Rapid development of accurate artificial intelligence scoring for colitis disease activity using applied data science...

Area of Science:

• Artificial intelligence in medical imaging
• Gastroenterology and endoscopy
• Machine learning for disease scoring

Background:

• Accurate endoscopic scoring of colitis activity is crucial for clinical and research applications.
• Previous AI models struggled with detailed scoring, such as subscores for the Mayo Endoscopic Score (MES) or ulcerative colitis endoscopic index of severity (UCEIS).
• Developing AI for colitis scoring often requires large datasets, limiting its accessibility.

Purpose of the Study:

• To develop and evaluate a machine-learning algorithm (MLA) for scoring ulcerative colitis endoscopic index of severity (UCEIS).
• To assess the accuracy and agreement of the MLA with human expert consensus.
• To demonstrate the feasibility of achieving high AI performance with a relatively smaller dataset.

Main Methods:

• A multi-task learning framework was employed for frame-by-frame analysis of endoscopic videos.
• The machine-learning algorithm (MLA) was trained on 38,124 frames from 73 patients with biopsy-proven ulcerative colitis.
• Performance was evaluated by comparing MLA-generated UCEIS scores against consensus scores from three independent human reviewers.

Main Results:

• The MLA achieved high accuracy in differentiating normal mucosa from active inflammation (UCEIS 0 vs ≥1; accuracy 0.90, κ=0.90).
• Excellent accuracy was observed in distinguishing mild from moderate-severe inflammation (UCEIS 0-3 vs ≥4; accuracy 0.98, κ=0.96).
• High agreement was found for the total UCEIS score (κ=0.92) and its subdomains (vascular pattern κ=0.80, bleeding κ=0.83, erosions κ=0.88).

Conclusions:

• Modified data science techniques enable AI to achieve high accuracy and agreement with human reviewers in colitis endoscopic scoring, even with smaller datasets.
• The developed AI demonstrates near-perfect performance in certain differentiation tasks, validating its potential clinical utility.
• This approach offers a pathway for faster AI development in endoscopic scoring and other medical imaging tasks.