A prediction model for genetic cholestatic disease in infancy using the machine learning approach

Area of Science:

• Pediatric Hepatology
• Medical Informatics
• Genetics

Background:

• Infant cholestasis diagnosis is challenging, particularly for genetic causes.
• Timely identification of genetic etiologies is crucial for effective pediatric liver disease management.

Purpose of the Study:

• To develop machine learning (ML) models, named Jaundice Diagnosis Easy for Baby (JADE-B), for predicting genetic causes of cholestasis in infants.
• To identify infants at higher risk for genetic disorders, enabling targeted diagnostic approaches.

Main Methods:

• Utilized a tertiary medical center's database (2006-2018) with 1845 infants diagnosed with cholestatic liver disease.
• Employed 47 clinical and laboratory parameters to train four ML classifiers: Logistic Regression, XGBoost, LightGBM, and Random Forests.
• Developed a refined JADE-B model using 20 accessible parameters for early prediction.

Main Results:

• Included 1008 infants under 1 year with cholestatic liver disease in the analysis.
• Achieved high performance with models yielding an area under the receiver operating characteristic curve ranging from 0.836 to 0.884.
• The final JADE-B model effectively predicts genetic disorders using 20 parameters within the first month of disease onset.

Conclusions:

• The JADE-B ML model aids in prioritizing infants for genetic testing and specialist referrals.
• Optimizes the allocation of genetic diagnostic resources for pediatric cholestasis.
• Facilitates earlier and more accurate identification of genetic liver diseases in infants.