A novel cardiovascular risk stratification model incorporating ECG and heart rate variability for patients presenting to the emergency department with chest pain

Design and setting

We conducted a prospective, non-randomized, observational study of patients presenting
to the ED with chest pain from March 2010 until August 2015. This study was performed
at the ED of the Singapore General Hospital (SGH), a tertiary care hospital in Singapore.
ED triage is performed by nurses using the national Singaporean patient acuity category
scale (PACS), a symptom-based triage system without strict physiological criteria.
ED patients are classified with a PACS score, which ranges from 1 to 4 and represents
the degree of urgency in patient attendance. Patients with PACS 1 are the most critically
ill, those with PACS 2 are non-ambulant, those with PACS 3 are ambulant, and those
with PACS 4 are non-emergencies. Our study focuses on patients presenting with chest
pain, who routinely receive a 12-lead ECG investigation (Philips PageWriter TC50 Cardiograph)
during triage and are placed in PACS 1 or 2 units where they receive further ECG monitoring
(ZOLL X Series Monitor defibrillator). The study was approved by the local ethics
committee (SingHealth Centralised Institutional Review Board, Singapore) with a waiver
of patient consent.

Patient recruitment and eligibility

All patients older than 21 years of age with a primary complaint of non-traumatic
chest pain were eligible. Patients presenting in non-sinus rhythm (arrhythmias, asystole,
complete heart blocks, or pacemaker rhythms) were excluded due to interference of
these phenomena with the interpretation of QRS complexes. Similarly, patients with
a high percentage of artifacts, ectopic beats, and non-sinus beats (30 % of ECG recordings)
were excluded due to their potential biasing effect on the HRV calculations 27]. Finally, patients who were lost to follow up or transferred to other (private) hospitals
within the 30-day time frame were excluded, on account of inability to ascertain whether
these patients had reached our primary endpoint.

Data collection and processing

All data were collected on standardized forms in a Research Electronic Data Capture
(REDCap) database. The electronic medical records (EMRs) were analyzed for demographic
characteristics, medical history, presenting symptoms, clinical information, and laboratory
results.

A trained research coordinator prospectively downloaded 12-lead ECG tracings from
the ZOLL X Series monitor defibrillator on a daily basis. We use our in-house software
package for ECG signal processing and parameter calculation 28]. Noise was manually removed from the lead II ECG tracing and its sample of 6 minutes
was stored in an Excel (Microsoft Office 2007; Microsoft, Redmond, WA, USA) file for
further processing. A 5–28 Hz band-pass filter was applied to the lead II sample to
facilitate peak detection 29]. QRS complexes were detected using a threshold-plus-derivative method that has been
previously validated 28]. Time domain and frequency domain HRV parameters were calculated in accordance with
the guidelines outlined by the Taskforce of the European Society of Cardiology 30].

Vital signs were recorded at initial ED patient presentation using the Propaq CS Vital
Signs Monitor (Welch Allyn, Skaneateles, NY, USA). The first set of complete vital
signs obtained at initial presentation was used for this study. The 12-lead ECG tracings
recorded during triage were used for the evaluation of ECG variables. These tracings
were recorded using a Philips PageWriter TC50 cardiograph and subsequently extracted
for analysis and storage. A trained research associate, blinded to patient outcomes,
ascertained whether the patient was in sinus rhythm, and evaluated the 12-lead ECG
tracings for abnormalities. We followed the definitions of ECG variables as described
in John Hampton’s book “The ECG Made Easy”.

We tested the SEDRSM against the TIMI score for unstable angina (UA)/non-ST elevation
myocardial infarction (NSTEMI), a score that has been employed on the ED to predict
MACE within 30 days of presentation to the ED with chest pain 31]. Data pertaining to the TIMI score criteria were retrieved from the EMRs and used
to construct the TIMI score.

Outcomes

The primary endpoint of this study, MACE, was a composite outcome of death, acute
myocardial infarction, and revascularization, including coronary artery bypass graft
(CABG) or percutaneous coronary intervention (PCI), within 30 days of presentation
to the ED. Patients were followed up and EMRs were reviewed to ascertain whether the
patient had experienced an endpoint criterion within 30 days after presentation.

Statistical analysis

SPSS (version 21.0; SPSS Inc, Chicago, IL, USA) software was used for statistical
analysis. Derivation and validation of the SEDRSM was done in the same cohort. Univariate
relationships between baseline characteristics and MACE were assessed using the appropriate
statistical test, based on type and distribution of data. We tested normality of distribution
by inspecting normality graphs and interpreting the Kolmogorov-Smirnov quantitative
normality test.

A total of 16 HRV parameters, 13 ECG variables, 7 vital signs, and 3 demographic variables
(age, gender, and race) were screened as candidate predictors of MACE using the same
univariate analytical method described previously. Variables associated with a p value 0.05 were selected and categorized in order to facilitate scoring and increase
applicability at the ED. HRV parameter category cutoffs were chosen based on the visual
comparison between HRV parameter value and frequency of MACE occurrence. ECG variables
were dichotomous. Vital signs and demographics were categorized based on recognized
(physiological) cutoff values.

We introduced the categorized candidate variables into an automated likelihood ratio
backward stepwise logistic regression model. The retained candidate variables were
used to construct the SEDRSM. All unstandardized coefficients were normalized by dividing
them by the smallest coefficient, and subsequently rounded off to the nearest integer.
The SEDRSM score was then calculated by a simple arithmetic sum of the integers assigned
to the criteria satisfied.

The overall goodness of fit of the model was assessed by the Hosmer-Lemeshow test.
The predictive accuracy of the SEDRSM and TIMI score was assessed using the area under
the receiver operating characteristic (AUROC) curve. Discriminatory values (i.e.,
sensitivity, specificity, positive predictive value, and negative predictive value)
were also determined for both risk stratification models.