Introduction
Out-of-hospital cardiac arrest (OHCA) is a critical public health burden affecting approximately 400,000 persons in the United States annually where only 10% survive [
1]. While advances in resuscitation science have improved survival rates, mortality varies widely by geography, emergency medical services (EMS) agency, and hospital [
2]. While some of the variation has been attributed to OHCA characteristics (i.e., presenting rhythm, age, receipt of bystander cardiopulmonary resuscitation), variations in post-cardiac-arrest hospital care, such as use of coronary angiography and revascularization, when needed, may explain some of the heterogeneity seen when comparing survival and good neurological outcome [
3].
Machine Learning (ML) is a subfield of artificial Intelligence (AI) where algorithms learn tasks by studying high volumes of data [
4,
5]. The advent of big data and use of electronic health records enable us to pursue solutions to critical health issues. While traditional statistical methods are the standard for investigating patient and treatment intervention and associated outcomes, studies have suggested that ML algorithms provide greater insights across a wide variety of clinical settings [
6]. AI models use data to predict future events on the basis of the statistical weight of historical correlations and identify sensitive points within the system of care to direct strategic allocation of resources to improve disparities in clinical outcomes. ML has already proved useful in healthcare applications including medical imaging [
7], disease outbreak prediction [
8], drug discovery/usage [
9,
10], and hospital workflow optimization [
11,
12].
The OHCA care workflow is a time-sensitive process that requires quick and effective decision making throughout the chain of survival. ML has been applied at several stages of the care workflow to aid in predicting risk and recognition of cardiac arrest. During calls to emergency centers, conversations can be monitored using a ML model to help identify a cardiac arrest [
13]. Wearable devices can monitor vitals to predict the occurrence of a cardiac arrest for high risk individuals [
14]. ML has also been used to predict in-hospital cardiac arrests (IHCA) based on vital monitoring [
15].
Previous ML studies in OHCA have been limited by small population size and lack of diversity [
16], absence of pre-hospital data in model development [
17], and by not discriminating overall survival from survival with good neurologic outcomes [
17]. In addition, there remain significant research gaps on the utilization of ML models for post-return of spontaneous circulation (ROSC) decision-making throughout the OHCA workflow and their impact on survival outcomes. Powerful and affordable computer technologies enable us to combine big data to evaluate interactions that affect decision-making and survival. This study aims to develop ML models that effectively predict hospital’s post-ROSC practice to perform coronary angiography in adult patients with ROSC after OHCA and subsequent neurologic outcomes.
Results
The training set consists of 957 OHCA patients. Of these patients, 209 receive a Coronary Angiography, and 198 have a neurological outcome of Class 0. The training set is used to develop the ML models. Each model utilizes the training data to learn the underlying patterns in the data with the objective of performing the classification task. For the decision making model, the objective is to learn how the decision to perform a Coronary Angiography is being made based on the real instances recorded in the data.
The validation set consists of 241 cardiac arrest patients, of which, 52 receive a Coronary Angiography, and 51 have a neurological outcome of Class 0. Additional file
1: Table S3 provides detailed information about the optimized hyperparameters for each model. Table
1 presents the AUROCs for each ML model on the tasks of modeling CA and CPC. The EFCN model achieves the best AUROC for CA and CPC with scores of 0.8836 and 0.9272, respectively. Additional file
1: Table S4 provides complete modeling results of all models with a variety of metrics and complete confusion matrices.
Table 1
Results on validation set in terms of AUROC
LightGBM | 0.7050 | 0.7462 |
Random Forest | 0.6641 | 0.6437 |
XGBoost | 0.6639 | 0.7561 |
Gradient Boost | 0.6619 | 0.6977 |
Decision Tree | 0.6415 | 0.6537 |
k-Nearest Neighbor | 0.6781 | 0.7077 |
Logistic Regression | 0.6937 | 0.7417 |
EFCN | 0.8836 | 0.9272 |
The testing set consists of 600 cardiac arrest patients, of which 130 receive a CA, and 130 have a neurological outcome of Class 0. The AUROC of the CA and CPC EFCN models on the testing set are 0.9079 and 0.8967, respectively. Additional file
1: Table S2 provides SHAP value averages for each model broken by class. Additional file
1: Table S4 provides information about the testing results of all models and additional evaluation metrics.
The cohort analysis set consists of 600 cardiac arrest patients. For these 600, 94 are removed in unused hospitals as discussed in the Cohort Analysis subsection. For the cohorts there are 132, 156, and 218 cardiac arrest patients in Cohorts 1–3, respectively. Each of these cohorts have 26, 34, and 43 patients that receive a CA, respectively. The CA EFCN model is reoptimized based on the Cohort 1 training and validation data. When evaluating the performance of the reoptimized model on the respective cohort sets, AUROCs of 0.9761, 0.6601, and 0.6371 are achieved for the respective cohorts. Table
2 demonstrates the expected model changes of the reoptimized Cohort 1 model on the other 2 cohorts. This table first shows the patients that did not receive a Coronary Angiography and then what happens to their expected output where a positive change means that a patient is now expected to survive with CPC1 or 2. The Cohort 1 model expected 33 of the 175 Cohort 3 patients without a CA to be given a CA. Then using the survival model, the Cohort 1 model predicts a positive change in survival for 10 of the 33 patients with a changed CA.
Table 2
Cohort analysis of the reoptimized Cohort 1 EFCN models on the other cohort data
Patients that were not initially given CA, that the model predicts to get CA 23/122 (18.85%) | No change in CPC class 18/23 (78.3%) | Patients that were not initially given CA, that the model predicts to get CA 33/175 (18.86%) | No change in CPC class 23/33 (69.7%) |
Positive change in CPC class 5/23 (21.7%) | Positive change in CPC class 10/33 (30.3%) |
Negative change in CPC class 0/23 (0%) | Negative change in CPC class 0/33 (0%) |
Discussion
Our decision and survival models achieve testing AUROCs of 0.9079 and 0.8967, respectively. These evaluation metrics are similar to the AUROCs of the validation set, which suggests these models do well in generalizing to unseen data. Our cohort analysis showed how modeled changes in decisions could impact OHCA survival. When evaluating the lowest tertile (Cohort 3) with models based on the highest tertile (Cohort 1), our model showed a change in the decision to perform CA for 18.86% of the patients and predicted a positive impact in CPC class for 30.3% of the patients with a changed decision. To our knowledge, this is the first study to show that ML modeling can not only effectively predict patient outcome after an OHCA but can also predict hospital practice to perform CA post-ROSC.
Because decision-making in healthcare often involves large amounts of data, ML and simulation can be useful tools to predict how different variable combinations affect patient outcomes [
6]. ML has been successfully used in prognosis, diagnosis, treatment, clinical workflow, and expanding the availability of clinical expertise [
41]. In our previous work, a ML model using data from the Chicago CARES dataset had an AUROC of 0.825 in predicting survival with favorable neurological outcomes among patients with a witnessed OHCA [
36]. A Korean study of deep learning ML models used electronic health record data to predict subsequent cardiac arrest in hospitalized patients with an AUROC of 0.850 [
15]. ML models of OHCA can also predict survival outcome. A study from the Swedish Registry of Cardiopulmonary Resuscitation (SRCR) reported an accuracy of 0.82 in predicting survival after OHCA [
42]. The Korean OHCA registry had a better performance in predicting neurologic outcomes with an AUROC of 0.953, but this study included only patients who had sustained ROSC [
43]. In our study utilizing data from patients who survived to hospitalization post-OHCA, the AUROC of 0.8967 for survival with functional CPC was better than both previous studies.
Perhaps the most powerful use of ML models is as virtual laboratories for examining the interaction of treatment strategies and interventions under different patient variables and systems of care circumstances that may be otherwise costly, time-consuming, or even unethical to manipulate in the real world. Such models can help decision-makers within OHCA systems of care make tactical decisions regarding resource allocation or adapt treatment guidelines to local context. Ours is the first model to predict decisions with a high accuracy level regarding provision of CA after OHCA with an AUROC 0.8836. Moreover, we were able to show that when hospitals in Cohort 3 (lowest CPC tertile) changed the decision-making regarding coronary angiography to resemble decisions made by Cohort 1 (highest CPC tertile), more patients could survive with functional neurologic outcome. OHCA systems of care can use ML models to critically review OHCA treatment guidelines and test how different decisions may affect patient outcomes before costly and time-consuming implementation and training.
Our findings also have significant implications for how emergency systems of care define cardiac arrest care centers (CACs). Some studies have suggested that hospital case volume and coronary angiography capabilities are associated with better outcomes [
44‐
47]. However, post-OHCA care is complex and requires the coordination of multiple specialties including neurointensivists, cardiologists, pulmonary, and critical care specialists, to name a few. Our study demonstrates the power of ML as a tool to inform decision-making for systems of care. In the future, EMS systems of care without formalized CAC agreements could develop ML models to identify which hospitals to preferentially transport patients post-OHCA. ML models can also be developed to perform continuous quality improvement of treatment in the field and on the hospital side of care. Hospitals can also use ML for benchmarking against other hospitals within the same system of care and to simulate how changes in local treatment guidelines impact patient outcomes before implementing these changes at a larger scale. These ML models can also be adapted to other emergency systems of care such as stroke, myocardial infarction, and trauma.
One of our study limitations is its limited generalizability as external validation is needed to further interrogate the performance of the final model. Another important limitation is our inability to define why hospitals make different interventions decisions although caring for patients with similar demographics and cardiac arrest characteristics. Specifically, the CARES data does not include data on the presence or absence of ST-elevation on EKG and it does not provide sufficient detail to measure and compare the utilization of other resources, such as expertise in neuroprognostication or the presence of a cardiac arrest champion. The CARES data set is also limited in that it does not include details on comorbid illnesses that influence prognostication and the decision to perform CA such as cancer and end stage renal disease. Despite the data limitations, our models show promise for ML as a tool to predict hospital variations in post-ROSC care and subsequent neurologic outcome in OHCA.
Acknowledgements
We would like to thank the Chicago Fire Department Emergency Medical Services for sharing their cardiac arrest data and for their leadership and vision; we especially want to acknowledge the support of Mary Sheridan and Joanne Farrell. We also would like to thank Dr. Bradshaw Bunney and Dr. Pavitra Kotini Shah of the Illinois Heart Rescue leadership team and Dr. Bryan McNally, executive director of CARES, for data management support and valuable insight in model development and manuscript preparation.
Open AccessThis article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit
http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (
http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.