Is AI ready to replace protocol guided biopsies in Barrett's surveillance? The first real-world experience

Introduction

We aim to investigate the real-world value of AI during Barrett’s surveillance in view of the recent ESGE position statement on the expected value of AI in endoscopy.

Methods

The development and validation of the regulatory-approved AI system in this study was described in a recent peer-reviewed publication by our group. The study was conducted at a single tertiary centre for Barrett’s neoplasia endotherapy. Statistical powering was performed to estimate the number of missed neoplasia by AI compared to Seattle protocol biopsies assuming 40% prevalence of neopalsia (based on our enriched population’s local data) and 10% miss rates by AI (based on pre-clinical validation data) using 95% confidence level and +/-5% precision level. Ground truth was expert endoscopist assessment and histology.

Results

A total of 231 consecutive patients, including 92 patients with Barrett’s neoplasia, were included. Histology of neoplastic lesions showed adenocarcinoma, HGD and LGD in 57.1%, 35.7%, and 7.2% of patients respectively. In the per-patient analysis, the sensitivity, specificity and NPV of AI-assisted neoplasia detection was 89.3%, 72.8% and 91.06% respectively. Neoplasia miss rate by AI compared to Seattle protocol biopsies was 10.7%, however the mean number of Seattle protocol biopsies and AI-targeted biopsies was 8.16 and 0.81 respectively.

Conclusion

This is the first real-world experience demonstrating the potential value of AI-assisted targeted biopsies in Barrett’s neoplasia surveillance. The specificity of AI neoplasia detection is less compared to previously published pre-clinical studies, highlighting the need to address the issue of false positive predictions by AI. This data needs validating in a multi-centre design given high prevalence of neoplasia in this enriched tertiary setting.