Interobserver agreement and accuracy of preoperative endoscopic ultrasound-guided biopsy for histological grading of pancreatic cancer

Background and study aim: Poorly differentiated/high grade pancreatic ductal adenocarcinoma (PDAC) is associated with an early unfavorable outcome, and patients with these tumors may be candidates for neo-adjuvant treatment. Endoscopic ultrasound-guided pancreatic fine-needle biopsy (EUS-FNB) may, in theory, allow preoperative assessment of PDAC histological grading. The aim of the current study was to assess the interobserver agreement and accuracy of preoperative PDAC grading from EUS-FNB specimens. Methods: Data from 42 postsurgical PDAC patients who had undergone preoperative EUS-FNB were retrieved. Four experienced pathologists independently reviewed the EUS-FNB slides and reported tumor grading (well, moderately, or poorly differentiated). Agreement among pathologists for grading of preoperative EUS-FNB samples was expressed by using Cohen's or Fleiss' kappa statistic, as appropriate. Postsurgical PDAC grading was used as the gold standard to assess the cumulative accuracy of EUS-FNB for the preoperative prediction of PDAC grading. Results: The kappa values for PDAC grading on EUS-FNB specimens ranged from 0.09 to 0.41. The total agreement among the four pathologists was only fair (κ = 0.27; 95 % confidence interval [CI] 0.14 - 0.38). When tumor grades were grouped as well or moderately differentiated vs. poorly differentiated, kappa values ranged from 0.19 to 0.50, with only a fair overall agreement (κ = 0.27; 95 %CI 0.21 - 0.49). The accuracy of preoperative grading from EUS-FNB was 56 % (75/134 readings; 95 %CI 40 % - 65 %), with mean sensitivity and specificity to detect a high grade, poorly differentiated tumor of 41 % (95 %CI 19 % - 54 %) and 78 % (53/68 readings; 95 %CI 60 % - 99 %), respectively. Conclusions: Preoperative EUS-FNB-based histological grading of PDAC is unreliable, and current results do not support the use of this information in clinical practice. This appears to be due to suboptimal interobserver agreement among pathologists and an overall low accuracy in predicting postsurgical grading.