Decision tree model for predicting in‐hospital cardiac arrest among patients admitted with acute coronary syndrome

Abstract Background In‐hospital cardiac arrest (IHCA) may be preventable, with patients often showing signs of physiological deterioration before an event. Our objective was to develop and validate a simple clinical prediction model to identify the IHCA risk among cardiac arrest (CA) patients hospitalized with acute coronary syndrome (ACS). Hypothesis A predicting model could help to identify the risk of IHCA among patients admitted with ACS. Methods We conducted a case‐control study and analyzed 21 337 adult ACS patients, of whom 164 had experienced CA. Vital signs, demographic, and laboratory data were extracted from the electronic health record. Decision tree analysis was applied with 10‐fold cross‐validation to predict the risk of IHCA. Results The decision tree analysis detected seven explanatory variables, and the variables' importance is as follows: VitalPAC Early Warning Score (ViEWS), fatal arrhythmia, Killip class, cardiac troponin I, blood urea nitrogen, age, and diabetes. The development decision tree model demonstrated a sensitivity of 0.762, a specificity of 0.882, and an area under the receiver operating characteristic curve (AUC) of 0.844 (95% CI, 0.805 to 0.849). A 10‐fold cross‐validated risk estimate was 0.198, while the optimism‐corrected AUC was 0.823 (95% CI, 0.786 to 0.860). Conclusions We have developed and internally validated a good discrimination decision tree model to predict the risk of IHCA. This simple prediction model may provide healthcare workers with a practical bedside tool and could positively impact decision‐making with regard to deteriorating patients with ACS.


| Study design and setting
We conducted a case-cohort study review on all adult ACS patients discharged from January 2012 to December 2016 from three tertiary hospitals in Fujian province, China. These participant hospitals included two comprehensive hospitals and one specialist hospital, with approximately 1200, 2500, and 1900 annual admissions of ACS patients, respectively. All physicians and nurses were required to receive Advance Cardiac Life Support training to ensure their ability to resuscitate patients.

| Study populations
At the participating hospitals, we identified a total of 21 337 ACS patients who had undergone a resuscitation attempt (chest compression and/or defibrillation) complicated by CA, from January 2012 to December 2016. In the case group, the inclusion criteria were: patients aged 18 years or older who had experienced a CA and had been admitted to hospital at least 24 hours later. Patients who were diagnosed with unstable angina (UA), acute ST-segment elevation myocardial infarction (STEMI) and acute non-ST-segment elevation myocardial infarction (NSTEMI) were included. CA was defined by unresponsiveness, apnea, and the absence of a central palpable pulse due to pulseless ventricular tachycardia or ventricular fibrillation, pulseless electrical activity (PEA), or asystole. In the case group, the exclusion criteria were: patients with CA whose family caregivers had refused resuscitation, patients with missing data, or patients with prior out-of-hospital CA and who were transported to hospital with ongoing resuscitation, or CA that had occurred during an operation, or patients whose situation was complicated due to multiple organ failure and who were in the terminal stage of their disease. For patients with more than one IHCA during the same hospitalization, only the first arrest was included.
All ACS patients residing at a participant hospital who had not had a previous CA were eligible for inclusion in the control group.
Patients were excluded if they had been discharged "against advice" or had missing data. Using a random number generator, we randomly selected the control subjects in a 1:3 ratio, with an ACS diagnosis, hospitalized in the same year and in the same hospital department (on the ward or in the ICU) as the CA patients in the case group.

| Data collection
Our study group of seven, who collected data from April 2015 to January 2017, consisted of a cardiologist, anesthetist, nurse, nursing master's student, epidemiology master's student, and two nursing interns. The information was retrieved from patients' electronic medical records.
All data that were routinely collected included: Demographics (age, gender, height, weight, body mass index, smoking, length of day, others parameters, in addition to systolic blood pressure, pulse rate, respiratory rate, body temperature, and AVPU score. The scores vary from 0 to 3 for each parameter, and from 0 to 21 in total value. 11 In our previous study, we reported that ViEWS was the most accurate tool for predicting CA, in comparison with the Modified Early Warning Score (MEWS) and National Early Warning Score (NEWS). 12 Laboratory values (white cell count, hemoglobin, platelet count, red cell count, lactate, sodium, potassium, chloride, bicarbonate, anion gap, blood urea nitrogen, glucose, serum creatinine, creatine kinase, creatine kinase-MB, brain natriuretic peptide, cardiac troponin I, others).

| Statistical analysis
Descriptive statistics were reported as mean ± SD, or median [interquartile range (IQR)] for continuous variables. For categorical variables, the percentages of patients in each category were calculated. Comparisons between categorical data were done by χ 2 test, and comparisons between continuous variables were done by Student's t test. As an estimate of effect size and variability, we have reported the odds ratio (OR) with a 95% confidence interval (CI).
We performed recursive portioning analysis to develop a decision tree model for IHCA prediction, which was used to separate patients into different homogeneous risk groups and to determine predictors for CA. 13 We assessed overall model discrimination by using the area under the receiver operating characteristic curve (ROC). To cope with the overfitting and instability inherent in the decision tree, a 10-fold cross-validation procedure was applied. Missing values were treated with imputation by random forest. Statistical analyses were conducted using R software, version 3.4.3 (Chicago, Illinois). All significance tests used a two-sided P value < 0.05.

| Baseline characteristics between the two groups
The baseline characteristics analysis between the two groups is shown in Table 1. There were no significant differences in hospital setting, hospital department, gender, diagnosis of ACS, hypertension, smoking, etc. (P > .05), however, age, diabetes, CCI and culprit artery were significantly different in the two groups (P < .05).
According to previous reports, age, 5,14 diabetes, 15,16 and CCI 17,18 were risk factors in predicting CA, so these may also be candidate predictors in our further analysis. The culprit artery of ACS cannot be clearly defined until an angiocardiography is done, which takes time and is unsuitable for early prediction, and so was not included in our analysis.  Table 2.  The most crucial predictor for CA was ViEWS. The National Institute for Health and Clinical Excellence recommended that physiologic tracking and trigger systems should be used to monitor all adult patients in an acute hospital setting 19 ; most such systems are also known as early warning scores. ViEWS is an early warning score established by Prytherch et al. 11 It was the best performing early warning score in a recent study that compared it to 33 other systems.

| The discrimination of the development and internal validation model
Our previous study also proved that ViEWS is better at discriminating the risk of CA compared to two other common early warning scores. 12 The second crucial predictor we evaluated was fatal arrhythmia.
Studies have shown that among 45% to 55% of patients with inferior acute myocardial infarction, 8% to 35% of these patients have varying degrees of atrioventricular block. 20  not included in the statistical analysis because they were missing more than 60% of the data, and this would call for more in-depth studies.

| CONCLUSIONS
We have developed and internally validated a good discrimination decision tree model based on seven factors to predict the risk of IHCA, enabling patients to be readily stratified into three groups of high, intermediate, and low risk of CA. This simple prediction model may provide healthcare workers with a practical bedside tool, and will impact decision-making for patients with ACS whose health is deteriorating.

ACKNOWLEDGMENT
This study was supported by the Individual Fund for Innovation Lead-