Interinstitutional variation in predictive value of the ThyroSeq v2 genomic classifier for cytologically indeterminate thyroid nodules

Abstract Background The ThyroSeq v2 next-generation sequencing assay estimates the probability of malignancy in indeterminate thyroid nodules. Its diagnostic accuracy in different practice settings and patient populations is not well understood. Methods We analyzed 273 Bethesda III/IV indeterminate thyroid nodules evaluated with ThyroSeq at 4 institutions: 2 comprehensive cancer centers ( n  = 98 and 102), a multicenter health care system ( n  = 60), and an academic medical center ( n  = 13). The positive and negative predictive values of ThyroSeq and distribution of final pathologic diagnoses were analyzed and compared with values predicted by Bayes theorem. Results Across 4 institutions, the positive predictive value was 35% (22%–43%) and negative predictive value was 93% (88%–100%). Predictive values correlated closely with Bayes theorem estimates ( r 2  = 0.84), although positive predictive values were lower than expected. RAS mutations were the most common molecular alteration. Among 84 RAS -mutated nodules, malignancy risk was variable (25%, range 10%–37%) and distribution of benign diagnoses differed across institutions (adenoma/hyperplasia 12%–85%, noninvasive follicular thyroid neoplasm with papillary-like nuclear features 5%–46%). Conclusion In a multi-institutional analysis, ThyroSeq positive predictive values were variable and lower than expected. This is attributable to differences in the prevalence of malignancy and variability in pathologist interpretations of noninvasive tumors. It is important that clinicians understand ThyroSeq performance in their practice setting when evaluating these results.