Application of machine learning in constructing a diagnostic model for neonatal biliary atresia

Area of Science:

• Pediatric Gastroenterology
• Medical Informatics
• Neonatal Medicine

Background:

• Early diagnosis of biliary atresia (BA) is crucial for timely surgical intervention, such as the Kasai operation, significantly impacting patient prognosis.
• Developing accurate diagnostic tools for BA in neonates remains a clinical challenge.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for the early diagnosis of BA in neonates.
• To identify key clinical and serological predictors for neonatal BA.

Main Methods:

• Trained five ML models (logistic regression, random forest, support vector machine classifier, multilayer perceptron, extreme gradient boosting) on clinical and laboratory data from neonates with pathological jaundice.
• Utilized a stacking classifier (SC) algorithm to ensemble the best-performing models.
• Included 85 neonates (42 with BA) from January 2013 to December 2023.

Main Results:

• XGBoost and Random Forest models achieved perfect diagnostic performance (AUC = 1.000).
• The ensemble SC model also demonstrated excellent diagnostic accuracy (AUC = 1.000).
• Key predictors identified were elevated gamma-glutamyl transpeptidase (GGT), increased platelet (PLT) counts, and acholic stools.

Conclusions:

• ML models, particularly XGBoost and RF, show high accuracy in diagnosing BA in neonates.
• Elevated GGT, increased PLT counts, and acholic stools are significant predictors for BA.
• This ML-driven approach holds promise for earlier BA identification and intervention in the neonatal period.