Pooled individual patient data from five countries were used to derive a clinical prediction rule for coronary artery disease in primary care

Abstract

Objective

To construct a clinical prediction rule for coronary artery disease (CAD) presenting with chest pain in primary care.

Study Design and Setting

Meta-Analysis using 3,099 patients from five studies. To identify candidate predictors, we used random forest trees, multiple imputation of missing values, and logistic regression within individual studies. To generate a prediction rule on the pooled data, we applied a regression model that took account of the differing standard data sets collected by the five studies.

Results

The most parsimonious rule included six equally weighted predictors: age ≥55 (males) or ≥65 (females) (+1); attending physician suspected a serious diagnosis (+1); history of CAD (+1); pain brought on by exertion (+1); pain feels like “pressure” (+1); pain reproducible by palpation (−1). CAD was considered absent if the prediction score is <2. The area under the ROC curve was 0.84. We applied this rule to a study setting with a CAD prevalence of 13.2% using a prediction score cutoff of <2 (i.e., −1, 0, or +1). When the score was <2, the probability of CAD was 2.1% (95% CI: 1.1–3.9%); when the score was ≥ 2, it was 43.0% (95% CI: 35.8–50.4%).

Conclusions

Clinical prediction rules are a key strategy for individualizing care. Large data sets based on electronic health records from diverse sites create opportunities for improving their internal and external validity. Our patient-level meta-analysis from five primary care sites should improve external validity. Our strategy for addressing site-to-site systematic variation in missing data should improve internal validity. Using principles derived from decision theory, we also discuss the problem of setting the cutoff prediction score for taking action.

Introduction

Applying individual patient meta-analysis to create clinical prediction rules is methodologically difficult when primary studies, acting independently, do not collect the same standard data sets. Methods to summarize the measures of prediction (e.g., regression coefficients) across studies must account for the data that individual studies did not try to collect. We encountered this problem when we used data from five independent studies of chest pain to develop a clinical prediction rule for initial assessment of patients presenting to a primary care setting. Chest pain is an important diagnostic problem in primary care, where 0.7–2.7% of patient encounters are due to chest pain [1], [2], [3], and coronary artery disease is the cause of chest pain in 8.6–14.6% of patients [3], [4]. Clinical prediction rules developed in emergency departments, specialty clinics, or hospitals may not apply to primary care because diagnostic test results (e.g., an electrocardiogram) are incorporated in the prediction rule in those settings.