Effective combination of isolated symptom variables to help stratifying acute undifferentiated chest pain in the emergency department

Background Symptom is still indispensable for the stratification of chest pain in the emergency department. However, it is a sophisticated aggregation of several aspects of characteristics and effective combination of those variables remains deficient. We aimed to develop and validate a chest pain symptom score (CPSS) to address this issue. Hypothesis The CPSS may help stratifying acute undifferentiated chest pain in ED. Methods Patients with non‐ST segment elevation chest pain and negative cardiac troponin (cTn) over 3 hours after symptom onset were consecutively recruited as the derivation cohort. Logistic regression analyses identified statistical predictors from all symptom aspects for 30‐day acute myocardial infarction (AMI) or death. The performance of CPSS was compared with the symptom classification methods of the history variable in the history, electrocardiograph, age, risk factors, troponin (HEART) score. This new model was validated in a separated cohort of patients with negative cTn within 3 hours. Results Seven predictors in four aspects of chest pain symptom were identified. The CPSS was an independent predictor for 30‐day AMI or death (P < 0.001). In the derivation (n = 1434) and validation (n = 976) cohorts, the expected and observed event rates were well calibrated (Hosmer–Lemeshow test P > 0.30), and the c‐statistics of CPSS were 0.72 and 0.73, separately, significantly better than the previous history classifications in HEART score (P < 0.001). Replacing the history variable with the CPSS improved the discrimination and risk classification of HEART score significantly (P < 0.001). Conclusions The effective combination of isolated variables was meaningful to make the most stratification value of symptoms. This model should be considered as part of a comprehensive strategy for chest pain triage.


| INTRODUCTION
Chest pain and related symptoms rank the top reasons for visits to emergency department (ED) all over the world, 1,2 which are extremely heterogeneous with a wide spectrum of conditions ranging from lethal diseases such as acute myocardial infarction (AMI) to minor acute problems such as intercostal neuralgia. The inappropriate discharge of high-risk patients with AMI from the EDs would lead to nearly two times mortality when compared with admitted patients. 3 However, the majority of undifferentiated acute chest pain are low risk and not myocardial ischemic related, not requiring unnecessary further invasive tests or admission. 4,5 Therefore, rapid risk stratification models 2 | METHODS

| Study design
This is a prospective cohort study of acute non-ST segment elevation chest pain in the urban ED of the Qilu Hospital of Shandong University (a university-affiliated teaching hospital), and in the rural ED of the People's Hospital of Wenshang County from 24 August 2015 to 30 September 2017. The derivation cohort consisted of patients enrolled consecutively with negative initial cTn over 3 hours after symptom onset. And the validation cohort comprised chest pain patients with negative cTn within 3 hours after symptom onset. This study has been approved by the ethics committee at these two hospitals. Written informed consent was obtained from all participants.

| Patients enrollment
Any patient aged 18 or older, with acute nontraumatic chest pain occurring within the past 24 hours, and with negative initial cTn testing, was consecutively recruited.
Acute symptoms of myocardial ischemia, such as upper extremity, mandibular, or epigastric discomfort, or an ischemic equivalent such as dyspnea or fatigue, were also considered as chest pain according to the American Heart Association case definitions. 16 Feelings of actual pain in anatomical chest area or any symptom that indicated myocardial ischemia were identified as generalized chest pain in this study.
Negative cTn indicated that the value of initial testing was under the 99th percentile upper reference limit (URL), specific to the assay used by each site. The contemporary cTnI assays arranged by emergency physicians in their daily work were used to determine the eligibility of enrollment in the derivation and validation cohorts.
Patients were excluded if they had ST-elevation myocardial infarction (STEMI), or they were unable or unwilling to provide informed consent.

| Data collection and measurements
Data collection was prospectively conducted on a standardized case report form by trained research assistants, according to the standard definitions for each candidate variable from a published data dictionary. 16 Clinical symptoms were recorded as reported by the patient, and "no" was selected if certain situation was absent or unknown.
Detailed information about symptom complex included eight sections, namely character, location, radiation, severity, time course, associated symptoms, precipitating factors, and relieving factors (Table 1). Pain severity was quantified by using the numeric rating scale (NRS), integer from 0 for no pain to 10 for worst pain. 17 Heavy pain indicated the pain with NRS ≥ 5. Nocturnal onset indicated that the most significant symptoms occurred during the period from 8 PM to 8 AM.
Follow-up was conducted by trained research assistants through telephone interviews at 30 days after enrolment, to acquire information about major adverse cardiac events (MACE) and hospital attendances.
Calculation methods for the history and other variables of HEART score have been described in previous articles. 7,13,14 Four history classification methods are shown in Supporting Information Table S1.
These four methods provide more detailed information than other papers about which combinations of what variables are high-risk and which are low-risk. So, we translated these calculation methods into computer programs using SAS. ECG interpretation was conducted by two independent cardiologists blinded to the symptoms, cTn levels and events. And discrepancies were evaluated by a third cardiologist.
The HEART score was determined by the SAS programs to guarantee the consistency and comparability of scoring.

| Outcome
The primary outcome was the composite end point of death from all causes and AMI, including the index AMI and subsequent AMI, during the 30 days after presentation to the ED. Each event in MACE was independently adjudicated by two senior cardiologists of clinical

| Statistical analyses
Univariable logistic regression was used to determine whether each individual symptom characteristic helped to predict the 30-day MACE.
Odds ratios with 95% confidence intervals were reported. Variables with P value ≤0.2 were eligible to enter the stepwise multiple logistic regression. Separate multivariable logistic regression models were developed for each section of symptom profile. The variables with P < 0.05 in each section entered the next multiple logistic analysis and those with P < 0.05 were retained in the final model ( Figure S1).  (Figure 1).
Demographic characteristics, risk factors, medical history, and detailed symptom profiles of patients in derivation cohort are shown in Table 1.
Seven predictors in four aspects of chest pain symptom were identified after two stages of multivariable logistic regression with stepwise elimination (Table 2). For facilitated clinical use, integers were assigned for each of the seven predictors when they were present according to the relative numerical magnitudes of β coefficients.  Table 2). The expected and observed event rates were proved to be well calibrated using the Hosmer-Lemeshow test (P = 0.30) (Figure 2A).
The HEART score with CPSS replacing the history variable executed greater discrimination (AUC) than the standard HEART scores with different history measurements (P < 0.001) ( Figure 3B). And both NRI and IDI showed that the new model with CPSS resulted in significant improvement in predicting performance (P < 0.001) ( Table 3).
In the validation cohort, the CPSS performed well in calibration (P = 0.33) ( Figure 2B), and the discrimination was also excellent with an AUC reaching 0.73 (0.70, 0.76). In addition, CPSS improved the HEART discriminatory capacity and risk classification with a significant increase in c-statistic value, NRI, and IDI (Table 4).

| DISCUSSION
In this study, we developed a CPSS to assist in the prediction of 30-day AMI or death in acute undifferentiated chest pain presenting to the ED with negative initial cTn over 3 hours after symptom onset.
The CPSS served as an independent predictor and performed excellent in both the derivation and validation cohort. Moreover, the CPSS combined with the HEART risk score improved the discriminatory vomiting were shown to be significant predictors of AMI in accordance with other studies. [20][21][22] Some chest pain characteristics considered as "high-risk" in the HEART score were found to have no significant diagnostic or prognostic value, like exertional pain and relief with nitroglycerin. Besides, the reported "low-risk" variables 21 (stabbing/sharp character, influenced by posture/breathing/cough, etc.) were also shown to be of no importance.
Using the same measurements with the history variable of HEART score, the CPSS improved the discrimination and risk classification significantly. In general, the symptom variables were integrated into a predictive score or decision rule for clinical use. 7,9,[23][24][25] Symptom features were usually analyzed as isolated variable with risk factors, medical history, ECG, and cTn in the multivariable logistic regression and the ones with significant importance were selected into the final model. 8,9,25 However, there are some questions, whether single symptom variable was at the same level as ECG when using the multivariable regression analysis? Whether the organic integration of all aspects of symptom serve better than the isolated ones? Several studies have tried to establish a combinatorial chest pain score as one variable in the prediction model. However, the detailed development method and separate validation were not mentioned. 23,26,27 Therefore, in order to portrait the symptom profile from all aspects (character, location, severity, radiation, time course, and so on), we develop the CPSS by using two-step multivariable logistic regression. Interestingly, all of the variables except heaviness character in CPSS were identified as "typical" features in the HEART score, 13,14  For risk stratification of chest pain in ED, the CPSS could add predictive information to the negative cTn as an independent predictor.
Consistent with the kinetic profile of cTn releasing, 15 it is too early to detect troponin rise within 3 hours, leading to misdiagnosis of AMI.
The occurrence rate of AMI in patients with negative cTn within 3 hours after symptom onset was significantly higher than the rate in patients with negative cTn over 3 hours in this study. Undoubtedly, cTn should be retested later because of the relatively high misdiagnosis rate within 3 hours. But negative cTn over 3 hours after symptom was not safe enough, and there were still 6.5% major events missed, which was unacceptable for discharge. In the era of high sensitivity-cTn (hs-cTn), rule-in and rule-out of AMI in undifferentiated chest pain have been much more effective and safer than before. 4 But there are still some actual reasons for considering  3-hour pathway only using the hs-cTn to rule in and rule out AMI provides adequate diagnostic performance. 28 And data from one study has shown that the addition of clinical risk scores would improve the safety of pathways for early rule-out AMI. 29 Importantly, symptoms are the key components in these risk scores. Secondly, the accessibility of hs-cTn is insufficient around the world currently, especially in  Abbreviations: CPSS, chest pain symptom score; HEART, history, ECG, age, risk factors, troponin; HEART1~4, HEART scores with different history component classifications; HEART(CPSS), the HEART score with the CPSS replacing history component; IDI, integrated discrimination improvement; NRI, net reclassification improvement. the developing countries. Some in the Asia-Pacific tertiary centers still use contemporary standard assays. 30 As a result, since using hs-cTn is still a development trend, simple and economic diagnostic tools are indeed needed now, such as the rational use of symptoms. In our study, the CPSS improved the discrimination and risk classification of risk model significantly, providing important pretest probability and reminding physicians to interpret results of laboratory testing more carefully. Furthermore, the performance of CPSS remained excellent in patients with negative cTn within 3 hours, demonstrating the external utilization potentiality in this population.

| Limitations
This study also had several limitations. Firstly, the CPSS was developed from the ED chest pain patients in two hospitals of Shandong province in China. Although urban and rural hospitals were both covered and external validation was carried out in the same hospitals, the external generalization of the CPSS to wider patients should be determined by further studies in heterogeneous groups. Secondly, the HEART score was a very good reference substance here, but whether the CPSS-HEART should be used in clinical practice needs more assessments. Since symptom merely plays a certain part of roles in the diagnosis/prognosis of chest pain, the integration with other significant predictors is prerequisite, such as demographic characteristics, medical history, risk factors, ECG, and cardiac markers. The combination of these components and the weights for selected variables require many explorations. Thirdly, considering that the sensitivity of the hs-cTn is different from the contemporary cTn in patients assessed within 3 hours after onset of symptoms, we will take further evaluations based on the central testing of the hs-cTn to assess the performance of the CPSS.

| CONCLUSIONS
The CPSS performed well to assist in the prediction of 30-day AMI or death in acute undifferentiated chest pain presenting to the ED with negative initial cTn, better than the different symptom classification methods of the HEART score. And the HEART score with CPSS replacing the history variable executed greater discrimination and reclassification than the standard HEART scores. This demonstrated that the effective combination of isolated variables was meaningful to make the most diagnostic or predictive value of symptoms and made preparations for the development of complete stratification models with medical history, risk factors, symptoms, ECG, and troponins for chest pain triage.