Secondary Logo

Journal Logo

Cardiac arrest

prediction models in the early phase of hospitalization

Dumas, Florencea,b; Bougouin, Wulfranb,c; Cariou, Alainb,d

Current Opinion in Critical Care: June 2019 - Volume 25 - Issue 3 - p 204–210
doi: 10.1097/MCC.0000000000000613
CARDIOPULMONARY RESUSCITATION: Edited by Hans Friberg
Free
SDC

Purpose of review There is a need for an early assessment of outcome in patients with return of spontaneous circulation after cardiac arrest. During the last decade, several models were developed in order to identify predictive factors that may facilitate prognostication and stratification of outcome.

Recent findings In addition to prognostication tools that are used in intensive care, at least five scores were recently developed using large datasets, based on simple and immediately available parameters, such as circumstances of arrest and early in-hospital indicators. Regarding neurological outcome, predictive performance of these models is good and even excellent for some of them. These scores perform very well for identifying patients at high-risk of unfavorable outcome. The most important limitation of these scores remains the lack of replication in different communities. In addition, these scores were not developed for individual decision- making, but they could instead be useful for the description and comparison of different cohorts, and also to design trials targeting specific categories of patients regarding outcome. Finally, the recent development of big data allows extension of research in epidemiology of cardiac arrest, including the identification of new prognostic factors and the improvement of prediction according to the profile of populations.

Summary In addition to the development of artificial intelligence, the prediction approach based on adequate scores will further increase the knowledge in prognostication after cardiac arrest. This strategy may help to develop treatment strategies according to the predicted severity of the outcome.

aEmergency Department, Cochin University Hospital (APHP) and Paris Descartes University

bSudden Death Expertise Center, INSERM U970 (Team 4), PARCC, Paris

cRamsay Générale de Santé, Hôpital Privé Jacques Cartier

dMedical Intensive Care Unit, Cochin University Hospital (APHP) and Paris Descartes University, Paris, France

Correspondence to Florence Dumas, Emergency Department, Cochin University Hospital (APHP) and Paris Descartes University, 27 rue du Faubourg Saint Jacques, 75014 Paris, France. E-mail: florence.dumas@aphp.fr

Back to Top | Article Outline

INTRODUCTION

Even in patients with successful return of spontaneous circulation (ROSC), outcome after cardiac arrest remains poor. The overall in-hospital survival rate widely varies both worldwide [1] and across communities [2,3], from one-fold to four-folds according to circumstances of arrest and postresuscitation interventions.

Morbidity and mortality events observed during the postresuscitation phase are mainly explained by the postcardiac arrest syndrome, occurring during first hours and days following cardiac arrest. This syndrome combines in different proportions a whole-body ischemia–reperfusion that provokes a multiple organ failure and an anoxic–ischemic cerebral injury [4]. Several studies have already shown that early interventions performed after ROSC, such as treatment of the cause, targeted temperature management, optimal hemodynamic management and extracorporeal life support in selected patients, could improve the outcome in postcardiac arrest patients [5,6]. However, the decision-making process regarding the allocation of these resources, in parallel with the management of patients’ proxies, remains a complex challenge for physicians facing these situations. Early prognostication of patients could be helpful for treatment allocation, and for delivering adequate information to relatives.

Consequently, several prediction models and scores have been developed in order to stratify the risk of unfavorable outcome and to discriminate the best candidates for postresuscitation interventions. Most of these prediction models were developed and validated using observational data originating from registries or consortium databases, based on early classical prognostic factors. More recently, efforts were also developed to identify parameters using novel methods, such as big data management and artificial intelligence.

In the present review, we aimed to describe prediction models that are currently available and to develop new approaches that could improve early prognostication of postcardiac arrest patients in the coming years.

Box 1

Box 1

Back to Top | Article Outline

PREDICTION MODELS AND SCORES

Prognostic factors

After successful resuscitation, several factors have previously and consistently been identified as predictors of unfavorable outcome [7]. These parameters include both prehospital circumstances and early in-hospital parameters. For instance, initial shockable rhythm, bystander cardiopulmonary resuscitation, and age are all strong prognostic factors in predicting short-term (hospital discharge) and even long-term outcomes [8–10]. Some studies investigated other factors, such as duration of resuscitation, which is dramatically related with outcome [11]. In addition, parameters that are closely associated with a higher risk of a severe postcardiac syndrome are also strong predictors of unfavorable neurological outcome at discharge. Among biological markers, blood lactate concentrations are indeed associated with increased cardiac arrest mortality [12–14]. However, the first two links of the chain of survival (early recognition, alert and chest compressions) are undoubtedly the strongest determinants of later outcome (Table 1).

Table 1

Table 1

In order to combine indicators of outcome and to refine early assessment, several scores were developed based on prediction models of risk. Existing scores have been validated and are commonly used in the general population of critically ill patients. However, these nonspecific scores cannot be employed in cardiac arrest patients as they do not consider specific prognostic factors. Consequently, different scores were recently developed using cardiac arrest registries with internal, and for some of them, external, validation.

Back to Top | Article Outline

Unspecific intensive care scores

The APACHE II (Acute Physiology and Chronic Health Evaluation) score, described in 1981 by Knaus et al.[15] and the SAPS II (Simplified Acute Physiologic Score) by Le Gall et al.[16] are two severity illness classifications that are commonly used in ICU. They include both clinical and biological parameters, using the worst values within the first 24 h in the ICU for each one of them. However, even if widely used in many illness conditions, they appear inappropriate in cardiac arrest because of their complexity (requirement of dozens of criteria) and the 24 h wait for assessment. This time delay in evaluation is crucial and quite inadequate in postcardiac arrest, as many patients will experience a postcardiac arrest syndrome in the very first hours, influencing mortality drastically. Furthermore, usual ‘Utstein criteria’ are not part of the scores.

Back to Top | Article Outline

Early out-of hospital cardiac arrest prediction scores

To refine prediction in this specific setting, new tools were generated for early prediction of outcome, which relies on variables readily available at hospital admission.

The out-of hospital cardiac arrest (OHCA) score was the first model developed in a cohort of 130 consecutive unselected OHCA patients, and validated in a prospective sample set of 210 patients. Using seven simple and early variables, this model accurately predicted prognosis in four different French ICUs during two different periods [17]. More recently, Maupain et al.[18] created the cardiac arrest hospital prognosis (CAHP) score (including age, location, witnessed, resuscitation delays, initial rhythm, arterial pH and epinephrine dosage), based on a larger dataset, validated in a prospective period and in a retrospective cohort with good discrimination performance stratifying the risk of unfavorable outcome into three groups (<150; 150–200 and >200). For the last category, the proportion of unfavorable outcome was over 96%. For note, OHCA and CAHP scores were both tested in external cohorts in the same community with good discrimination.

On the basis of data from the targerted temperature management (TTM) trial [19], Martinell et al. used 10 variables at admission including prehospital and in-hospital parameters (age, location, initial rhythm, resuscitation delays, epinephrine dosage, eye reflex, blood gas and motor Glasgow) to predict neurological outcome. The main strength of this model relies on the quality of the initial database, originating from a clinical trial of high quality, even if no external validation has yet been performed [20].

Datasets from four different hospitals in Japan constituted development and validation cohorts to create another scoring system: the cardiac arrest syndrome for induced hypothermia (CAST) score. This score originally included specific in-hospital variables, such as biological results or gray-to-white matter attenuation ratio (GWR) on early brain computed tomography (CT) scan. Even if validated in small datasets, the CAST score had excellent performance with 93% of patients being correctly classified [21,22].

Focusing on a selected population, Bascom et al.[23▪] developed a bedside prediction tool in order to determine the risk of circulatory-etiology death. In a multicenter registry (INTCAR: International Cardiac Arrest Registry), 956 survivors from cardiac arrest without ST elevation on the postresuscitation ECG constituted the final cohort that was divided into a derivation and a validation cohort. Authors retained five variables available at admission that were included in the CREST score, the final model yielded a satisfying discrimination. The CREST model stratified patients immediately after resuscitation according to risk of a circulatory-etiology death. The tool may allow for estimation of circulatory risk and improves the triage of survivors of cardiac arrest without ST-segment-elevation myocardial infarction.

Back to Top | Article Outline

What model should be preferred?

Even if a rigorous methodology was used in order to generate models providing a substantial or high prediction performance, there is a wide heterogeneity between these scores. Several reasons may explain this heterogeneity. First, several in-hospital factors differed across validation studies, such as results of CT-scans or arterial pH at admission (Table 2). In addition, all the scoring system used retrospective analyses of datasets, from registries or a clinical trial, and usually studied unselected populations (except for the CREST score that focused on no ST segment elevation patients and for the CAST score that was validated in hypothermia-treated patients). Furthermore, the CAHP, the CREST and TTM scores were developed and validated in large datasets. Importantly, endpoints varied from hospital discharge to 6 months neurological recovery. Finally, external validation is often lacking and very few direct comparisons have been performed between these scores. As an example, the OHCA score, which was the first to be developed, is also the only one that was evaluated in external datasets providing validation of calibration and high discrimination in two American distinct populations [24]. Finally, Martinell et al. compared their score with OHCA and CAHP score, and they reported a better performance using their model (Table 3).

Table 2

Table 2

Table 3

Table 3

Prediction models and scoring systems have expanded in recent years to provide support for early prediction and risk stratification. On the whole, these tools share a good performance even if they vary widely regarding their respective components. However, they were validated in different populations, resulting in difficulties to obtain external validation and reproducibility. Despite the heterogeneity, these scores represent a real improvement in assessing severity of disease in cardiac arrest patients, strategy and in the epidemiology to compare cohorts of cardiac arrest patients. Furthermore, scores may be of help when designing clinical trials targeting a specific population.

Back to Top | Article Outline

INTERESTS OF BIG DATA IN CARDIAC ARREST RESEARCH

Prediction tools previously described mostly relied on ‘classical prognostic factors’ commonly used in the field of cardiac arrest, such as Utstein characteristics [7]. Indeed, most cardiac arrest prediction scores were developed using cardiac arrest registries (OHCA score [17], CAHP score [18], CREST score [23▪] or CAST score [22]), or trials (TTM [19]), describing and including these Utstein parameters. However, in a large cohort from the ROC consortium, Rea et al.[25] previously reported that Utstein elements only account for a modest portion of cardiac arrest survival variability. Therefore, identification of new prognostic factors in the field of cardiac arrest is mandatory.

In the last decade, large datasets have been widely available for clinical research, thanks to the increasing use of digital health records systems. It was recently reported the healthcare data volume is increasing at a rate of almost 50% annually [26] and is expected to reach over 2000 exabytes (1018 bytes) in 2020. This development, combined with increased computational power, offers an interesting perspective.

As an example, the Medical Information Mart for Intensive Care (MIMIC-III) database includes over 54 000 hospital admissions in critical care [27], with vital signs, laboratory tests, reports, medications, ICD (International Classification of Diseases) codes. Similarly, the eICU Collaborative Research Database includes patients admitted to critical care units with vital signs, laboratory tests and medications [28]. Similar databases, applied to cardiac arrest patients, might offer a perspective to describe more accurately subgroups of patients with their early prognostic factors. Indeed, as recently described in a review by Narayan et al.[29▪▪] establishing a better definition of sudden cardiac arrest, and identifying subgroups of phenotypes (instead of considering cardiac arrest as a monolithic event) is crucial to better address its pathology. Especially, regarding early prognostication at admission, a definition of precise phenotypes could enhance previously established models, and addition of initial vital signs and/or laboratory results could dramatically increase the predictive power of existing models. Tailoring specific treatments in the post-ROSC period (such as coronary angiogram or targeted temperature management), according to predicted prognosis (as previously described for coronary angiogram according to CAHP score [30]) could help to optimize efficiency of care in this setting, and to avoid unnecessary treatments.

In addition to these promising elements, management of massive datasets requires a specific, careful data management. A major pitfall can result from unlabelled and/or uncorrected data, with uncertain quality. Considering the usually automated gathering of data in dataset (for example, regarding Electronic Health Records or vital parameters), medical accuracy of data must be assessed. Moreover, several data cannot be used because they are unstructured (for example, free-text reports). Use of data from electronic medical records, thus requires prior assessment of their reliability.

Back to Top | Article Outline

ARTIFICIAL INTELLIGENCE IN CARDIOVASCULAR RESEARCH, AND IN THE FIELD OF CARDIAC ARREST

In order to provide more personalized care, and thanks to big datasets previously described, a new approach, such as machine learning could be used [31] These techniques have been previously described in other cardiologic diseases (such as electrocardiogram analysis [32,33] or echocardiogram [34]) and in the critical care setting (for example, to identify optimal treatment strategies for sepsis in the ICU [35▪]). To the best of our knowledge, the use of machine learning has not been reported in the aim of establishing new prediction models in cardiac arrest research.

Different approaches could be used. Machine learning can be either supervised or unsupervised [36▪]. In a supervised approach, the algorithm aims to accurately establish and predict a specific characteristic, such as outcome. Supervised machine learning could be used to predict early outcome (i.e. death or unfavourable neurological outcome at hospital discharge) after cardiac arrest, using parameters available during the first hours (Utstein parameters, but also clinical and biological data). By contrast, unsupervised methods can have exploratory goals, to identify hidden clusters or relations between variables. Unsupervised learning could be useful to identify phenotypes of patients with a certain risk profile after hospital admission.

Overall, enrichment of datasets offers great potential for new approaches in the field of cardiac arrest research, including machine learning (either with or without preconditions, thanks to unsupervised approach). Potential applications are promising, and development of large datasets should be promoted to allow research in cardiac arrest using an approach previously validated in other fields of cardiovascular research. However, whether development of artificial intelligence will make ‘doctors obsolete’ remains highly uncertain [37].

Back to Top | Article Outline

CONCLUSION

Early prediction of outcome after cardiac arrest is widely considered a critical issue for many reasons. Classifying patients and assessing the degree of severity constitute an important key for clinicians facing treatment decisions, families’ questions and also for clinical research. To address this question, many models were recently developed using large datasets and registries. Most of these scoring systems were found to have good performance for the prediction of outcome, but replication studies and external validation are lacking.

Back to Top | Article Outline

Acknowledgements

None.

Back to Top | Article Outline

Financial support and sponsorship

None.

Back to Top | Article Outline

Conflicts of interest

There are no conflicts of interest.

Back to Top | Article Outline

REFERENCES AND RECOMMENDED READING

Papers of particular interest, published within the annual period of review, have been highlighted as:

  • ▪ of special interest
  • ▪▪ of outstanding interest
Back to Top | Article Outline

REFERENCES

1. Nichol G, Thomas E, Callaway CW, et al. Resuscitation Outcomes Consortium Investigators. Regional variation in out-of-hospital cardiac arrest incidence and outcome. JAMA 2008; 300:1423–1431.
2. Chocron R, Bougouin W, Beganton F, et al. Are characteristics of hospitals associated with outcome after cardiac arrest? Insights from the Great Paris registry. Resuscitation 2017; 118:63–69.
3. van Diepen S, Girotra S, Abella BS, et al. Multistate 5-year initiative to improve care for out-of-hospital cardiac arrest: primary results from the HeartRescue Project. J Am Heart Assoc 2017; 6: pii: e005716.
4. Neumar RW, Nolan JP, Adrie C, et al. Postcardiac arrest syndrome: epidemiology, pathophysiology, treatment, and prognostication. A consensus statement from the International Liaison Committee on Resuscitation (American Heart Association, Australian and New Zealand Council on Resuscitation, European Resuscitation Council, Heart and Stroke Foundation of Canada, InterAmerican Heart Foundation, Resuscitation Council of Asia, and the Resuscitation Council of Southern Africa); the American Heart Association Emergency Cardiovascular Care Committee; the Council on Cardiovascular Surgery and Anesthesia; the Council on Cardiopulmonary, Perioperative, and Critical Care; the Council on Clinical Cardiology; and the Stroke Council. Circulation 2008; 118:2452–2483.
5. Chelly J, Mongardon N, Dumas F, et al. Benefit of an early and systematic imaging procedure after cardiac arrest: insights from the PROCAT (Parisian Region Out of Hospital Cardiac Arrest) registry. Resuscitation 2012; 83:1444–1450.
6. Nolan JP, Soar J, Cariou A, et al. European Resuscitation Council and European Society of Intensive Care Medicine 2015 guidelines for postresuscitation care. Intensive Care Med 2015; 41:2039–2056.
7. Perkins GD, Jacobs IG, Nadkarni VM, et al. Cardiac arrest and cardiopulmonary resuscitation outcome reports: update of the Utstein Resuscitation Registry Templates for Out-of-Hospital Cardiac Arrest: a statement for healthcare professionals from a task force of the International Liaison Committee on Resuscitation (American Heart Association, European Resuscitation Council, Australian and New Zealand Council on Resuscitation, Heart and Stroke Foundation of Canada, InterAmerican Heart Foundation, Resuscitation Council of Southern Africa, Resuscitation Council of Asia); and the American Heart Association Emergency Cardiovascular Care Committee and the Council on Cardiopulmonary, Critical Care, Perioperative and Resuscitation. Circulation 2015; 132:1286–1300.
8. Dumas F, Rea TD, Fahrenbruch C, et al. Chest compression alone cardiopulmonary resuscitation is associated with better long-term survival compared with standard cardiopulmonary resuscitation. Circulation 2013; 127:435–441.
9. Geri G, Dumas F, Bougouin W, et al. Immediate percutaneous coronary intervention is associated with improved short- and long-term survival after out-of-hospital cardiac arrest. Circ Cardiovasc Interv 2015; 8: pii: e002303.
10. Dumas F, Rea TD. Long-term prognosis following resuscitation from out-of-hospital cardiac arrest: role of aetiology and presenting arrest rhythm. Resuscitation 2012; 83:1001–1005.
11. Reynolds JC, Frisch A, Rittenberger JC, Callaway CW. Duration of resuscitation efforts and functional outcome after out-of-hospital cardiac arrest: when should we change to novel therapies? Circulation 2013; 128:2488–2494.
12. Dell’Anna AM, Sandroni C, Lamanna I, et al. Prognostic implications of blood lactate concentrations after cardiac arrest: a retrospective study. Ann Intensive Care 2017; 7:101.
13. Donnino MW, Andersen LW, Giberson T, et al. Initial lactate and lactate change in postcardiac arrest: a multicenter validation study. Crit Care Med 2014; 42:1804–1811.
14. Hayashida K, Suzuki M, Yonemoto N, et al. SOS-KANTO 2012 Study Group. Early lactate clearance is associated with improved outcomes in patients with postcardiac arrest syndrome: a prospective, multicenter observational study (SOS-KANTO 2012 Study). Crit Care Med 2017; 45:e559–e566.
15. Knaus WA, Draper EA, Wagner DP, Zimmerman JE. APACHE II: a severity of disease classification system. Crit Care Med 1985; 13:818–829.
16. Le Gall JR, Loirat P, Alperovitch A. Simplified acute physiological score for intensive care patients. Lancet 1983; 2:741.
17. Adrie C, Cariou A, Mourvillier B, et al. Predicting survival with good neurological recovery at hospital admission after successful resuscitation of out-of-hospital cardiac arrest: the OHCA score. Eur Heart J 2006; 27:2840–2845.
18. Maupain C, Bougouin W, Lamhaut L, et al. The CAHP (Cardiac Arrest Hospital Prognosis) score: a tool for risk stratification after out-of-hospital cardiac arrest. Eur Heart J 2016; 37:3222–3228.
19. Nielsen N, Wetterslev J, Cronberg T, et al. Targeted temperature management at 33°C versus 36°C after cardiac arrest. N Engl J Med 2013; 369:2197–2206.
20. Martinell L, Nielsen N, Herlitz J, et al. Early predictors of poor outcome after out-of-hospital cardiac arrest. Crit Care 2017; 21:96.
21. Nishikimi M, Matsuda N, Matsui K, et al. CAST: a new score for early prediction of neurological outcomes after cardiac arrest before therapeutic hypothermia with high accuracy. Intensive Care Med 2016; 42:2106–2107.
22. Nishikimi M, Matsuda N, Matsui K, et al. A novel scoring system for predicting the neurologic prognosis prior to the initiation of induced hypothermia in cases of postcardiac arrest syndrome: the CAST score. Scand J Trauma Resusc Emerg Med 2017; 25:49.
23▪. Bascom KE, Dziodzio J, Vasaiwala S, et al. Derivation and validation of the CREST model for very early prediction of circulatory etiology death in patients without ST-segment-elevation myocardial infarction after cardiac arrest. Circulation 2018; 137:273–282.

Prediction model including patients from 44 international centres.

24. Hunziker S, Bivens MJ, Cocchi MN, et al. International validation of the out-of-hospital cardiac arrest score in the United States. Crit Care Med 2011; 39:1670–1674.
25. Rea TD, Cook AJ, Stiell IG, et al. Resuscitation Outcomes Consortium Investigators. Predicting survival after out-of-hospital cardiac arrest: role of the Utstein data elements. Ann Emerg Med 2010; 55:249–257.
26. Stanford Medicine 2017 Health Trends Report - Harnessing the Power of Data in Health.
27. Johnson AEW, Pollard TJ, Shen L, et al. MIMIC-III, a freely accessible critical care database. Sci Data 2016; 3:160035.
28. eICU. cited 11 February 2019. Available at: https://eicu-crd.mit.edu/about/eicu/. [Accessed 11 February 2019].
29▪▪. Narayan SM, Wang PJ, Daubert JP. New concepts in sudden cardiac arrest to address an intractable epidemic: JACC state-of-the-art review. J Am Coll Cardiol 2019; 73:70–88.

Very interesting review with a state-of-the-art in the field of sudden death, identifying future areas of research.

30. Bougouin W, Dumas F, Karam N, et al. Sudden Death Expertise Center. Should we perform an immediate coronary angiogram in all patients after cardiac arrest?: Insights from a large French Registry. JACC Cardiovasc Interv 2018; 11:249–256.
31. Beam AL, Kohane IS. Big data and machine learning in healthcare. JAMA 2018; 319:1317–1318.
32. Hannun AY, Rajpurkar P, Haghpanahi M, et al. Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nat Med 2019; 25:65–69.
33. Attia ZI, Kapa S, Lopez-Jimenez F, et al. Screening for cardiac contractile dysfunction using an artificial intelligence–enabled electrocardiogram. Nat Med 2019; 25:70–74.
34. Zhang J, Gajjala S, Agrawal P, et al. Fully automated echocardiogram interpretation in clinical practice. Circulation 2018; 138:1623–1635.
35▪. Komorowski M, Celi LA, Badawi O, et al. The artificial intelligence clinician learns optimal treatment strategies for sepsis in intensive care. Nat Med 2018; 24:1716–1720.

Nice example of the potential use of artificial intelligence in the treatment of sepsis, based on a large US database.

36▪. Johnson KW, Torres Soto J, Glicksberg BS, et al. Artificial intelligence in cardiology. J Am Coll Cardiol 2018; 71:2668–2679.

In this review, application of artificial intelligence is described in general, and especially its application in cardiology.

37. Goldhahn J, Rampton V, Spinas GA. Could artificial intelligence make doctors obsolete? BMJ 2018; 363:k4563.
Keywords:

cardiac arrest; outcome; prediction; scores

Copyright © 2019 YEAR Wolters Kluwer Health, Inc. All rights reserved.