If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Despite improvements in short-term survival for Out-of-Hospital Cardiac Arrest (OHCA) in the past two decades, long-term survival is still not well studied. Furthermore, the contribution of different variables on long-term survival have not been fully investigated.
Aim
Examine the 1-year prognosis of patients discharged from hospital after an OHCA. Furthermore, identify factors predicting re-arrest and/or death during 1-year follow-up.
Methods
All patients 18 years or older surviving an OHCA and discharged from the hospital were identified from the Swedish Register for Cardiopulmonary Resuscitation (SRCR). Data on diagnoses, medications and socioeconomic factors was gathered from other Swedish registers. A machine learning model was constructed with 886 variables and evaluated for its predictive capabilities. Variable importance was gathered from the model and new models with the most important variables were created.
Results
Out of the 5098 patients included, 902 (∼18%) suffered a recurrent cardiac arrest or death within a year. For the outcome death or re-arrest within 1 year from discharge the model achieved an ROC (receiver operating characteristics) AUC (area under the curve) of 0.73. A model with the 15 most important variables achieved an AUC of 0.69.
Conclusions
Survivors of an OHCA have a high risk of suffering a re-arrest or death within 1 year from hospital discharge. A machine learning model with 15 different variables, among which age, socioeconomic factors and neurofunctional status at hospital discharge, achieved almost the same predictive capabilities with reasonable precision as the full model with 886 variables.
Rawshani A, Herlitz J, Lindqvist J, Aune S, Strömsöe A. Svenska Hjärt- och Lungregistret - Årsrapport 2020 [Swedish Register for Cardioplumonary Resuscitation - Year Report 2020], 2021.
Survival rates were around 4–5% in the 1990s, followed by a rapid increase in survival during the early 2000s. No improvements have been observed since year 2010.
Rawshani A, Herlitz J, Lindqvist J, Aune S, Strömsöe A. Svenska Hjärt- och Lungregistret - Årsrapport 2020 [Swedish Register for Cardioplumonary Resuscitation - Year Report 2020], 2021.
Considering the dismal prognosis in OHCA much of the scientific work has been focused on short-term outcomes and less work has been done on long-term survival.
With regards to survivors of OHCA, previous studies have indicated that the risk of recurrent cardiac arrest and death peaks during the first year following the event, with mortality rates five years post-arrest approaching that of the general population.
Known predictors of long-term survival are percutaneous coronary intervention (PCI), cardiac aetiology as well as exhibiting a shockable initial rhythm.
Long-term prognosis following resuscitation from out of hospital cardiac arrest: role of percutaneous coronary intervention and therapeutic hypothermia.
Despite the advances in the management of OHCA more work is needed to investigate survival in the long-term perspective. In this study, we investigated if recurrent cardiac arrest or death could be predicted in patients surviving an OHCA to be discharged from hospital. The clinical utility of such a tool is obvious as it could enable risk stratification of survivors. We constructed and evaluated a large number of machine learning models with the aim to create a clinical prediction tool.
Methods
Ethics
The study was approved by the Swedish Ethical Review Authority (registration number 2021-06459-02).
Study population
All patients with an OHCA between January 1, 2010, and July 5, 2021, were identified from the Swedish Register for Cardiopulmonary Resuscitation (SRCR). Only patients 18 years or older who were discharged from hospital alive following were eligible.
Data sources
The SRCR has previous been described and validated.
Rawshani A, Herlitz J, Lindqvist J, Aune S, Strömsöe A. Svenska Hjärt- och Lungregistret - Årsrapport 2020 [Swedish Register for Cardioplumonary Resuscitation - Year Report 2020], 2021.
Cardiac arrest in COVID-19: characteristics and outcomes of in- and out-of-hospital cardiac arrest. A report from the Swedish Registry for Cardiopulmonary Resuscitation.
Discriminatory cardiac arrest care? Patients with low socioeconomic status receive delayed cardiopulmonary resuscitation and are less likely to survive an in-hospital cardiac arrest.
All Swedish EMS units report to the register. For OHCA, all cases where the emergency medical services (EMS) crews initiate or continue cardiopulmonary resuscitation are reported to the register. The SRCR contains data on the critical time intervals from the cardiac arrest (e.g., time from cardiac arrest to ambulance arrival), to the subsequent steps in the chain of survival
as well as initial rhythm, presumed cause and place of the arrest as well as some pre-hospital and hospital interventions.
For the patients identified in the SRCR, data was linked to the Swedish National Patient Register for data on comorbidities reported by specialist care units.
Information on education level, country of birth and income levels was gathered from Statistics Sweden’s Longitudinal integrated database for health insurance and labour market studies (LISA). These variables are a proxy of the concept of socioeconomic status, the social standing or class of individuals or groups.
Data on prescribed medications was gathered from the Swedish National Prescribed Drug Register. From the mentioned data sources, a total of 886 variables were gathered and used for analysis.
Outcomes
Outcomes include death or re-arrest occurring within one year from the first cardiac arrest. Recurrent arrest (re-arrest) includes both OHCA and IHCA occurring after the index event (first OHCA).
For OHCA, SRCR only includes OHCA cases where resuscitation has been attempted and OHCA being a disorder often leading to death for the affected patients, we deemed it suitable to combine death and re-arrest as an outcome variable.
Statistical analyses
Extreme gradient boosting (XGBoost), a tree-based ensemble method with boosting,
was used to predict the binary outcome of death or re-arrest within one year from the initial OHCA. Model performance was evaluated using AUC (area under the curve) of ROC (receiver operating characteristics) curves.
To assess the relative importance of each variable, we computed variable importance using the default method for XGBoost (model gain). In short, gain is the contribution of a single variable to the model, relative to each of the other variables.
By assessing variable importance, parsimonious classification models with the top 5, 10 or 15 variables (i.e. those with the greatest importance) were constructed to evaluate whether more parsimonious models were sufficient, as compared to the full model with hundreds of variables.
Kaplan-Meier estimators were calculated to describe survival.
All statistical models were computed using R (R Foundation for Statistical Computing, version 4.1.2). The XGBoost models were constructed within the tidymodels framework in R.
Kuhn et al. Tidymodels: a collection of packages for modeling and machine learning using tidyverse principles. 2020 [cited June 29, 2022]. Available from: https://www.tidymodels.org.
Table 1 shows baseline characteristics of the study population overall and stratified by outcome. For the patient outcome, an event refers to the patient suffering a recurrent arrest or death within 1 year from their initial OHCA. For the 5098 patients included, the mean age was 63.4 years and 75.3% were male. The index OHCA had a presumed cardiac cause in 73.3% of all cases. The group with an event were older at baseline (67.9 vs 62.5). The disposable income on both family and individual level was lower in the event group. Patients suffering an event less often had a presumed cardiac cause for their OHCA (68.8% vs 74.3%), they also received an ICD (implantable cardioverter defibrillator) to a lesser extent (13.2% vs 26.9%). The event group underwent PCI to a lesser degree (34.7% vs 50.3%). Some of the variables examined at baseline have been chosen from the variables with the highest importance, see Fig. 1.
Table 1Baseline characteristics of the entire study population and stratified by whether an event had occurred within 365 days.
Overall
Death or re-arrest
No event
p
n
5098
902
4196
Male (%)
3839 (75.3)
658 (72.9)
3181 (75.8)
0.078
Age in years (mean (SD))
63.42 (15.98)
67.86 (16.67)
62.46 (15.67)
<0.001
Cardiac cause for the cardiac arrest (%)
3738 (73.3)
621 (68.8)
3117 (74.3)
0.001
Disposable income, family level, in hundreds SEK (mean (SD))
8690.43 (3726.78)
8022.58 (3731.26)
8834.00 (3710.60)
<0.001
Disposable income, individual level, in hundreds SEK (mean (SD))
4793.46 (2654.59)
4502.71 (2778.97)
4855.97 (2623.22)
<0.001
Education status of bystander (%)
0.022
Laymen, not CPR educated
2042 (40.1)
348 (38.6)
1694 (40.4)
Laymen, CPR educated
2066 (40.5)
349 (38.7)
1717 (40.9)
Health care professional
990 (19.4)
205 (22.7)
785 (18.7)
Year for cardiac arrest (median [IQR])
2015.00 [2012.00, 2017.00]
2015.00 [2013.00, 2018.00]
2015.00 [2012.00, 2017.00]
<0.001
Time from cardiac arrest to start of CPR, minutes (median [IQR])
1.00 [0.00, 3.00]
1.00 [0.00, 3.75]
1.00 [0.00, 3.00]
0.099
Time from cardiac arrest to alarm out, minutes (median [IQR])
2.00 [1.00, 3.00]
2.00 [1.00, 4.00]
2.00 [1.00, 3.00]
0.037
Time from cardiac arrest to ambulance arrival, minutes (median [IQR])
9.00 [6.00, 13.00]
9.00 [6.00, 14.00]
9.00 [6.00, 13.00]
0.248
Time from alarm out to ambulance arrival, minutes (median [IQR])
8.00 [5.00, 13.00]
8.00 [6.00, 13.00]
8.00 [5.00, 13.00]
0.041
First registered cardiac rhythm (%)
<0.001
VF/pVT
3952 (77.5)
554 (61.4)
3398 (81.0)
PEA
455 (8.9)
138 (15.3)
317 (7.6)
Asystole
691 (13.6)
210 (23.3)
481 (11.5)
Adrenaline given pre-hospital (%)
1834 (36.0)
386 (42.8)
1448 (34.5)
<0.001
Coronary angiography being performed during hospital care (%)
2156 (42.3)
408 (45.2)
1748 (41.7)
0.053
Percutaneous coronary intervention being performed during hospital care (%)
2425 (47.6)
313 (34.7)
2112 (50.3)
<0.001
ICD implanted during hospital stay (%)
1246 (24.4)
119 (13.2)
1127 (26.9)
<0.001
Cerebral Performance Category at hospital discharge (%)
Fig. 1Variable importance (gain) extracted from the full XGBoost model. a) Importance for the individual variables. Top 15 variables displayed. Further explanations can be found in Supplementary Table 1. b) Importance for classes of variables. Further explanations can be found in Supplementary Table 2. CPC: Cerebral Performance Category, CPR: Cardiopulmonary resuscitation,
Following hyperparameter tuning, the classification model with all 886 variables achieved an AUC of 0.78 on the training data and an AUC of 0.73 on test data (data not shown). From this model variable importance was computed. Fig. 1a and 1b presents the relative importance of all 886 variables in the classification model.
Among the fifteen most important variables, age had the highest importance followed by Cerebral Performance Category (CPC) score of 4 and variable disposable income on an individual level (Fig. 1a).
Grouping all variables by different classes, comorbidities (diagnoses at baseline according to International Statistical Classification of Diseases and Related Health Problems 10, ICD 10) had the highest relative importance followed by spatiotemporal factors (e.g., time of cardiac arrest, local for cardiac arrest and which hospital managed the cardiac arrest) (Fig. 1b). See Supplementary Tables 1 and 2 for a further description of the top 15 variables and classes of variables respectively.
From the 15 variables with the highest importance (see Fig. 1a), three models were created and evaluated with the top 5, 10 or 15 variables respectively. These three models achieved a ROC AUC of 0.69, 0.64 and 0.62 for the top 15, top 10 and top 5 variables, respectively when evaluated on test data. For comparison the model built with all 886 variables achieved a ROC AUC of 0.73 (Fig. 2).
Fig. 2Comparison of ROC AUC value predicting the test data for models based on all 886, the top 15, top 10 and top 5 variables, respectively. Longitudinal dashed line corresponding to an AUC of 0.5.
Regarding re-arrest or death for the study population, a Kaplan-Meier curve shows that most of the events have occurred close to the date for the index OHCA (Fig. 3a). The cumulative number of events within one year from the index OHCA was 902, corresponding to roughly 18% of the study population. Kaplan-Meier curves stratified by sex, age and causes indicates a higher proportion of events in females, higher ages and for non-cardiac causes (Fig. 3b-d).
Fig. 3Kaplan Meier plots illustrating the outcome re-arrest or death in the study population for 365 days. Survival probability in regard to re-arrest or death is shown on the Y-axis. Time in days is shown on the X-axis. a) Plot showing re-arrest or death for entire study population. People at risk is shown in the upper table. Cumulative number of events in the lower table. b) Survival stratified by sex. Table showing cumulative number of events. c) Survival stratified by age. Table showing cumulative number of events. d) Survival stratified by causes. Table showing cumulative number of events.
Cardiac arrest is a highly stochastic event, making it very difficult to make long-term predictions of the risk of cardiac arrest. Research efforts have failed to create reliable long-term prediction models. Thus, it has recently been suggested that risk prediction should move from long-term models to near-term prediction.
The fundamental explanation for this observation is the fact that a cardiac arrest is a stochastic event, typically requiring the coincidence of a multitude of pro-arrest factors. However, this study demonstrated that long-term prediction of recurrent arrest or death is possible in the subgroup of patients that survive an initial OHCA. We show that these events can be predicted with a ROC-AUC of 0.73, which is arguably adequate for clinical use.
This prediction model should be considered in all patients discharged after OHCA.
The advantage of using machine learning algorithms is their ability to handle many variables, as well as capture non-linear relationships in data. We examined 886 candidate predictors of death or recurrent cardiac arrest in patients surviving an OHCA to discharge. The full model achieved an AUC of 0.73 on test data, corresponding to a 73% chance of correctly classifying the outcome in each case. For the parsimonious models constructed using the top 5, 10 and 15 variables, the AUC was reduced to 0.62, 0.64 and 0.69, respectively. This suggest that a reliable prediction requires a large number of predictors, although reasonable predictions can be made using models with relatively few variables. Hence, it should be considered a feasible task to collect these variables and perform the risk calculation.
For variable importance (Fig. 1a), the most important predictor was age. This is line with the fact that all-cause mortality and cardiac arrest both correlate with increasing age.
Rawshani A, Herlitz J, Lindqvist J, Aune S, Strömsöe A. Svenska Hjärt- och Lungregistret - Årsrapport 2020 [Swedish Register for Cardioplumonary Resuscitation - Year Report 2020], 2021.
A previously published paper also found that age of 65 years or older was associated with a higher probability of recurrent OHCA compared to lower ages.
Furthermore, a CPC of 4 (severe neurological sequelae) was the second most important predictor. CPC is a scoring system which is commonly used for measuring neurological outcome following a cardiac, ranging from no or minimal disability (CPC-1) to brain death (CPC-5).
CPC or other measures of neurological outcome are important to report considering the importance of function for patients surviving an OHCA, as noted by the latest Utstein guidelines for OHCA.
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.
As can be seen in Fig. 1a, all CPC scores except CPC-5 (i.e., brain death) appear among the 15 most important predictors. The reason for multiple levels of the same categorical variable appearing is due to the XGBoost algorithm only accepts numerical data.
Accordingly, one may dichotomize categorical variables to 0 or 1 for each level representing yes or no, which is called one-hot encoding. This means that instead of one variable for CPC we end up with 4 variables which may assume the value of either 0 or 1. This highlights the importance of assessing CPC.
Following age and CPC, we note that disposable family and individual income were the third and fourth most important predictors.
As evident in the baseline table (Table 1), the disposable income was significantly lower in patients who experienced an event. Socioeconomic status is a complex concept reflecting the social standing of an individual or group, and it is consistently associated with survival across the spectrum of diseases.
The association between socioeconomic status and survival after an OHCA has been previously studied; studies have found a correlation between higher socioeconomic status and higher survival rates.
The mechanisms of how low SES is associated with a poorer outcome in OHCA are not fully elucidated. Previous studies have suggested that disparity in survival may be due to lower rates of CPR training resulting in a lower tendency to perform CPR and less knowledge about how to alert the EMS.
Socio-economic differences in incidence, bystander cardiopulmonary resuscitation and survival from out-of-hospital cardiac arrest: A systematic review.
Lower income has also been associated with a lower rate of early coronary angiography in OHCA, mainly explained by a lower rate of shockable rhythms as presenting rhythm. This may indicate a different spectrum of causes for the OHCA for different income levels or perhaps a longer delay to CPR resulting in shockable rhythms transforming to non-shockable rhythms.
Corroborating these findings, socioeconomic inequities exist in pre-hospital care where patients with a lower SES have a longer delay in stroke care, which in turn may contribute to the socioeconomic disparities seen in stroke outcomes.
Altogether, this indicates that when accounting for comorbidities, medications, peri-arrest factors and demographic factors, higher socioeconomic status is associated with a higher 1-year survival rate. Furthermore, socioeconomic factors are among the most important factors for predicting re-arrest or death in the first year following an OHCA, implying their importance in survival following an OHCA.
Initial presenting rhythm was also among the most important variables. The patients with an event within one year had a higher degree of PEA (pulseless electrical activity) or asystole, i.e. non-shockable rhythms (Table 1). In line with this, a previous study found a lower five-year survival rate in patients presenting with a non-shockable rhythm.
Interestingly, looking at variable importance for classes of variables, comorbidities have the highest importance despite no single diagnosis being among the most important variables. Whether it is a combination of diagnoses or the number of diagnoses which are important cannot be concluded from this study but would be interesting to look further into in this population.
Long-term prognosis following resuscitation from out of hospital cardiac arrest: role of percutaneous coronary intervention and therapeutic hypothermia.
the population with no event within a year had a cardiac cause for their OHCA more frequently and PCI was more frequently performed despite coronary angiography being performed to a similar degree; see Table 1. This supports PCI and cardiac cause being associated with a more favourable outcome. It is probable that PCI is being performed to a higher degree due to there being a cardiac cause more frequently in the group without a recurrent cardiac arrest or death within 1 year.
In the study population, roughly 18% (902outof5096) suffered a recurrent cardiac arrest or died within one year from the index OHCA. Previous studies on long-term survival in patients discharged from hospital after an OHCA have found a 1-year survival rate of 92% (95% CI: 91–93%) and 1-year OHCA recurrence rate of 2.4% (95% CI: 2.0–3.0%).
In comparison to these studies, a higher event rate was observed, which in part may be due to also including in-hospital cardiac arrest as an outcome.
Limitations
In the SRCR, OHCAs and IHCAs are only reported in patients where resuscitation is attempted; sudden cardiac deaths without resuscitation attempts are not enrolled in the registry. We capture this in the outcome model by using a composite of recurrent cardiac arrest or death. Regarding cause of cardiac arrest, this variable is reported by EMS personnel with limited data, making this variable somewhat uncertain.
Clinical implications
An increasing number of individuals are surviving OHCA. Previous studies indicate that the risk of recurrent OHCA and mortality peaks in the first year following an OHCA.
By being able to identify high-risk individuals, we will be able to tailor follow-up, select patients for ICD implantation and other interventions. Importantly, the performance of a model with the top 15 variables is similar to that of a model with 886 variables. Thus, we have elaborated a clinically useful prediction models that can be used at discharge after OHCA.
Conclusions
In patients discharged alive from hospital following an OHCA, roughly 18% suffered a recurrent arrest or death within 1 year. Using machine learning algorithms, we have achieved a relatively well performing prediction model predicting death or recurrent arrest in the year following an OHCA (AUC = 0.73 for 886 variables) as well as comprehensive models with similar predictive capabilities despite substantially fewer variables used (AUC = 0.69 for 15 variables). Age is the most important predictor of death or recurrent arrest, followed by CPC-4 and disposable income. Comorbidities are as a group the most important predictors, but no single diagnosis is among the most important predictors.
Funding
The Wallenberg Centre for Molecular and Translational Medicine. Swedish Research Council (2019–02019), Swedish state under the agreement between the Swedish government, and the county councils (ALFGBG-971482).
Rawshani A, Herlitz J, Lindqvist J, Aune S, Strömsöe A. Svenska Hjärt- och Lungregistret - Årsrapport 2020 [Swedish Register for Cardioplumonary Resuscitation - Year Report 2020], 2021.
Long-term prognosis following resuscitation from out of hospital cardiac arrest: role of percutaneous coronary intervention and therapeutic hypothermia.
Cardiac arrest in COVID-19: characteristics and outcomes of in- and out-of-hospital cardiac arrest. A report from the Swedish Registry for Cardiopulmonary Resuscitation.
Discriminatory cardiac arrest care? Patients with low socioeconomic status receive delayed cardiopulmonary resuscitation and are less likely to survive an in-hospital cardiac arrest.
Kuhn et al. Tidymodels: a collection of packages for modeling and machine learning using tidyverse principles. 2020 [cited June 29, 2022]. Available from: https://www.tidymodels.org.
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.
Socio-economic differences in incidence, bystander cardiopulmonary resuscitation and survival from out-of-hospital cardiac arrest: A systematic review.
00751-1&id=gr1.jpg){kind=link}
00751-1&id=gr2.jpg){kind=link}
00751-1&id=gr3.jpg){kind=link}