Interobserver agreement of the modified Paris classification and histology prediction of colorectal lesions in patients with inflammatory bowel disease

Background and Aims: SCENIC (International Consensus Statement on Surveillance and Management of Dysplasia in IBD) guidelines recommend that visible dysplasia in patients with longstanding inflammatory bowel disease (IBD) should be endoscopically characterized using a modified Paris classification. This study aimed to determine the interobserver agreement (IOA) of the modified Paris classification and endoscopists’ accuracy for pathology prediction of IBD visible lesions. Methods: One hundred deidentified endoscopic still images and 30 videos of IBD visible colorectal lesions were graded by 10 senior and 4 trainee endoscopists from 5 tertiary care centers. Endoscopists were asked to assign 4 classifications for each image: the standard Paris classification, modified Paris classification, pathology prediction, and lesion border. Agreement was measured using Light's kappa coefficient. Consensus of ratings was assessed according to strict majority. Results: The overall Light's kappa for all study endpoints was between .32 and .49. In a subgroup analysis between junior and senior endoscopists, Light's kappa continued to be less than .6 with a slightly higher agreement among juniors. Lesions with the lowest agreement and no consensus were mostly classified as Is, IIa, and mixed Paris classification and sessile and superficial elevated for modified Paris classification. Endoscopist accuracy for prediction of dysplastic, nondysplastic, and serrated pathology was 77%, 56%, and 30%, respectively. There was a strong association (P < .001) between the given morphology classification and the predicted pathology with Ip lesions carrying a much lower expectation of dysplasia than Is/IIc/III and mixed lesions. The agreement for border prediction was .5 for junior and .3 for senior endoscopists. Conclusions: This study demonstrates very low IOA for Paris and modified Paris classifications and low accuracy and IOA for lesion histopathology prediction. Revisions of these classifications are required to create a clinically useful risk stratification tool and enable eventual application of augmented intelligence tools.