Development and validation of a minimally invasive diagnostic model for biliary atresia using artificial intelligence

Comprehensive Summary

The study evaluates whether an AI based diagnostic model that combines ultrasound imaging with serum MMP-7 can accurately differentiate biliary atresia from other causes of infantile cholestasis. The researchers carried out a multicenter diagnostic study that included a retrospective training cohort and a prospective validation cohort across six medical centers in China. Data from 348 infants with obstructive jaundice, including ultrasound images and serum MMP-7 levels, were used to develop and validate a logistic regression based AI model, and its performance was compared with an AI model based on ultrasound alone and biomarker testing alone. The results showed that the combined AI model achieved very high diagnostic accuracy, with strong sensitivity and specificity in both the training and validation cohorts. The combined model consistently outperformed ultrasound alone and serum MMP-7 testing alone, with an AUROC of 0.985 and 0.949 in the training and validation cohorts respectively versus 0.945 and 0.909 in the ultrasound only group and 0.916 and 0.907 in the serum MMP-7 group. Performance remained stable across participating centers, supporting the generalizability of the model. In the discussion, the authors note that the study was limited by a significantly different prevalence in biliary atresia in the training and validation cohorts. Even so, they explain that integrating objective biomarker data with AI supported ultrasound analysis improves diagnostic reliability while reducing reliance on operator experience. They also highlight the successful development of a mobile application, which supports the potential for real world clinical use.

Outcomes and Implications

This research is important because early and accurate diagnosis of biliary atresia is critical for timely surgical intervention, which directly affects long term outcomes in affected infants. Current diagnostic approaches can be invasive or highly dependent on specialist expertise, leading to delays in care. Clinically, this AI based model could support noninvasive early screening by improving diagnostic confidence using routinely available ultrasound images and a serum biomarker. The findings suggest that this approach could be integrated into clinical workflows, including mobile based decision support tools, to assist clinicians in identifying high risk infants and supporting faster referral for surgery. While further implementation and broader access to MMP 7 testing are needed, the study indicates that clinical use could occur in the near term, particularly in specialized pediatric centers.