A ghost in the machine: is machine learning necessary for prediction of choledocholithiasis?

Endoscopic retrograde cholangiopancreatography (ERCP) remains the standard of care for management of biliary lithiasis; however, ERCP carries a substantial risk of post-procedure pancreatitis, warranting its judicious use. Several manuscripts over the years have attempted to create predictive models for the presence of choledocholithiasis, all of which had certain limitations. With the continued push for technological advancement, more recent attempts have utilized advanced modeling or artificial intelligence, though these have also been unable to gain a substantial clinical foothold. At present, guidelines from American Society for Gastrointestinal Endoscopy (ASGE) and European Society of Gastrointestinal Endoscopy (ESGE) remain widely utilized to guide management for patients with suspected common bile duct (CBD) stones.

“The authors have aimed to mitigate the risk of unnecessary ERCP, a praiseworthy goal and welcome concept by the advanced endoscopy community; however, whether this task has been accomplished remains unclear.”

In this issue of Endoscopy, Steinway et al. [11] present data using a gradient boosting machine learning model to predict the presence of CBD stones in patients with suspected choledocholithiasis. The goal of this research was to improve upon current stratification methods in patients with suspected choledocholithiasis to mitigate the risks associated with unnecessary ERCP procedures. The model parameters were developed from two previously published cohort studies with patient level data from nearly 1400 unique patients, ~60% of whom had choledocholithiasis [22] [33]. Eight variables (presence of CBD stone on ultrasound, index total bilirubin, index alkaline phosphatase, concomitant acute pancreatitis, age, index alanine transaminase, index aspartate transaminase, and CBD diameter >6 mm on ultrasound) were utilized in the final model. The variables were then input into an “app-based” program that provided the likelihood of choledocholithiasis in any given patient. In the study, the model outperformed both ASGE and ESGE guidelines with regard to diagnostic accuracy (71.5% vs. 62.4% vs. 62.8%, respectively).

The authors have aimed to mitigate the risk of unnecessary ERCP, a praiseworthy goal and welcome concept by the advanced endoscopy community; however, whether this task has been accomplished remains unclear. The superior diagnostic accuracy is a relatively marginal improvement on existing consensus guidelines, with several concerns noted. The question of ERCP is not always “A. Yes” or “B. No”; it may be “C. Need more information.” The inability to account for the moderate-risk cohort who would undergo endoscopic ultrasound or magnetic resonance cholangiopancreatography (MRCP) prior to determining the utility of ERCP is a substantial study limitation that renders the results difficult to interpret. Additionally, the authors made assumptions that may not reflect generalizable clinical practice. The equivalence of a CBD stone and biliary sludge “raises an eyebrow.” The authors cite data suggesting that spontaneous stone passage occurs in over 50% of patients, which is likely to relate to smaller CBD stones [44]. With this concept in mind, it would seem that biliary sludge alone is likely to pass at a much higher rate and less likely to become obstructive. We recognize that the authors constructed the model with two large cohort datasets where biliary sludge and stones were deemed equivalent; however, future iterations of the model may want to reconsider this concept. Finally, the use of CBD diameter of 6 mm as a cutoff to determine CBD dilation is likely to be unreliable and inaccurate. Recent data from Beyer et al. [55] using a population-based, cross-sectional cohort study based on MRCP demonstrated that a normal upper reference limit should be 8 mm in patients aged <65 years and 11 mm in patients ≥65 years. These findings are concurrent with what is used in clinical practice and may improve overall modeling.

While the authors’ current iteration may fall short, they should be commended on their efforts thus far. It is expected that future models would take into consideration noninvasive imaging for moderate-risk patients, as well as fine-tuning the model parameters. Additionally, the model should be externally validated on a dataset with patient level data as opposed to validation via modeling techniques. If ultimately successful, this machine learning platform could be a monumental step forward in risk stratification related to ERCP for patients with suspected choledocholithiasis.

Publication History

Article published online:
23 January 2024

• References

• 1 Steinway SN, Tang B, Telezing J. et al. A machine learning-based choledocholithiasis prediction tool to improve ERCP decision making: a proof-of-concept study. Endoscopy 2024; 56 DOI: 10.1055/a-2174-0534. (PMID: 37699524)
• 2 Yu CY, Roth N, Jani N. et al. Dynamic liver test patterns do not predict bile duct stones. Surg Endosc 2019; 33: 3300-3313 DOI: 10.1007/s00464-018-06620-x. (PMID: 30911921)
• 3 Suarez AL, LaBarre NT, Cotton PB. et al. An assessment of existing risk stratification guidelines for the evaluation of patients with suspected choledocholithiasis. Surg Endosc 2016; 30: 4613-4618
• 4 Tranter SE, Thompson MH. Spontaneous passage of bile duct stones: frequency of occurrence and relation to clinical presentation. Ann R Coll Surg Engl 2003; 85: 174-177 DOI: 10.1308/003588403321661325. (PMID: 12831489)
• 5 Beyer G, Kasprowicz F, Hanneman A. et al. Definition of age-dependent reference values for the diameter of the common bile duct and pancreatic duct on MRCP: a population-based, cross-sectional cohort study. Gut 2023; 72: 1738-1744