Probabilistic reporting of EUS‐FNA cytology

BACKGROUND The objectives of this study were to determine threshold probabilities needed to perform endoscopic ultrasound‐guided fine‐needle aspiration (EUS‐FNA) and those needed to treat patients suspected of having malignancy and then to compare these thresholds to the pre‐ and posttest probabilities of malignancy associated with benign, atypical, suspicious, and malignant diagnoses. The goal was to aid endoscopists in making appropriate clinical decisions based on both quantitative and qualitative approaches. METHODS The study included 633 consecutive patients. A decision tree was constructed to estimate the “treatment” threshold. Using treatment threshold and likelihood ratios, the authors determined the “no‐test‐test” and “test‐treatment” thresholds. Pretest probability was compared with no‐test‐test and test‐treatment thresholds, and the post‐EUS‐FNA probability of malignancy for each diagnostic category with the treatment threshold. Results were stratified by lesion site, lesion size, and cytopathologist. RESULTS EUS‐FNA has a wide range of pretest probabilities within which it could be performed (0.06–0.98). The posttest probabilities for malignancy, 0.99 (95% confidence interval [CI], 0.967–0.996) and 0.09 (95% CI, 0.057–0.126), after a positive or a negative result, respectively, were significantly different from the treatment threshold but not those of suspicious, 0.92 (95% CI, 0.767–0.994) diagnosis. The posttest probability of atypical diagnosis, 0.60 (95% CI, 0.407–0.772), was not significantly different from that of pretest probability. Results did not vary by lesion size, organ site, or cytopathologist. CONCLUSION The authors demonstrated the uncertainty associated with EUS‐FNA diagnostic categories and used the threshold approach to qualify quantitatively the decision to perform EUS‐FNA and the decision to treat patients suspected of having malignancy. Cancer (Cancer Cytopathol) 2006. © 2006 American Cancer Society. This article presents the first study of its kind to assess probabilistic reporting and threshold modeling of clinical decisions in cytodiagnosis.