SSAT - COMMON BILE DUCT STONES: A MACHINE LEARNING MODEL ON CYSTIC DUCT DIAMETER AS A PREDICTOR OF CHOLEDOCHOLITHIASIS

Back to 2025 Posters

COMMON BILE DUCT STONES: A MACHINE LEARNING MODEL ON CYSTIC DUCT DIAMETER AS A PREDICTOR OF CHOLEDOCHOLITHIASIS
Amir H. Sohail^*¹, Hazim Hakmi², Chinyere Okpara², Megan Winner², David Halpern²
¹University of New Mexico Health Sciences Center, Albuquerque, NM; ²New York University, New York, NY

Introduction:
Cholecystectomy is a common surgical procedure for gallbladder-related conditions, but the detection of common bile duct (CBD) stones during cholecystectomy remains a challenge, especially without intra-operative cholangiography (IOC). Presently, there are no pre-operative tools to reliably identify patients at high-risk for CBD stones undergoing cholecystectomy.

Methods:
We conducted a retrospective study on patients undergoing cholecystectomy between 2011 and 2022. The presence of retained common bile duct (CBD) stones was confirmed by endoscopic retrograde cholangiopancreatography (ERCP).

Several machine learning models, including ensemble methods, were evaluated to predict CBD stone occurrence. Predictive algorithms were trained on pre- and perioperative factors such, as cystic duct diameter size (abstracted from surgical pathology), liver function panel, laboratory work-up, comorbids, lifestyle factors. Prediction performance was assessed using recall and precision, given the relatively low incidence of CBD stones. Then, a decision tree surrogate model was fit to interpret the best-performing machine learning model and establish a framework for clinical decision-making in CBD stone screening.

Results:
Over the 12 year study period, out of the 3335 cholecystectomy patients, 278 (8.3%) required ERCP for retained CBD stones. Patients with retained CBD stones had a significantly larger intraoperative cystic duct diameter (0.35mm vs. 0.3mm, p<0.01). However, the strongest indicator of retained stones was elevations in liver function markers including AST, ALT, ALP, and serum bilirubin (p<0.001) (Figure 1).

With an overall recall of 0.84 and a precision of 0.24, our final three-class prediction classified patients into high risk (indicating an increased need for procedural intervention), low risk (no intervention), and elevated risk (unclear need for procedural intervation) (Table 1).

Conclusion:
The results of our prediction model establish a robust framework for improving the detection of CBD stones prior to surgery, outperforming previous methods. While research is needed with a larger population to determine precise cutoff values, our findings suggest that easily obtainable laboratory values can effectively predict the presence of CBD stones preoperatively. This potential for scalability supports the feasibility of future multi-site studies.

Back to 2025 Posters