Machine Learning Creates a Simple Endoscopic Classification System that Improves Dysplasia Detection in Barrett's...

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Medical Diagnosis
• Gastrointestinal Endoscopy

Background:

• Barrett's oesophagus (BE) is a precursor to oesophageal adenocarcinoma (OAC), necessitating endoscopic surveillance for dysplasia.
• Current surveillance endoscopy lacks standardized guidelines for performer competence.
• Machine learning (ML), specifically decision trees (DTs), offers a novel approach to codify expert knowledge.

Purpose of the Study:

• To hypothesize that a DT derived from expert endoscopists can improve dysplasia detection rates in non-expert endoscopists.
• To evaluate the efficacy of ML-generated decision rules for dysplasia prediction in BE.
• To assess the impact of standardized training using ML algorithms on endoscopist competence.

Main Methods:

• Collected high-definition endoscopy videos of non-dysplastic (ND-BE) and dysplastic (D-BE) Barrett's oesophagus.
• Expert endoscopists interpreted videos, and data were used to construct a ML decision tree (DT).
• Non-expert endoscopists (trainees, students) scored videos before and after web-based training with the DT.

Main Results:

• The ML decision tree achieved 92% accuracy, with 97% sensitivity and 88% specificity for dysplasia prediction.
• Web-based training using the DT significantly improved dysplasia detection accuracy in both trainees and students.
• Trainee sensitivity increased from 71% to 83%, and student specificity rose from 31% to 49% post-training.

Conclusions:

• ML can effectively generate simple algorithms from expert opinion to accurately predict dysplasia in BE.
• Training non-experts with these ML-derived algorithms significantly enhances their dysplasia detection capabilities.
• This approach facilitates standardized training, competence assessment, and potentially shortens the learning curve for endoscopy in BE.