Conventional small-bowel capsule endoscopy reading vs proprietary artificial intelligence auxiliary systems: Systematic review and meta-analysis.

Abstract

Background and study aims: Small-bowel capsule endoscopy (SBCE) is the gold standard for diagnosing small bowel (SB) pathologies, but its time-consuming nature and potential for human error make it challenging. Several proprietary artificial intelligence (AI) auxiliary systems based on convolutional neural networks (CNNs) that are integrated into SBCE reading platforms are available on the market and offer the opportunity to improve lesion detection and reduce reading times. This meta-analysis aimed to evaluate performance of proprietary AI auxiliary platforms in SBCE compared with conventional, human-only reading.

Methods: A systematic literature search was conducted to identify studies comparing AI-assisted SBCE readings with conventional readings by gastroenterologists. Performance measures such as accuracy, sensitivity, specificity, and reading times were extracted and analyzed. Methodological transparency was assessed using the Methodological Index for Non-randomized Studies (MINORS) assessment tool.

Results: Of 669 identified studies, 104 met the inclusion criteria and six were included in the analysis. Quality assessment revealed high methodological transparency for all included studies. Pooled analysis showed that AI-assisted reading achieved significantly higher sensitivity and comparable specificity to conventional reading, with a higher log diagnostic odds ratio and no substantial heterogeneity. In addition, AI integration substantially reduced reading times, with a mean decrease of 12-fold compared with conventional reading.

Conclusions: AI-assisted SBCE reading outperforms conventional human review in terms of detection accuracy and sensitivity, remarkably reducing reading times. AI in this setting could be a game-changer in reducing endoscopy service workload and supporting novice reader training.