The role of AI models in enhancing choledocholithiasis diagnosis: A systematic review and meta-analysis

Accurate choledocholithiasis (CDL) diagnosis is essential to avoid delayed treatment, prevent complications and reduce unnecessary interventions. Traditional guidelines aid in risk stratification but may lack precision. Notably, artificial intelligence (AI) and machine learning (ML) offer innovative tools that may enhance the accuracy and timeliness of CDL prediction. The present study aimed to systematically evaluate the diagnostic performance of AI-assisted tools in predicting CDL and to compare it to traditional guideline-based methods. A comprehensive search was conducted in MEDLINE, EMBASE, PubMed and Web of Science, identifying 578 studies. After screening and application of the inclusion criteria, 11 studies were analyzed. A bivariate random-effects model was used to pool sensitivity, specificity and positive likelihood ratios (LR+). Summary receiver operating characteristic (SROC) curves were also generated. Meta-analysis showed an overall high pool sensitivity and specificity of AI-assisted models: 83.2% [95% confidence interval (CI): 68.9; 91.8] and 91.1% [95% CI: 84.7; 95.0], respectively. The LR+ from the common effect model was 8.39 [95% CI: 7.4; 9.5], suggesting that AI models have a moderately strong ability to predict CDL. AI models demonstrated higher diagnostic performance than traditional American Society for Gastrointestinal Endoscopy guidelines, as evidenced by SROC comparisons. In conclusion, AI-assisted tools show promise in enhancing CDL diagnosis through high sensitivity and specificity. Innovative AI and ML tools may serve as predictive tools and therapeutic decision-support systems deserving further clinical validation.