A Retrospective Study: Quick Scoring of Symptoms to Estimate the Risk of Cardiac Arrest in the Emergency Department

Purpose At present, not enough is known about the symptoms before cardiac arrest. The purpose of this study is to describe the precursor symptoms of cardiac arrest, focusing on the relationship between symptoms and cardiac arrest, and to establish a quick scoring model of symptoms for predicting cardiac arrest. Patients and Methods. A retrospective case-control study was carried out on cardiac arrest patients who visited the emergency department of Peking University Third Hospital from January 2018 to June 2019. Symptoms that occurred or were obviously aggravated within the 14 days before CA were defined as warning symptoms. Results More than half the cardiac arrest patients experienced warning symptoms within 14 days before cardiac arrest. Dyspnea (p < 0.001) was found to be associated with cardiac arrest; syncope and cold sweat are other symptoms that may have particular clinical significance. Gender (p < 0.001), age (p < 0.001), history of heart failure (p=0.006), chronic kidney disease (p=0.011), and hyperlipidemia (p=0.004) were other factors contributing to our model. Conclusions Warning symptoms during the 14 days prior to cardiac arrest are common for CA patients. The Quick Scoring Model for Cardiac Arrest (QSM-CA) was developed to help emergency physicians and emergency medical services (EMS) personnel quickly identify patients with a high risk of cardiac arrest.


Introduction
Cardiac arrest (CA) refers to the sudden termination of cardiac ejection that result in interruption of systemic blood circulation, respiratory arrest, and loss of consciousness [1]. As a potentially fatal cardiovascular event with poor prognosis, especially when it occurs outside of a hospital, CA has become a public health threat around the world. It is generally accepted that 4-6 minutes after the occurrence is the ideal time for treatment. However, there are several early warning symptoms of cardiac arrest, including chest pain, dyspnea, syncope, cold sweat, backache, abdominal pain, and palpitation [2,3], and the presence of these prodromal symptoms could persuade people to contact their doctors early enough to prevent cardiac arrest [4] if the public were aware of them. Yet no efective early warning model has been broadly employed, though it is of great signifcance to identify the high-risk population and provide timely warning of imminent cardiac arrest. Establishing a warning model based on prodromal symptoms would earn time for prevention and intervention of out-of-hospital cardiac arrest [5] by emergency medical services (EMS), and could ultimately promote public awareness. Tis study explored the correlations between CA and warning symptoms and therefore developed a symptoms-based quick scoring model for CA using one and a half years of emergency room data.

Patients and Study Design.
Tis single-center retrospective case-control study was conducted at Peking University Tird Hospital, which has about 240,000 emergency visits per year. From January 2018 to June 2019, we extracted data for 309 consecutive cardiac arrest (CA) patients. Te incidence of cardiac arrest was defned as the cessation of mechanical cardiac activity as confrmed by the absence of signs of circulation.
A total of 150 patients over 18 years old were enrolled in the case group after excluding patients with a history of trauma, drug overdose, drowning, pregnancies, end-stage malignancies, and patients with severe instances of missing data. Patients who were pronounced clinically dead before admission were also excluded because of the low proportion of sudden CAs, as confrmed by autopsy. All enrolled patients had detailed records of rescue and short-term outcomes after CA. Of the case group, 98 were out-of-hospital cardiac arrest (OHCA) and 52 were inhospital CA (IHCA). A control group of noncardiac arrest patients who visited the emergency department on the same date were randomly enrolled in a 1 : 3 ratio, totaling 450 cases in the control group. Te exclusion criteria for the case group were also used for the control group. Patients in a perioperative period were also excluded from the control group.

Data
Collection. Data were collected from the electronic health records of Peking University Tird Hospital by three authorized researchers on standard forms and double keyed into an Epidata 3.1 database (Odense, Denmark). Te case report form was designed based on the Utstein-style uniform guidelines for out-of-hospital CA, which includes the patient variables, arrest variables, prodromal symptoms and outcomes [6] shown in Table 1.
Identifcation of the conditions underlying CA is one of the interests of this study, which, according to previous research, includes age, cardiovascular diseases, respiratory diseases, and diabetes. Close attention was paid to patients' history of smoking and drinking. Symptoms that occurred or were obviously aggravated within 14 days before CA were defned as prodromal symptoms. All symptoms were recorded, with particular attention to dyspnea, chest pain, abdominal pain, backache, syncope, cold sweat, and palpitation [7]. Data on the duration and character of the symptoms were also collected [8].
Te primary outcome of this study was the incidence of cardiac arrest. Gender, age, underlying diseases, and early warning symptoms and were included in the model for clinical use.

Statistical
Analysis. R4.0.3 statistical software was used for statistical analysis. Te LRM function was used to perform binary logistic regression analysis. When the independent variable was the classifcation variable, the minimum value group was taken as the reference group. When the independent variable is a continuous variable, the continuous variable is directly included in the binary logistic regression model. Te enter regression was used to screen variables, and the nomogram and calibration curve were made. Te pROC package was used to draw the ROC curve. A p value < 0.05 was considered statistically signifcant. Diferent combinations were set up according to statistical stability and the researchers' clinical experience. Five-fold cross-validation was used for model evaluation.

Epidemiological and Clinical Characteristics of Cardiac
Arrest Patients. Te patient enrollment fow is shown in Figure 1. Of the 150-patient case group, 98 were OHCA and 52 were in-hospital cardiac arrest (IHCA). All 35 patients in the case group survived to discharge. Tough only 22 (22.4%) of the 98 OHCA patients had bystander CPR, OHCA patients had a 28.6% (n � 28) return of spontaneous circulation (ROSC), and a discharge survival rate of 23.5% (n � 23). A comparison of demographic and clinical characteristics between the cardiac arrest group and control group is shown in Table 2. Patients who sufered from cardiac arrest had more incidence of underlying disease than the control group.
We also presented several important clinical predictors and factors with statistical signifcance (Table 2), which indicates that more than half of the CA patients sufered from prodromal symptoms. Te seven listed symptoms are considered potential risk factors based on previous research: dyspnea, chest pain, abdominal pain, backache, syncope, cold sweat, and palpitation [4]. Tis study found four of them to show no signifcance, but dyspnea, syncope, and cold sweat were found to be related to cardiac arrest.

Risk Factors and the Quick Scoring Model for Cardiac
Arrest. Factors with statistical signifcance were included as independent variables to explore risk factors. Logistic regression analysis was used to select which independent variables were most suitable for the model. Our results concurred with previous research [6] in fnding that cardiac arrest is associated with gender, age, and three of the previously tested diseases. We also found dyspnea is another independent risk factor. Although stroke showed a signifcant correlation, it was not included in our model considering the small samples. Te model thus includes six independent risk factors: gender, age, heart failure history, chronic kidney disease, hyperlipidemia, and dyspnea (Table 3). Te characteristics of cardiac arrest patients in the emergency department can be well described by this model, which became our new scoring model, the Quick Scoring Model for Cardiac Arrest (QSM-CA), for emergency room physicians to identify high cardiac arrest risk patients early.
With logistic analysis, we were able to assign values to each of the variables (shown in Table 3) and total them as a predictive risk score. Te prediction model was constructed by nomograph ( Figure 2). Te AUC (C Index) of the area under the ROC curve was calculated to be 0.834. One of the standard methods for evaluating a model is through fve-fold cross-validation. Te results of all the folds are combined and reported. Sensitivity and specifcity were 39% and 95%, respectively, and the AUC was 0.829.

Discussion
Te incidence of OHCA in adults is 100.1/100,000 per year in Beijing, of which only 19.7% received cardiopulmonary resuscitation provided by emergency centers, and only 1.3% survived [9]. Because this study did not include pre-hospital    [7,10]. Nearly, 61% of the 150 patients in the current study had early warning or prodrome symptoms within 2 weeks before CA, which matches with Binz and Andrea [10] (56%). In our study, more than 50% of the symptoms occurred within 24 hours of CA, as shown in (Table 4). Tese fndings highlight the potential for developing a new model base on consideration of warning symptoms. Te incidence rate of Korean prodrome is almost the same as domestic statistics [7]. Compared with previous studies, the incidence of chest pain before cardiac arrest was 3.33%, lower than the Japanese study (20.7%) and another study (48.8%). While the incidence of dyspnea was 32%, higher than in the Japanese study and similar to another studies. Syncope (3.3%) was slightly less common, and cold sweat (8.7%) was more common.  (Table 3) and total them as a predictive risk score.
(b) ROC of the model, the AUC (C index) of the area under the ROC curve was 0.834.
As for the selection of the control group, it is difcult to select the perfect control group because the probability of symptoms before cardiac arrest is very low in healthy people and requires a large sample size. Terefore, we selected the population of outpatient emergency treatment as the control group, whose population is suitable for the medical applications (people who have some health problems).
Te model has great predictive value for cardiac arrest and the AUC was 0.834, which includes age, gender, underlying disease, and symptoms [9]. Te underlying diseases found to be associated with cardiac arrest were consistent with previous studies, including acute myocardial infarction, cerebrovascular disease, hypertension, heart failure, diabetes, chronic kidney disease, hyperlipidemia, and asthma. Only heart failure, chronic kidney disease, and hyperlipidemia contributed to the model. Diabetes and hypertension in the CA group (35.4% and 57.85%) and control group (18.24% and 33.62%) have a high prevalence rate in both CA, and are often accompanied by other symptoms, so it has not been included in the model. In order to improve the specifcity of the model and recognize deteriorating patients, we fnally chose not to include these basic diseases. According to previous research, a healthy lifestyle might reduce the risk of cardiac arrest [11]. Our study shows that both smoking history and alcohol history are risk factors for CA, but neither contributed to the model because a lack of ongoing records on these factors might cause systematic error and selection bias. Te conclusions of this research still need further confrmation and amendment by relevant studies.
We found that the incidence rate of chest pain and palpitation is very low. Te possible reasons are as follows: frst, because not all selected patients with CA are cardiogenic, of which cardiogenic accounts for 41.3% and noncardiogenic accounts for 58.6%; second, palpitation and chest pain, as subjective symptoms, are difcult to obtain after cardiac arrest. Dyspnea, as a symptom that can be observed and recorded by bystanders, appears more frequently in the main complaint. In the process of incorporating symptoms into the model, due to changes in consciousness and seizures often accompanied by dyspnea, they were not included in the model at last.
Te study has several strengths. Te model could improve the stratifcation of patients, enabling more immediate action for those known to be at higher risk to prevent further deterioration because most patients have antecedents before the attack and early intervention is associated with decreased unexpected cardiac arrests and unexpected deaths according to Chen's research [12].
Te combination of multiple laboratory tests and vital sign observations are used to develop predictive models such as National Early Warning Score (NEWS) [13]. In addition, the Modifed Early Warning Score (MEWS) [14] and the electronic Cardiac Arrest Risk Triage (eCART) score [15] have been investigated in a number of previous studies dealing with in-hospital patients [16]. Tese scoring methods, however, use CA variables that rely on accurate laboratory examination results and continuous monitoring of vital signs [17], which are difcult to obtain for OHCA. Te innovation of our Quick Score Model for Cardiac Arrest (QSM-CA) is that it is based on simple warning symptoms that can be quickly acquired in short conversations with patients or their families in an emergency department or EMS. Tis model is thus a useful tool to identify those at high CA risk and is applicable to the emergency environment.
It is the frst study in which warning symptoms are included as predictive factors. Strong correlations were found between several symptoms such as dyspnea, syncope, cold sweat, and cardiac arrest. Although we had an insufcient number of samples of syncope and cold sweat in this experiment, the models in which they were included showed remarkable predictability. Despite the inadequate sample, the results still have some clinical value for subsequent research. Vettor's research also suggested exercise-induced syncope as an important and alarming symptom of arrhythmic cardiac arrest [18]. Syncope and cold sweat were incorporated into two other models, however, both of them were unstable. Surprisingly, neither chest pain nor the diferent nature of chest pain showed signifcance in our study. With only 5 chest pains (3.33%) reported before CA in our study, which is markedly lower than previous studies (22%-48%) [3,19], it is far too risky to draw conclusive from one set of results.
Tis study has a few other limitations as well. It was a single-center retrospective study with limited data sources, all the data coming from the clinical records of emergency department patients. Due to the urgent circumstances characteristic of that environment, emergency medical personnel may not be able to identify all symptoms in the process of data collection. As the prediction factors included in the model are limited, any incomplete clinical information could cause migration error. Records of symptoms were also somewhat general; symptoms of diferent severity and nature can Emergency Medicine International represent totally diferent clinical signifcance. For example, severe persistent primordial pain shows a higher predictive value for CA than does chest pain caused by cough. As no consensus is available on the defnitions of some of the other warning symptoms, they can be challenged, which could have led to overestimation of their incidence in this study. Electronic in-hospital patient safety net systems like between the Flags have improved the early recognition of deteriorating patients [13] and the reduction in CA and CArelated mortality [20]. One of the challenges with any out-ofhospital health monitoring system is timely emergency calls from patients, families, or by-standers. Simple health warning systems embedded in personal terminals are necessary. Taking the occurrence of symptoms into account can help people decide when to seek help, such as whether to go to the emergency room or call emergency services. Our model initially shows the feasibility of establishing an out-ofhospital early warning system. It also helps EMS identify patients at high risk of CA and can be applied to the triage of patients in an emergency department.
In our next step in this experimental research, we plan to use EMS testing with current clinical data to validate and modify the model, which will be conducive to further development and improvement of the future application.

Conclusions
More than half of the patients in this study experienced warning symptoms within 24 hours before cardiac arrest. Dyspnea, syncope, and cold sweat may have specifc signifcance in patients presenting to an emergency department. Te Quick Score Model for Cardiac Arrest (QSM-CA) is set up by a combination of six independent risk factors: gender, age, heart failure history, chronic kidney disease, hyperlipidemia, and dyspnea, which helps with the timely identifcation of patients with high risk of cardiac arrest.

Data Availability
Te characteristics of patients data used to support the fndings of this study have not been made available because the data contains personal information on human subjects.

Ethical Approval
Tis study was approved by the Peking University Tird Hospital Medical Science Research Ethics Committee (IRB00006761-M2019353). An exemption from informed consent was acquired and the study was conducted according to the Declaration of Helsinki.