Diagnostic accuracy of updated risk assessment criteria and development of novel computational prediction models for patients with suspected choledocholithiasis

Abstract

Background: There are risks of choledocholithiasis in symptomatic gallstones, and some surgeons have proposed the identification of choledocholithiasis before cholecystectomy. Our goal was to evaluate the diagnostic accuracy of the latest guidelines and create computational prediction models for the accurate prediction of choledocholithiasis.

Methods: We retrospectively reviewed symptomatic gallstone patients hospitalized with suspected choledocholithiasis. The diagnostic performance of 2019 and 2010 guidelines of the American Society for Gastrointestinal Endoscopy (ASGE) and 2019 guideline of the European Society of Gastrointestinal Endoscopy (ESGE) in different risks. Lastly, we developed novel prediction models based on the preoperative predictors.

Results: A total of 1199 patients were identified and 681 (56.8%) had concurrent choledocholithiasis and were included in the analysis. The specificity of the 2019 ASGE, 2010 ASGE, and 2019 ESGE high-risk criteria was 85.91%, 72.2%, and 88.42%, respectively, and their positive predictive values were 85.5%, 77.4%, and 87.3%, respectively. For Mid-risk patients who followed 2019 ASGE about 61.8% of them did not have CBD stones in our study. On the choice of surgical procedure, laparoscopic cholecystectomy + laparoscopic transcystic common bile duct exploration can be considered the optimal treatment choice for cholecysto-choledocholithiasis instead of Endoscopic Retrograde Cholangio-Pancreatography (ERCP). We build seven machine learning models and an AI diagnosis prediction model (ModelArts). The area under the receiver operating curve of the machine learning models was from 0.77 to 0.81. ModelArts AI model showed predictive accuracy of 0.97, recall of 0.97, precision of 0.971, and F1 score of 0.97, surpassing any other available methods.

Conclusion: The 2019 ASGE guideline and 2019 ESGE guideline have demonstrated higher specificity and positive predictive value for high-risk criteria compared to the 2010 ASGE guideline. The excellent diagnostic performance of the new artificial intelligence prediction model may make it a better choice than traditional guidelines for managing patients with suspected choledocholithiasis in future.

Keywords: Artificial intelligence; Machine learning; Suspected choledocholithiasis; The American Society for Gastrointestinal Endoscopy; The European Society of Gastrointestinal Endoscopy.