Introduction Barrett’s Oesophagus (BE) is the pre-cursor to oesophageal adenocarcinoma. Endoscopic surveillance is performed to detect dysplasia in BE as it is likely to be treatable. Machine Learning (ML) is a technology that generates simple rules, known as a Decision Tree (DT). Using a DT generated from Expert Endoscopists (EE), we hypothesised that this could be used to improve dysplasia detection in Non-Expert Endoscopists (NEE).
Methods Endoscopic videos of Non-Dysplastic (ND-BE) and Dysplastic (D-BE) BE were recorded. Areas of interest were biopsied. Videos were shown to 3 EE (blinded) who interpreted mucosal & vascular patterns, presence of nodularity/ulceration & suspected diagnosis. Acetic Acid (ACA) was sometimes used. EE answers were inputted into the WEKA package to identify the most important attributes and generate a DT to predict dysplasia. NEE (GI registrars and medical students) scored these videos online before & after online training using the DT (Fig 1). Outcomes were calculated before & after training. Student’s t-test was used (p < 0.05).
Results Videos from 40 patients (11 pre/post ACA) were collected (23 ND-BE, 17 D-BE). EE mean accuracy of dysplasia prediction was 96% using the DT. Mean sensitivity/specificty were 93%/99%. Neither vascular pattern nor ACA improved dysplasia detection. Students had a high sensitivity but poor specificity as they ‘overcalled’ normal areas. GI registrars did the opposite. Training significantly improved sensitivity of dysplasia detection amongst registrars without loss of specificity. (Table 1). Specificity rose in students without loss of sensitivity and significant improvement in overall detection.
Conclusion ML can generate a simple algorithm from EE to accurately predict dysplasia. Once taught to NEE, it yields a significantly higher rate of dysplasia detection. This opens the door to standardised training and assessment of competence in those performing endoscopy in BE.
Disclosure of Interest None Declared