Reply to Chen et al.

We thank Chen et al. for their thoughtful insights into our study, in which we developed a machine learning-based choledocholithiasis prediction model. We share their excitement about the implementation of machine learning in gastroenterology.

To reiterate, we tested three versions of our model, one with the laboratory test values from the initial presentation, a second model that incorporated initial laboratory test values and follow-up test values to see whether a second set of test results would improve prediction, and a third model that incorporated the difference between the first and second set of test values. The two-value test model only slightly improved the model performance, and given the need for several extra inputs, we felt the single laboratory test model was the more parsimonious equivalent.

Regarding the question of handling missing data in the two-lab value model, this was handled by the mean imputation method.

Regarding the question of hepatoprotective measures that would affect follow-up lab values, this is a good point and might explain why an additional set of lab values did not substantially improve model predictivity, though this kind of data heterogeneity is one of the many challenges of clinical model creation.

Given the limited spectrum of this algorithm and the preliminary nature of the study, we deferred the calibration component for now and plan to do the calibration in the larger definitive study with multiple decision components (surgery, endoscopic ultrasound, magnetic resonance cholangiopancreatography, endoscopic retrograde cholangiopancreatography, etc).

Finally, we will be including the source of the model and the updates made during the larger definitive study in the GitHub public repository. We are excited that you wish to prospectively validate our model and we hope this aids in your work.

Publication History

Article published online:
27 June 2024