New diagnostic prediction model with individualized absolute probability score of PE presence

Introduction and methods

Background

Pulmonary embolism (PE) can be confirmed or ruled out with imaging tests such as CT pulmonary angiography or ventilation-perfusion scanning. However, these tests have several disadvantages such as radiation exposure, risk of contrast reactions or nephropathy, and high costs. Diagnostic algorithms, which utilize medical history and physical examination findings with D-dimer testing, can be used to exclude PE in patients with suspected PE, and thereby greatly reduce the need for imaging tests [1]. Nevertheless, a high proportion of patients with suspected PE with non-low clinical probability and elevated D-dimer are referred for imaging, and PE is not diagnosed in most of these patients [2-5], indicating that better risk stratification methods are needed.

Aim of the study

The aim of the study was to develop a clinical prediction model that accurately calculates an absolute PE probability for each patient using readily available clinical items and D-dimer levels.

-

Methods

An individual patient dataset was generated using data of studies published in MEDLINE from January 1995 until January 2021 [6]. Eligible studies had a prospective or cross-sectional design, structured clinical pretest probability, measured D-dimer levels, and performed imaging at baseline in all patients or clinical follow-up of at least 30 days in patients not undergoing imaging at baseline. A total of 16 studies with 28,305 unique patients with suspected PE were included (n=23,899 patients without PE, and n=4406 patients with PE). Patients from various clinical settings were enrolled, including from self-referral emergency care, primary care, secondary care, and hospitalized or nursing homes. The following variables were considered in the new model: age (in years), sex, previous venous thromboembolism (VTE), recent surgery or immobilization, hemoptysis, cancer, clinical signs of DVT, tachycardia, inpatient status, D-dimer level (in µg/L), and interaction between age and D-dimer. Subjective variables such as ‘PE is the most likely diagnosis’ or unstructured PE probability estimates were not considered. Variables with P-value of > 0.10 in more than half of the imputed datasets were excluded from the model. The performance of the model was assessed with internal-external cross validation on study level.

Efficiency was defined as the number of patients in whom PE can be considered excluded based on the diagnostic model without imaging relative to all patients with suspected PE. Safety was defined by as the number of patients in whom PE is present relative to patients in whom PE was considered excluded by the model (false negative).

Outcomes

The outcome was a diagnosis of PE confirmed by imaging at baseline or VTE during 30 to 90 day follow-up. Fatal PE was also considered as outcome.

Main results

Model development and performance

• All candidate predictors were included in the model, with the exception of tachycardia.
• Discrimination performance of the model was good across all validation studies (pooled c-statistic: 0.87; 95%CI: 0.85-0.89; prediction intervals [PI]: 0.77-0.93).
• Calibration performance of the model was excellent (pooled outcome:expected ratio: 0.99; 95%CI: 0.87-1.14), but there was some heterogeneity across studies (PI: 0.55-1.79).
• There was an overall good agreement between the estimated probabilities of the model and observed prevalence of PE. However, the model overestimated the probability of PE by about 1% in the clinically relevant range of probabilities of 0%-3%.

Comparisons with current algorithms

• In a dataset of 14 studies, the new model scored better on performance (pooled c-statistic: 0.87; 95%CI: 0.84-0.89; PI: 0.76-0.93) compared with the Wells score with age-adjusted D-dimer testing (pooled c-statistic: 0.73; 95%CI: 0.70-0.75; PI: 0.62-0.82) and the Wells score with three-tier D-dimer testing (pooled c-statistic: 0.79; 95%CI: 0.76-0.81; PI: 0.66-0.88).
• Similar to the new model, the Well score models also overestimated the probably of PE in the lower clinically relevant range of probabilities (between 1%-10%).
• The efficiency and safety of the new model was comparable to currently used algorithms.

Mean vs. individual probability estimates

• The new model identified a high proportion of patients with an estimated individual probability of PE ≥2% in patients who were classified as ‘imaging not indicated’ by current diagnostic algorithms. This occurred in 28% of the patients when PE was excluded using the Wells score with age-adjusted D-dimer testing and in 44% of the patients when PE was excluded using the Wells score with a threshold based on clinical pretest probability.

Conclusion

References

Find this article online at Eur Heart J.