Article, Cardiology

An alternative tool for triaging patients with possible acute coronary symptoms before admission to a chest pain unit

a b s t r a c t

Objective: This study aimed to develop a Triage tool to more effectively triage possible ACS patients presenting to the emergency department (ED) before admission to a protocol-driven Chest pain unit (CPU).

Methods: Seven hundred ninety-three Clinical cases, randomly selected from 7962 possible ACS cases, were used to develop and test an ACS triage model using cluster analysis and stepwise logistic regression.

Results: The ACS triage model, logit (suspected ACS patient) = -5.283 + 1.894 x chest pain + 1.612 x age + 1.222

x male + 0.958 x proximal radiation pain + 0.962 x shock + 0.519 x acute heart failure, with a threshold value set at 2.5, was developed to triage patients. Compared to four existing methods, the chest-pain strategy, the Zarich’s strategy, the flowchart, and the heart broken index (HBI), the ACS triage model had better performance.

Conclusion: This study developed an ACS triage model for triaging possible ACS patients. The model could be used as

a rapid tool in EDs to reduce the workloads of ED nurses and physicians in relation to admissions to the CPU.

(C) 2017

Introduction

Acute coronary syndrome (ACS), a common presentation at the emergency department (ED), represents a difficult diagnostic challenge. Because a missed diagnosis of ACS may lead to further Ischemic events and a potentially Preventable death or disability, ACS patients need im- mediate care once they present to EDs. However, ACS patients may ini- tially appear well, without significant presenting symptoms, and thus may be easily ignored or the severity of their condition underestimated. Hence, to reduce the mortality and morbidity of ACS patients, a rapid tri- age is critical to distinguish them for immediate care from those who can safely wait in overcrowded EDs.

To triage suspected ACS patients, several methods have been pro- posed. Chest pain has been accepted as a critical symptom to rapidly tri- age possible ACS patients for fast and efficient protocol-driven diagnostic testing, called the “accelerated Diagnostic protocol” (ADP). More specifically, the chest pain unit (CPU), suggested by the College of Cardiology/American Heart Association guidelines [1], is commonly applied by hospitals around the world to rule in and rule out suspected ACS patients [2,3-5]. However, as mentioned by many studies [6-9], the use of chest pain alone is inadequate and can result in false-positive test results and unnecessary downstream procedures. To make the triage

? Source: Taiwan Ministry of Science and Technology (MOST103-2221-E-155-053-MY3).

* Corresponding author.

E-mail address: [email protected] (R.F. Lin).

more effective, Zarich, Sachdeva [10], Sanchez, Lopez [7], and Lopez, Sanchez [11] have used criteria in addition to chest pain to triage pa- tients prior to admission to CPUs. Nonetheless, certain limitations remained while using these methods to identify possible ACS patients, driving the need for further investigation.

To increase the accuracy of triaging possible ACS patients seen in a CPU or any observation unit (OU), this study uses cluster analysis and stepwise logistic regression to establish a triage tool, called the “ACS tri- age model.” The ultimate purpose here was to avoid unnecessary delays and improve resource utilization by reducing the number of patients seen in CPUs. The ACS triage model was expected to increase the iden- tification of suspected ACS patients, rather than the detection of coro- nary artery diseases. Optimal interpretation and diagnosis rely on the CPU, along with clinical judgment.

Material and methods

Study design and setting

This study was conducted in the Far Eastern Memorial Hospital, a 1000-bed regional teaching hospital with a percutaneous coronary in- tervention (PCI) center located in Northern Taiwan. Since 2007, the hos- pital has employed several methods for ACS patients. These methods include the introduction of an audit program, re-design of the patient processing flow, modification of bottle-neck processes using six-sigma methodology, and the triage of patients before admission to an ACS OU.

https://doi.org/10.1016/j.ajem.2017.12.026

0735-6757/(C) 2017

To triage possible ACS patients, the hospital employs a self-devel- oped triage model [12], called the “heart broken index” (HBI). The HBI model consists of four critical symptoms, comprising chest pain, epigas- tric pain, cold sweating, and dyspnea. Triage nurses in the ED used the HBI, embedded in the ED information system, to triage patients over 18 years of age by asking if he/she had any of the four critical symptoms. However, due to the variance of the patient descriptions, these symp- toms actually had broader meanings. Chest pain represented symptoms such as genuine chest pain, discomfort, and any sensations of heaviness, pressure, tightness, or compression. Epigastric pain included pain or dis- comfort in the area between the lower margin of the sternum and the umbilicus. Cold sweating referred to the patient sweating profusely, which may wet their clothing, despite being stationary. Dyspnea re- ferred to symptoms like shortness of breath, and feelings of suffocating or air hunger. If the patient was experiencing chest pain or any two of the other three symptoms, he/she was considered a suspected ACS pa- tient and admitted to the ACS OU with an ADP. In the ACS OU, patients underwent several tests, comprising the 12-lead ECG, blood biomarkers, and X-rays. The ED physician made a diagnosis based on patients’ clini- cal presentations and lab findings. Patients with objective evidence of ACS were immediately admitted for urgent therapy. However, if the pa- tient was ruled out by the protocol, she/he was discharged from the ED. Those suspected of being a victim of ACS were kept in the OU or sent to the cardiology ward (if available) for confirmation of the diagnosis. Such a patient may spend several hours in the ACS OU. The time spent in the ACS OU varied from 30 min to 72 h. The objective of applying the HBI was to avoid failing to diagnose an ACS patient, especially in cases that are not sensitive to chest pain.

Data collection

A total of 810 clinical cases, from June 2012 to December 2014, were randomly selected from 7962 clinical cases that applied the HBI in the ED of the Far Eastern Memorial Hospital. Our preliminary study showed that about 15% of patients who were administered the HBI should be ad- mitted to the ACS OU. Hence, the sampling size was determined and given a tolerance interval (12.5%, 17.5%) and a significance level of

0.05. The number of total clinical cases that needed to be collected

was 768. Therefore, 810 cases were collected, so that even with possible problems of duplicated records or incomplete information, the study sample size would be sufficient. To collect the data, ED nurses who were trained research assistants went to the medical record room to take pre-randomly-selected paper-based medical records, and then typed relevant information into an Excel file. For each case, gender, age, prior history of hypertension or diabetes mellitus, past history (both congenital and surgical) of heart disease or kidney disease, pre- senting symptoms at the ED, and final diagnosis made by physicians were recorded. All data were collected in binary format, except age, which was normalized with a minimum value of 18 years and a maxi-

mum value of 100 years (i.e., (age – 18)/(100 – 18)). This study re- ceived formal approval from the ethics committee and no personal

information was retained. The study was acknowledged by the institu- tional review board.

Several pre-analysis tasks were performed before the model’s devel- opment. First, because a presenting symptom could be written on paper-based medical records with different names and different for- mats by different ED physicians or at different times, an ED physician (who was part of the research team) categorized and summarized all the written symptoms into standardized symptom names. Furthermore, because different symptoms might indicate the same function failure, the ED physician also combined symptoms under a representative com- bined name. For example, as shown in Table 1, “chest pain” was defined as chest pain, chest discomfort, chest heaviness, chest pressure, Chest tightness, and chest compression. Second, the collected cases were checked for data validity. Among the 810 cases, 17 cases were invalid due to duplicate records and incomplete information. In total, 793

cases were deemed valid and used in further analyses. Finally, based on the final diagnosis of each valid clinical case, the ED physician made a judgment of whether the patient was a “suspected ACS patient” (the predicted state in further analyses) and should be admitted to the ACS or OU. Suspected ACS patients were those diagnosed with ACS, acute myocardial infarction (AMI), myocardial infarction (MI), ST-seg- ment elevation myocardial infarction (STEMI), non-ST-segment eleva- tion myocardial infarction (NSTEMI), and unstable angina .

Model development and validation

A model development procedure (Fig. 1) was implemented to devel- op and validate the ACS triage model. Techniques used in the procedure included cluster analysis, stepwise logistic regression, signal detection theory, and receiver operating characteristic (ROC) methodology. All statistical analyses were performed using SPSS 18. To alleviate concerns about practical usage and patient safety, the model was developed to meet three criteria: 1) achieve 100% sensitivity for identifying STEMI pa- tients, 2) achieve at least 90% sensitivity for identifying ACS patients, and 3) consist of no more than six predictors.

In total, eight steps were executed to develop and test the ACS triage model. In Step 1, the unconstrained cluster analysis was applied to all 793 clinical cases to extract key predictors for suspected ACS patients. In Step 2, the 793 cases were split into a modeling data set, with 594 cases randomly selected (approximately three fourths of the total sam- ple) and a testing data set of 199 cases (approximately one fourth of the total sample). This step was repeated five times (as Step 3 and Step 4) to generate five sets of modeling data and testing data. The five-selected modeling data sets in Step 3 were used to build five preliminary models, using stepwise logistic regression. Here, the key predictors extracted in Step 1 served as initial predictors in the preliminary models. In Step 4, the five preliminary models were tested for consistency and efficiency. In Step 5, the intermediate model was built based on these preliminary models, with the best combination of the selected predictors in the pre- liminary models. In Step 6, the ACS triage model was determined, based on the intermediate model with a threshold, log(c). The threshold was set so that the ACS model could achieve desired sensitivities (i.e., N 90%). In Steps 7 and 8, the ACS triage model was compared to the chest-pain model, the Zarich’s model [10], the flowchart model [7,11], and the HBI model [12]. The details of each step are stated in Appendix A.

Results

Clinical case analyses

Forty-two symptoms and 98 diagnoses were reported across the 793 clinical cases. In Table 1, the demographic and clinical features that oc- curred more than four times in the overall sample, and in the ACS and non-ACS groups, are listed. Also, Table 1 shows the symptoms included in the symptom categories determined by the ED physician. These symptoms included chest pain, acute heart failure, shock, distal radia- tion pain, proximal radiation pain, nausea, fever, consciousness change, and sub-acute heart failure. Chest pain (i.e., chest pain, discomfort, heaviness, pressure, tightness, and compression) was the most fre- quently occurring symptom (96.60% of the total sample). Among the 98 diagnoses, 21 diagnoses occurred twice and 30 diagnoses occurred once. Table 2 summarizes the 47 diagnoses that occurred more than twice. The case frequencies of the five diagnoses that required admis- sion to a CPU were 99 ACS cases, 22 MI cases, seven STEMI cases, six NSTEMI cases, and three UA cases; a total of 137 (17.28% signal rate) cases. Note that although MI, STEMI, NSTEMI, and UA were all in the ACS category, the diagnoses were made by the physician with sufficient evidence, whereas diagnoses of ACS were made with insufficient evidence.

Table 1

Demographic characteristics and clinical features that occurred more than four times in the overall sample, and in the ACS and non-ACS groups.

Variables

Relevant symptoms Overall

(n = 973)

ACS (n = 132)

Non-ACS (n = 661)

Age(year), mean +- SD

– 57.87 +- 15.96

60.52 +- 13.57

57.34 +- 16.37

Gender(male), n (%)

– 511 (64.44)

111 (84.09)

400 (60.51)

History of heart disease, n (%)

– 211 (26.61)

44 (33.33)

167 (25.26)

History of hypertension, n (%)

– 321 (40.48)

66 (50.00)

255 (38.58)

History of diabetes, n (%)

– 166 (20.93)

31 (23.48)

135 (40.34)

Chest pain, n (%)

Chest pain/Discomfort/ Heaviness/Pressure/ Tightness/Compression

766 (96.60)

131 (99.24)

635 (96.07)

Acute heart failure, n (%)

Dyspnea/Shortness of breath/Difficulty breathing

291 (36.70)

59 (44.70)

232 (35.10)

Shock, n (%)

Cold sweating/Pallor/

Diaphoresis/Shock

179 (22.57)

49 (37.12)

130 (19.67)

Distal radiation pain, n (%)

Teeth pain/Gingival pain/

Abdominal pain/Epigastric pain

105 (13.24)

16 (12.12)

89 (13.46)

Proximal radiation pain, n (%)

Radiation pain/Left arm

soreness/Neck compression

93 (11.73)

26 (19.70)

67 (10.14)

Nausea, n (%)

Nausea/Vomiting

72 (9.08)

16 (12.12)

56 (8.47)

Dizziness, n (%)

71 (8.95)

9 (6.82)

62 (9.38)

Palpitation, n (%)

41 (5.17)

2 (1.52)

39 (5.9)

Cough, n (%)

33 (4.16)

1 (0.76)

32 (4.84)

Weakness, n (%)

27 (3.41)

3 (2.27)

24 (3.63)

Fever, n (%)

Fever/Chillness

21 (2.65)

1 (0.76)

20 (3.03)

Headache, n (%)

Consciousness change, n (%)

Collapse/Consciousness

11 (1.39)

9 (1.14)

1 (0.76)

3 (2.27)

10 (1.51)

6 (0.91)

Diarrhea, n (%)

Loss/Presyncope/Syncope

9 (1.14)

1 (0.76)

8 (1.21)

Back pain, n (%)

Sub-acute heart failure, n (%)

Dyspnea on exertion/

7 (0.88)

6 (7.57)

1 (0.76)

0 (0)

6 (0.91)

6 (9.08)

Leg edema, n (%)

Difficult to walk

5 (0.63)

2 (1.52)

3 (0.45)

Note: Shaded symptoms indicate several similar symptoms could be combined as a general symptom name.

Developed of the preliminary and intermediate models

The critical predictors and corresponding weights of the five prelim- inary models are shown in Table 3. The five preliminary models consisted of different combinations of a total seven predictors, compris- ing chest pain, age, male, proximal radiation pain, shock, acute heart failure, and hypertension. Age, proximal radiation pain, shock, and male were found to be critical in all the five models. Chest pain was crit- ical in the four preliminary models, whereas acute heart failure and hy- pertension were critical in just one model. In all the preliminary models, the coefficients of critical symptoms that occurred more than once (i.e., chest pain, age, male, proximal radiation pain, and shock) were statisti- cally indifferent, at a significance level of 0.10.

The prediction of each preliminary model was tested with the 594 modeling cases using ROC methodology. The area under the curve (AUC) showed the discriminability of each model. The AUCs and their 95% upper and lower confidence intervals for the five data sets were 0.722 (0.669, 0.775), 0.729 (0.674, 0.783), 0.715 (0.658, 0.772), and

0.739 (0.687, 0.790). Comparison of these AUCs showed that with the same set of initial potential predictors selected by the clustering method (Step 1), the discrimination rates of all the resulting intermediate models were similar. This indicates that the preliminary models were consistent.

In total, five intermediate models were tested, and a model with five predictors was chosen as the final intermediate model. As shown in Table 3, the five preliminary models resulted in four combinations of predictors. Four preliminary models with these combinations, and an additional model that included all the seven predictors, were tested

for their Predictive performance. Among these five models, the final in- termediate model consisted of six predictors (Eq. 1), comprising chest pain, age, male, proximal radiation pain, shock, and acute heart failure, had the best predictive performance. Furthermore, the number of pre- dictors in the intermediate model met the first criterion we set for prac- tical usage.

Intermediate odds ratio = -5.283 + 1.894 x chest pain + 1.612

x age + 1.222 x male + 0.958

x proximal radiation pain + 0.962

x shock + 0.519 x acute heart failure (1)

Development of the ACS triage model

The ACS triage model developed (Eq. (2)) was built based on the in- termediate model (Eq. (1)) and the threshold value of log(c). Specifical- ly, chest pain represented a patient who presented with any of the following symptoms: chest pain, chest discomfort, chest heaviness, chest pressure, chest tightness, or chest compression. Age was given as a normalized age value with a minimum value of 18 years old and a maximum value of 100 years old. Male represented a male patient. Proximal radiation pain represented a patient who had any of the fol- lowing symptoms: radiation pain, left arm soreness, or neck compres- sion. Shock represented a patient who had any of the following symptoms: cold sweating, pallor, diaphoresis, or shock. Acute heart fail- ure represented a patient who had any of the following symptoms: dys- pnea, shortness of breath, and difficulty breathing. To meet the desired

Fig. 1. The eight-stEP model development procedure of the ACS triage model.

sensitivities of 95%, 98%, and 99%, log(c) was set at 2.05, 2.45, and 2.50. The method for identifying a suspected ACS patient was adj. odds ratio N 0.

Adj.odds ratio = -5.283 + 1.894 x chest pain + 1.612 x age

+ 1.222 x male + 0.958

x proximal radiation pain + 0.962 x shock

+ 0.519 x acute heart failure + log(c) (2)

ACS triage model validation and comparison

The predictive performance of the ACS triage model was evaluated by prediction of the states of the five sets of 199 testing clinic cases. Fig. (2) shows the model’s performance using ROC methodology. The area under the ROC curve was 0.731 (95% CI; 0.691, 0.770). To illustrate the relationship between critical value c, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), three sets of results for different log (c) corresponding to sensitivities above 95%,

98%, and 99% were analysed to compare the predictive performance with those of the other four triage models.

The ACS triage model had better predictive performance compared with the other four triage models. The predictive performance of the five models, including the chest-pain model, the Zarich’s model, the flowchart model, the HBI model, and the ACS triage model with three threshold values, are summarized in Table 4 and Fig. 3. All the models had 100% sensitivity for STEMI cases and had N 90% sensitivity for ACS patients. The ACS triage model with the threshold value set at 2.5, com- pared with the chest-pain model, the Zarich’s model, and the HBI model, had the highest discriminability index (d?), sensitivity, specificity, PPV, and NPV. Although the flowchart model had better specificity, com- pared with the ACS triage model with the threshold value set at 2.5, the ACS triage model with the threshold value set at 2.05 had better d

?, sensitivity, specificity, PPV, and NPV. Although the chest-pain and HBI models had high sensitivity (99.24%), they had very low specificities (3.93% and 4.08%). Although the flowchart model had good specificity (17.25%), it had very low sensitivity (93.18%). The Zarich’s model had moderate performance in terms of d?, sensitivity, specificity, PPV, and NPV.

Table 2

Forty-seven frequent (occurring more than twice) diagnoses reported in the 793 clinical cases.

Number

Diagnosis name

Frequency

Occurrence

1

Chest pain

296

37.28%

2

Acute coronary syndrome

99

12.47%

3

Coronary artery disease

73

9.19%

4

Chest tightness

37

4.66%

5

Gastroesophageal reflux disease

34

4.28%

6

Anxiety

23

2.90%

7

Arrhythmia

22

2.77%

8

Myocardial infarction

22

2.77%

9

Angina

21

2.64%

10

Congestive heart failure

21

2.64%

11

Dyspnea

18

2.27%

12

Palpitation

15

1.89%

13

Hyperventilation

15

1.89%

14

Dizziness

14

1.76%

15

Hypertension

14

1.76%

16

Pneumonia

14

1.76%

17

Chest discomfort

12

1.51%

18

Aortic dissection

11

1.39%

19

Peptic ulcer disease

11

1.39%

20

Abdominal pain

11

1.39%

21

Mitral valve prolapse

10

1.26%

22

upper respiratory infection

10

1.26%

23

Shortness of breath

8

1.01%

24

Epigastric pain

7

0.88%

25

Atypical chest pain

7

0.88%

26

St segment elevation myocardial infarction

7

0.88%

27

None

7

0.88%

28

Diabete mellitus

7

0.88%

29

Vertigo

6

0.76%

30

Non-St segment elevation myocardial infarction

6

0.76%

31

Paroxysmal supraventricular tachycardia

5

0.63%

32

acute gastroenteritis

5

0.63%

33

Gastritis

5

0.63%

34

Cerebral vascular accident

5

0.63%

35

Fever

4

0.50%

36

Chronic obstructive pulmonary disease

4

0.50%

37

Headache

4

0.50%

38

Bronchitis

4

0.50%

39

Anemia

4

0.50%

40

General discomfort

4

0.50%

41

lung edema

3

0.38%

42

Heart failure

3

0.38%

43

Hepatocellular carcinoma

3

0.38%

44

Unstable angina

3

0.38%

45

Chronic kidney disease

3

0.38%

46

end stage renal disease

3

0.38%

47

Myofascial pain

3

0.38%

Note: Shaded diagnoses indicate that the patient who should be admitted to a CPU or a OU.

Discussion

In attempting to triage possible ACS patients before admitting them to CPUs (or any OU), this study used a model development procedure to develop a rapid triage tool, called the “ACS triage model.” Although CPUs and other OUs with ACDs were diversely established to fit the needs of individual EDs in different parts of the world, such as the United States, Europe, Spain, and Germany [13], the use of chest pain to triage suspected patients for admission to CPUs is essential. However, the use of chest pain alone has resulted in a high false alarm rate [6-9]. To this end, Zarich, Sachdeva [10], Sanchez, Lopez [7], Lopez, Sanchez [11], and Hsu, Chen [12] were a few studies that attempted to improve the triage method by introducing additional factors and rules. However, because these studies developed their models mainly based on the clin- ical experiences, their models provided limited improvements, and have not been widely applied by other EDs. To the best of the authors’ knowledge, this research is the first to apply systematic statistical

methods to develop the triage model. The application of statistical methods helps to scientifically select critical factors and assign adequate weights of importance. The ACS triage model predicted suspected ACS patients based on six presenting symptoms. These six critical symp- toms, ranked in order of importance (i.e., the weights), were chest pain, age, male, proximal radiation pain, shock, and acute heart failure. Of these criteria, except for age and male, chest pain, proximal radiation pain, shock, and acute heart failure were not specifically defined. In- stead, these criteria individually represented several similar symptoms. The unspecific definitions of these criteria could be helpful for the ED nurses to process rapid triages, because of certain constraints (e.g., lim- ited time and patient abilities) of the interactions between triage nurses and ED patients.

Of the six factors of the ACS triage model, chest pain was the most important factor, and was used in all the five models. Studies [14,15] have shown its importance when triaging ACS patients. As shown in Table 2, the chest-pain model, which simply uses chest pain as the

Table 3

Critical predictors and the corresponding weights of the five preliminary models.

Parameter

Model 1

Model 2

Model 3

Model 4

Model 5

Average

Intercept

-5.684

-5.844

-3.792

-5.741

-5.352

-5.283

Chest pain

1.884

1.872

2.061

1.761

1.894

(0.141, 3.628)

(0.116, 3.627)

(0.307, 3.816)

(0.028, 3.493)

Age

1.875

2.091

1.514

1.224

1.356

1.612

(0.858, 2.893)

(1.035, 3.146)

(0.477, 2.550)

(0.201, 2.247)

(0.332, 2.379)

Male

1.290

1.153

1.237

1.330

1.100

1.222

(0.802, 1.778)

(0.678, 1.627)

(0.734, 1.740)

(0.842, 1.818)

(0.633, 1.568)

Proximal radiation pain

0.764

1.293

1.146

0.859

0.726

0.958

(0.241, 1.287)

(0.784, 1.802)

(0.639, 1.653)

(0.325, 1.394)

(0.181, 1.271)

Shock

0.936

0.981

0.879

1.108

0.907

0.962

(0.523, 1.349)

(0.556, 1.405)

(0.459, 1.298)

(0.711, 1.506)

(0.497, 1.318)

Acute heart failure

0.519

0.519

Hypertension

0.410

(0.137, 0.901)

0.410

(0.026, 0.795)

Note: values in brackets show 95% confidence intervals.

triaging criterion, has good sensitivity (99.24%) for triaging ACS pa- tients. This is the reason why the American Heart Association Scientific Statement [14] has suggested that hospitals use it as a single symptom to triage patients for executing ECG inspections. However, in addition to resulting in a high number of False alarms, merely using chest pain to triage patients may miss ACS patients. Jayes, Beshansky [16] and Douglas and Ginsburg [17] reported that chest pain is not significant for women. According to Brieger, Eagle [18], patients with ACS who present without chest pain, especially females, are frequently misdiagnosed and undertreated. According to our results, the chest- pain model had a miss rate of 0.76%. Thus, there is a motivation to use additional predictors to improve the model’s prediction.

Other than chest pain, the factors of age, male, proximal radiation pain, shock, and acute heart failure were found to be critical when de- veloping the ACS triage model. Among these predictors, shock and acute heart failure are considered by the HBI model, under the symptom names of “cold sweating” and “dyspnea.” As a result of discussions with our ED physicians, we used “shock” to cover the symptoms of cold sweating, pallor, diaphoresis, and shock, and used “acute heart failure” to cover the symptoms of dyspnea, shortness of breath, and difficulty breathing, because these symptoms deliver similar messages. This

study confirms that the use of cold sweating and dyspnea by the HBI model is reasonable. The importance of age was supported by Anderson, Adams [19], and was considered in the HBI model, the Zarich model, and the flowchart model, by setting age as a binary factor. Different from these models, the ACS triage model used a normalized-continuous age value to better describe the effect of age. Moreover, the ACS triage model showed that male and proximal radiation pain are other influen- tial factors. The importance of these factors is supported by previous studies. For example, Simon, Griffin [2] applied shoulder or Arm pain as a critical symptom when triaging ACS patients. Furthermore, ACS oc- curs more frequently in male patients [18,20]. The use of the model de- velopment procedure not only confirms the effects of these reported factors when triaging possible ACS patients, but also indicates the extent to which the factors affect the likelihood of judgment.

Based on the collected clinical cases, the ACS triage model had better predictive performance than the other four triage models. As expected, the chest pain strategy had high sensitivity (99.24%) but the lowest specificity (3.93%). Compared to the chest-pain model, the HBI model that included an additional three symptoms (epigastric pain, cold sweating, and dyspnea) should have had better prediction performance. However, the results showed slight improvements in terms of

Fig. 2. Receiver operator characteristic curve of the ACS triage model to predict ACS patients.

Table 4 Comparison of the predictive performance of the ACS triage model with the other four tri- age models

Model

d?

Sensitivity

Specificity

PPV

NPV

Sensitivity of STEMI

Chest-pain

0.69

99.24%

3.93%

17.10%

96.30%

100%

Zarich’s

0.84

96.97%

14.98%

18.55%

96.12%

100%

Flowchart

0.55

93.18%

17.25%

18.36%

92.68%

100%

Heart broken index

0.69

99.24%

4.08%

17.12%

96.43%

100%

ACS (threshold: 2.05)

1.33

96.97%

29.12%

21.36%

97.98%

100%

ACS (threshold: 2.45)

1.32

98.79%

16.97%

19.11%

98.60%

100%

ACS (threshold: 2.5)

1.49

99.39%

15.40%

18.92%

99.22%

100%

Note: d?, discriminability index; PPV, positive predictive value; NPV, negative prediction value; ACS, acute coronary syndrome; STEMI, ST-segment elevation myocardial infarction.

specificity, PPV, and NPV. The Zarich’s model that expanded the criteria to any male patient over the age of 35 years and any female patient over the age of 40 years presenting with any non-traumatic chest pain, as ex- pected, had the higher specificity (with improvements of 1.05%) than the chest-pain model. More improvements of 13.21% obtained by the flowchart model that ruled out ACS patients without chest pain, or with chest pain but without 1) age of 40 years or younger, 2) diabetes,

3) coronary artery disease (not previously known), and 4) non- retrosternal pain. However, both models sacrificed a certain degree of sensitivity, reducing sensitivity by 2.27% and 6.06%, respectively. Com- pared to the four abovementioned triage models, the ACS triage model had better performance. Aimed at achieving a high sensitivity, with a threshold value at -2.5, the model’s highest sensitivity is 99.39% and

maintains a reasonable specificity of 84.60%. Aimed at a highest specific-

ity, with a threshold value set at -2.05, the model can reach a specific- ity of 29.12% and maintain a high sensitivity of 96.97%.

The ACS triage model aims to triage possible ACS patients seen in CPUs (or any OU) when presenting at the ED. If EDs triage patients with a very low threshold, or do not triage at all, this results in unneces- sary ECG inspections and cardiac biomarker measurements. These pro- cedures expose patients to ionizing radiation and increase ED costs and the workloads of ED physicians and nurses. Using the model as a rapid triage tool provides the first assessment to prioritize patients stressed

by the considerable demands of emergency care and limited ED re- sources. Although more accurate diagnoses of ACS can be obtained in CPUs, without appropriate screening, EDs can become busy with non- coronary patients. To this end, the ACS triage model is an intervention that could be easily executed universally at EDs without considerable investments in new resources. The ACS triage model could be imple- mented as an app installed on a mobile device, or as a program built into an ED information system for triage nurses to triage possible ACS patients. The setting of the threshold value gives the model flexibility to tradeoff the costs of false alarms and misses that could be determined by EDs. The rule-in suspected ACS patients should be admitted to CPUs for further evaluation. The rule-out patients should follow a normal ED process. Thus, the ACS triage model does not decrease admission rates for patients who have ACS, but is expected to minimize unnecessary ECG inspections, cardiac biomarker measurements, and hospitalization, to reduce crowding of EDs and the workloads of ED nurses and physicians.

Limitations

The sample of clinical cases collected by this study has several limi- tations, and the study results should be interpreted within the scope and limitations of its design. Although we attempted to eliminate bias using data collected during a 31-month period, the total sampling pool of 7962 clinical cases was restricted to patients who were adminis- tered the HBI. Hence, the data has a limited scope of interpretation. First, our sample may not represent the whole ED population, because all the patients had at least one of the four critical symptoms: chest pain, epi- gastric pain, cold sweating, and dyspnea. It is possible that suspected ACS patients did not have any of the four HBI symptoms. Second, unlike Lopez, Sanchez [11] who validated the flowchart by collecting one- month re-visit follow-up data, our true state was limited to the diagno- ses made by ED physicians before the patients left the ED. The patients’ statuses may have been inconsistent if follow-up procedures had been applied. Third, although we tested the Zarich’s model for model com- parison, the results we obtained do not represent those of the five- step triage non-ACS flowchart proposed by Sanchez, Lopez [7] and

Fig. 3. Comparison of the sensitivity, specificity, positive predictive value, and negative predictive value of the ACS triage model with the other four triage models.

Lopez, Sanchez [11]. Due to limitations of our sample of clinical cases, the Zarich’s model used in this study did not use the factor of non-op- pressive pain. Hence, the original five-step flowchart model may have better predictive performance, especially a lower false alarm rate. Final- ly, this was a single-hospital system study and the collected data were specific to our institution. This data may differ from data collected by other hospitals in other places, especially in different countries.

The abovementioned limitations suggest two avenues of future re- search. The first avenue is to collect data that are more representative. Ideally, the sampling pool would reflect all patients presenting to the ED. However, this sampling method is time intensive, and may not be practical if a hospital does not have an electronic record system. An al- ternative method would be to use a sampling pool of presenting pa- tients who have any symptoms of ACS. The second avenue for future research is to validate the ACS triage model using practical tests, with follow-up data collection. The results of the ACS triage model, compared with the other four triage models, are encouraging us to test the practi- cal performance of the ACS triage model.

Conclusions

This study developed an ACS triage model, using the factors of chest pain, age, male, proximal radiation pain, shock, and acute heart failure, for triaging suspected ACS patients who need to be admitted to a CPU (or any OU). The ACS triage model was shown to have better predictive

the predicted state was calculated as shown in Eq. (A1).

Odds ratio = ?0 + ?i(significant predictors) (A1)

X

i

The parameters ?0 and ?i were coefficients of intercept and signifi- cant symptoms, respectively.

Step 4. The five preliminary models were tested for consistency and efficiency. First, predictor coefficients in the five preliminary models were examined for consistency. Second, the ROC curves of each prelim- inary model were presented to illustrate their efficiency and consistency.

Step 5. The five preliminary models obtained were used to deter- mine an intermediate model. The models obtained in Step 4 may have had different combinations of predictors, so any occurring combination of predictors was treated as a possible intermediate model, and tested to predict the states of the five sets of 199 testing cases generated from Step 2. When testing these possible intermediate models, the predictor coefficients were averaged values in the five preliminary models, if all estimated coefficients were within 90% confidence intervals. Otherwise, only coefficients within 90% confidence intervals were considered to de- termine the averaged coefficients. The signal detection theory was ap- plied to test the performance of these potential intermediate models. The model that had the best predictive performance was selected as the final intermediate model. This intermediate model outputs the in- termediate odds ratio for predicted state, as shown in Eq. (A2).

performance than other four triage models, and as such could be imple- mented by EDs as a rapid tool to triage patients for admission to a CPU or wait for a regular admission. The implementation of the ACS model is

Intermediate odds ratio = ?0

+ ?i(other predictors), ?iN1 (A2)

i

X

expected to reduce costs and ED workloads associated with diagnostic testing and hospitalization.

Declarations of interest

This study received grant support from the Taiwan Ministry of Sci- ence and Technology (MOST103-2221-E-155-053-MY3) for funding the research and paper submission. There are no other declarations of interest to declare.

Acknowledgements

We thank the Taiwan Ministry of Science and Technology (MOST103-2221-E-155-053-MY3) for their grant support. We also thank Rachel Pang, Pei-Li Chung, and Ming-Fen Guo for data collection.

Appendix A. The eight steps of the development of the ACS triage

Step 6. The ACS triage model was built based on the final intermedi- ate model (Step 5) and introduced the threshold log(c), with “c” representing the costs ratio of misidentification and false alarm (see Eq. (A3)). The threshold value of log(c) was determined from five re- served testing data sets and constrained by the corrected prediction of a patient’s true state of suspected ACS patient (i.e., sensitivity). The ACS triage model predicted the likelihood that a presenting patient was a suspected ACS patient who should be admitted to the CPU. When costs of missed identification and false alarm are equal, the way of identifying a given patient’s state was to see if the probability of being a suspected ACS patient was greater than he/she not being a suspected ACS patient (i.e., intermediate odds ratio N 0). The selection criteria may vary from hospital to hospital, to tradeoff the ED costs. In this study, three threshold values were suggested to show different combinations of sensitivity and false alarm rates. The adjusted odds ratio (adj. odds ratio), Eq. (A4), was calculated based on the intermedi- ate odds ratio and the costs ratio c. Thus, a suspected ACS patient was determined by the decision rule-in Eq. (A5).

model are detailed as follows

Step 1. The unconstrained cluster analysis was first applied to 793

P(ACS) = exp(intermediate odds ratio) (1 + exp(intermediate odds ratio))

(A3)

clinical cases, to exclude symptoms unrelated to the predicted state

(Suspected ACS Patient). Symptoms that were identified as related to

Adj.odds ratio = Intermediate odds ratio + log(c) (A4)

the predicted state were selected for further model development. The distance between each symptom was measured using the average link- age method.

Suspected ACS patient = Yes, if adj.odds ratioN0

(A5)

Step 2. Five hundred ninety-four cases (approximately three fourths of the total sample) were randomly selected from 793 valid cases as modeling data, whereas the remaining 199 cases (approximately one fourth of the total sample) were saved as testing data. This step was re- peated five times, to avoid the mover fitting problem.

No, otherwise

Step 3. Five preliminary models were built, based on the five sets of

594 cases using stepwise logistic regression, where the entering signif- icance equaled 0.05 and the leaving significance equaled 0.1. The possi- ble critical symptoms used to build these preliminary models were the predictors extracted from Step 1. The logistic regression equation of

Step 7. Four triage models, comprising the chest-pain model, the Zarich’s model, the flowchart model, and the HBI model, were tested to predict the states of all the 793 clinical cases. The chest-pain model triaged possible ACS patients simply using chest pain. The Zarich’s model triaged patients based on chest pain, with an additional age crite- rion: male patients over the age of 35 years or female patients over the age of 40 years. The flowchart model ruled out ACS patients without chest pain, or with chest pain but without 1) age of 40 years or younger,

2) diabetes, 3) coronary artery disease (not previously known), and 4) non-retrosternal pain. Note that because information on non-

oppressive pain was not available in our collected dataset, this criterion was removed from the flowchart model used in the comparison. The HBI model triaged possible ACS patients based on chest pain (with an age of 30 years or older), epigastric pain, cold sweating, and dyspnea; suspected ACS patients were those who had the symptom of chest pain or any two of the other three symptoms.

Step 8. To demonstrate the effectiveness of the developed ACS triage model, its predictive performance was compared with that of the other four triage models.

References

  1. Wright RS, Anderson JL, Adams CD, et al. 2011 ACCF/AHA focused update incorporat- ed into the ACC/AHA 2007 guidelines for the management of patients with unstable angina/non-ST-elevation myocardial infarction: a report of the American College of Cardiology Foundation/American Heart Association Task Force on Practice Guide- lines. J Am Coll Cardiol 2011;57:e215-367.
  2. Simon EL, Griffin P, Medepalli K, et al. Door-to-balloon times from freestanding emergency departments meet ST-segment elevation myocardial infarction reperfu- sion guidelines. J Emerg Med 2014;46:734-40.
  3. Blankenship JC, Haldis TA, Wood GC, Skelding KA, Scott T, Menapace FJ. Rapid triage and transport of patients with ST-elevation myocardial infarction for percutaneous coronary intervention in a rural health system. Am J Cardiol 2007;100:944-8.
  4. Diercks DB, Kirk JD, Lindsell CJ, et al. Door-to-ECG time in patients with chest pain presenting to the ED. Am J Emerg Med 2006;24:1-7.
  5. Hess EP, Knoedler MA, Shah ND, et al. The chest pain choice decision aid. Circ Cardiovasc Qual Outcomes 2012;5:251-9.
  6. Hess EP, Brison RJ, Perry JJ, et al. Development of a clinical prediction rule for 30-day cardiac events in emergency department patients with chest pain and possible acute coronary syndrome. Ann Emerg Med 2012;59:115-25.
  7. Sanchez M, Lopez B, Bragulat E, et al. Triage flowchart to rule out acute coronary syndrome. Am J Emerg Med 2007;25:865-72.
  8. Pope JH, Aufderheide TP, Ruthazer R, et al. missed diagnoses of acute cardiac ische- mia in the emergency department. N Engl J Med 2000;342:1163-70.
  9. Rosenfeld AG, Knight EP, Steffen A, Burke L, Daya M, DeVon HA. Symptom clusters in patients presenting to the emergency department with possible acute coronary syn- drome differ by sex, age, and discharge diagnosis. Heart Lung 2015;44:368-75.
  10. Zarich SW, Sachdeva R, Fishman R, et al. Effectiveness of a multidisciplinary quality improvement initiative in reducing door-to-balloon times in primary angioplasty. J Interv Cardiol 2004;17:191-5.
  11. Lopez B, Sanchez M, Bragulat E, et al. Validation of a triage flowchart to rule out acute coronary syndrome. Emerg Med J 2010;28:841-6.
  12. Hsu J-C, Chen K-C, Cheng I-N, Li A-H. Using heart broken index to improve the diag- nostic accuracy of patient with STEMI and shorten door-to-balloon time on emer- gency department. American Heart Association 2011 Scientific Sessions. Orlando, Florida; 2011.
  13. Durand E, Chaib A, Danchin N. Chest pain and Chest pain units. The ESC Textbook of Intensive and Acute Cardiovascular Care 2015;56.
  14. Amsterdam EA, Kirk JD, Bluemke DA, et al. Testing of low-risk patients presenting to the emergency department with chest pain a scientific statement from the American Heart Association. Circulation 2010;122:1756-76.
  15. Farkouh ME, Aneja A, Reeder GS, et al. clinical risk stratification in the emergency department predicts long-term cardiovascular outcomes in a population-based co- hort presenting with acute chest pain: primary results of the Olmsted County chest pain study. Medicine 2009;88.
  16. Jayes RL, Beshansky JR, D’Agostino RB, Selker HP. Do patients’ coronary risk factor re- ports predict Acute cardiac ischemia in the emergency department? A multicenter study. J Clin Epidemiol 1992;45:621-6.
  17. Douglas PS, Ginsburg GS. The evaluation of chest pain in women. N Engl J Med 1996;

    334:1311-5.

    Brieger D, Eagle KA, Goodman SG, et al. Acute coronary syndromes without chest pain, an underdiagnosed and undertreated high-risk group: insights from the Global Registry of Acute coronary events. Chest 2004;126:461-9.

  18. Anderson JL, Adams CD, Antman EM, et al. ACC/AHA 2007 guidelines for the man- agement of patients with unstable angina. J Am Coll Cardiol 2007;50:e1-157.
  19. Fan C-M, Lai C-L, Li A-H, Tsai K-C, Chung K-P. Reducing the door-to-balloon time for ST-segment elevation myocardial infarction: the result of a hospital quality im- provement initiative. J Emerg Med 2010;12:107-13.

Leave a Reply

Your email address will not be published. Required fields are marked *