and no-ROSC (right). For ROSC cases the interval around ROSC onset is analysed, in the no-ROSC cases the 5 min before the end of episode (EOE) are shown. MEtCO2was calculated as the median EtCO2value of all ventilations in a 1-minute interval before the indicated time-stamp. 0 20 40 60 80 100 100-SP (%) 0 20 40 60 80 100 SE (%) (a) All PR/PEA segments ETCO2 (0.76) ECG (0.88) ECG+TI (0.90) ECG+ETCO2 (0.91) ECG+TI+ETCO2 (0.92) 0 20 40 60 80 100 100-SP (%) 0 20 40 60 80 100 SE (%) (b) Excluding PEA segments from ROSC cases ETCO2 (0.79) ECG (0.93) ECG+TI (0.94) ECG+ETCO2 (0.95) ECG+TI+ETCO2 (0.96) Figure 3: ROC curves of the RF classifier for different feature sets. Panel (a) shows results for the whole dataset, while panel (b) shows the curves after excluding the PEAs from episodes with ROSC. The AUC value for each classifier is shown between parentheses. -1 0 1 ECG (mV) PEA/PEA 996 997 998 999 Time (s) -50 0 50 TI (m ) 0 50 100 CO2 (mmHg) PR/PR 1086 1087 1088 Time (s) (a) Correctly identified patient with ROSC PR/PR 1106 1107 1108 Time (s) -1 0 1 ECG (mV) PEA/PEA 705 706 707 708 Time (s) -50 0 50 TI (m ) 0 50 100 CO2 (mmHg) PEA/PEA 772 773 774 Time (s) (b) Correctly identified patient with no-ROSC PEA/PEA 792 793 794 Time (s) -1 0 1 ECG (mV) PEA/PEA 1763 1764 1765 1766 Time (s) -50 0 50 TI (m ) 0 50 100 CO2 (mmHg) PEA/PR 1915 1916 1917 1918 Time (s) (c) Incorrectly identified patient with no-ROSC PEA/PR 1935 1936 1937 1938 Time (s) Figure 4: Examples of the case study. Panels (a), (b) and (c) show a correctly identified patient with ROSC, a correctly identified patient without ROSC, and a patient without ROSC incorrectly identified, respectively. Each panel depicts the three consecutive PEA/PR segments analysed. The text on top of each segment indicates its true label followed by the predicted label by the classifier. The capnogram corresponds to the minute before the onset of the segment and the dashed horizontal line represents the MEtCO2. -10 -8 -6 -4 -2 ROSC Time (min) 0 0.2 0.4 0.6 0.8 1 ppr Figure 5: Time evolution of ppr for the PEA segments as the patients approach ROSC. Blue dots indicate values for each segment, and the red curve is fitted to the median values of ppr every 2 minutes. Table Legends416 Table 1 ROC curve analysis of the machine learning classifier when the whole417 PR/PEA dataset is considered and when the PEAs from ROSC cases418 were excluded. The SE and SP are given for the optimal point according419 to the Youden index.420 21