Background: Verification bias occurs when test findings influence the decision to perform a gold standard test. It adversely influences diagnostic test accuracy by inflating sensitivity and deflating specificity.Westudied the impact of verification bias on the estimated accuracy of a test commonly used in suspected lung cancer. Methods: We studied 534 consecutive patients referred for [18F]fluorodeoxyglucose uptake by positron emission tomography (PET). Primary outcomes were tissue diagnoses of cancer and of mediastinal lymph node metastases. A secondary outcome was 3-year mortality. We accounted for verification bias using 2 validated methods. Results: The gold standard test, namely tissue acquisition, was performed in 419 patients (78%); mediastinal lymph node sampling occurred in 301 (56%). While the 410 patients with PET-diagnosed stage I cancer or higher were more likely than patients with negative PET scan findings to undergo tissue diagnosis testing (92% vs 34%) (P<.001), there was no association between PET findings and performance of mediastinal sampling. Without accounting for verification bias, the sensitivity and specificity of PET for diagnosis of cancer were 0.95 (95% confidence interval [CI], 0.92-0.97) and 0.31 (95% CI, 0.21-0.42), respectively. After adjustment, sensitivity fell to 0.85 (95% CI, 0.81-0.89), while specificity increased to 0.51 (95% CI, 0.40-0.60). For diagnosis of mediastinal disease, verification bias had slight effects on test accuracy. There were 224 deaths, with a strong gradient between PET stage and death (P<.001). Conclusion: The diagnostic accuracy of PET for assessment of suspected lung cancer is substantially affected by verification bias.
ASJC Scopus subject areas
- Internal Medicine