Feature‐Based Patterns Associated With Japan NBI Expert Team Type 1 Classification of Serrated Colorectal Lesions: A Post Hoc Decision Tree Analysis
Аннотация
ABSTRACT Objectives Japan NBI Expert Team (JNET) Type 1 classification is used for optical characterization of colorectal lesions with magnifying NBI, but reported Type 1‐related findings may differ between readers when assessing serrated lesions. This small‐scale post hoc reader‐based study explored reported feature‐classification patterns associated with JNET Type 1 selection among serrated colorectal lesions and compared them between expert and non‐expert endoscopists. Methods We analyzed de‐identified data from two web‐based image‐interpretation studies using the same colorectal lesion image dataset. Thirteen histologically confirmed serrated lesions (11 sessile serrated lesions [SSLs] and two hyperplastic polyps [HPs]) assessed by 27 expert and 49 non‐expert Japanese endoscopists were included. Simplified JNET responses were dichotomized as Type 1 versus non‐Type 1. Classification and regression tree analysis was used to visualize associations between this outcome and predefined features, including Type 1‐supportive findings and an anti‐Type 1 composite variable. Variability was assessed using entropy. Results In the expert and non‐expert core models, the root splits were anti‐Type 1 findings and regular dark or white spots, respectively. These root splits were stable in leave‐one‐reader‐out analyses, but the non‐expert root split changed to anti‐Type 1 findings after excluding unassessable responses. Median entropy was lower among experts than non‐experts (0.229 vs. 0.954). Conclusions This exploratory post hoc decision tree analysis suggested relatively stable expert‐associated feature‐classification patterns, whereas non‐expert patterns were influenced by the treatment of unassessable responses. These findings reflect exploratory associations in Type 1 assignment, not cognitive decision sequences, SSL–HP differentiation, or a validated diagnostic algorithm.