Predictors of Clinically Significant Echocardiography Findings in Older Adults with Syncope: A Secondary Analysis
BACKGROUND: Syncope is a common reason for visiting the emergency department (ED) and is associated with significant healthcare resource utilization.
OBJECTIVE: To develop a risk-stratification tool for clinically significant findings on echocardiography among older adults presenting to the ED with syncope or near-syncope. DESIGN: Prospective, observational cohort study from April 2013 to September 2016
SETTING: Eleven EDs in the United States
PATIENTS: We enrolled adults (≥60 years) who presented to the ED with syncope or near-syncope who underwent transthoracic echocardiography (TTE).
MEASUREMENTS: The primary outcome was a clinically significant finding on TTE. Clinical, electrocardiogram, and laboratory variables were also collected. Multivariable logistic regression analysis was used to identify predictors of significant findings on echocardiography.
RESULTS: A total of 3,686 patients were enrolled. Of these, 995 (27%) received echocardiography, and 215 (22%) had a significant finding on echocardiography. Regression analysis identified five predictors of significant findings: (1) history of congestive heart failure, (2) history of coronary artery disease, (3) abnormal electrocardiogram, (4) high-sensitivity troponin-T >14 pg/mL, and 5) N-terminal pro B-type natriuretic peptide >125 pg/mL. These five variables make up the ROMEO (Risk Of Major Echocardiography findings in Older adults with syncope) criteria. The sensitivity of a ROMEO score of zero for excluding significant findings on echocardiography was 99.5% (95% CI: 97.4%-99.9%) with a specificity of 15.4% (95% CI: 13.0%-18.1%).
CONCLUSIONS: If validated, this risk-stratification tool could help clinicians determine which syncope patients are at very low risk of having clinically significant findings on echocardiography.
REGISTRATION: ClinicalTrials.gov Identifier NCT01802398.
© 2018 Society of Hospital Medicine
Candidate Predictors
Potential candidate predictors were identified through a prior expert panel process.20,21 Candidate predictors included age, sex, abnormal heart sounds, exertional syncope, shortness of breath, chest pain, near-syncope, family history of sudden cardiac death, high (>180 mm Hg) or low (<90 mm Hg) systolic blood pressure, abnormal ECG, elevated hs-TnT, elevated NT-proBNP, and history of the following: hypertension, cardiac dysrhythmia, renal failure, diabetes, congestive heart failure (CHF), and coronary artery disease (CAD).
The first obtained ECG was abstracted by one of five research study physicians blinded to all clinical data. Research study physicians demonstrated high interrater reliability (kappa > 0.80) in distinguishing normal from abnormal ECGs in a training set of 50 ECGs. Abnormal ECG interpretations included nonsinus rhythms (including paced rhythms), multiple premature ventricular complexes, sinus bradycardias (≤40 bpm), ventricular hypertrophies, short PR segment intervals (<100 milliseconds [ms]), axis deviations, first degree blocks (>200 ms), complete bundle branch blocks, Brugada patterns, Wolff-Parkinson-White patterns, abnormal QRS duration (>120 ms) or abnormal QTc prolongations (>450 ms), and Q/ST/T segment abnormalities suggestive of acute or chronic ischemia.
Statistical Analysis
We calculated descriptive statistics for each predictor variable, stratified by the presence or absence of TTE findings. Chi-square and t-tests were used to test associations between categorical or continuous variables and TTE findings using a significance level of 0.05 and 2-sided hypothesis testing. To identify a robust set of predictors of the primary outcome, we used multivariate logistic regression with the LASSO (Least Absolute Shrinkage and Selection Operator) to fit a parsimonious model.22 The LASSO selects variables and shrinks the associated coefficients to avoid overfitting.23-25 We then used a bootstrap to generate confidence intervals for coefficient estimates. Cases with missing echocardiography reports were excluded from the analysis. Bootstrap results were summarized as the percentage of bootstrap iterations in which each variable’s coefficient was 1) chosen and negative, 2) shrunk to zero, or 3) chosen and positive.
We assessed different weighting schemes to generate a risk score from significant variables identified by regression modeling. These included weighting by regression coefficients rounded to the nearest integer and simple summation of the presence or absence of each variable.
Based on these results, a predictive score was developed to risk stratify patients on their probability of major, clinically significant findings on TTE. The sensitivity and specificity of a score of zero to predict findings on TTE was calculated. For confidence intervals, we used Wilson’s method for binomial confidence intervals.26 The receiver operating characteristic (ROC) curve and its associated area under the curve (AUC) were calculated, and a confidence interval for the AUC was obtained through bootstrap resampling with 2,000 iterations. As part of our sensitivity analyses, we also calculated the ROC curve and AUC after excluding the patients with a known history of CHF and significant finding on TTE. Data analyses were performed in R.27 Two sensitivity analyses were performed: 1) we used multiple imputation to impute 1,000 complete data sets and then used the same LASSO methodology as with the complete data to assess whether incorporating missing data changed the results; and 2) we simulated a conventional troponin assay by raising the positive threshold for hs-TnT to >30 pg/mL (corresponding to the limit of detection for conventional troponin).28
