Regression tree for choledocholithiasis prediction

Artificial intelligence in the detection of choledocholithiasis: a systematic review

IMPORTANCE

Choledocholithiasis is a potentially life-threatening manifestation of acute biliary dysfunction (ABD) often requiring magnetic resonance cholangiopancreatography (MRCP) for diagnosis when standard investigation findings are inconclusive. Machine learning models (MLMs) may offer alternatives to diagnose choledocholithiasis.

OBJECTIVE

This systematic review seeks to evaluate the performance of MLMs in predicting choledocholithiasis and to compare this performance with the American Society of Gastrointestinal Endoscopy (ASGE) guidelines.

REVIEW

This review adhered to PRISMA guidelines. Four databases were searched for relevant records published between January 2000 and April 2024. Two researchers appraised records. MLM performance and ASGE guideline efficacy were compared, and the clinical utility of MLMs was assessed.

FINDINGS

408 records were screened; eight were eligible. Model accuracy ranged from 19 % to 97 %. Several records demonstrated a moderate-to-high risk of bias; of those featuring low risk of bias, peak accuracies ranged from 70 % to 85 %. Most MLMs outperformed ASGE guidelines. Important predictor variables included age, total bilirubin, and common bile duct diameter.

CONCLUSIONS

MLMs outperform ASGE guidelines in predicting choledocholithiasis. Nonetheless, biases in study design and reporting limit their prospective applicability. Current MLMs do not yet rival MRCP in detecting choledocholithiasis. Future guideline development should consider MLM-driven insights for better risk prediction.