Developing an explainable machine learning model to predict false-negative citrin deficiency cases in newborn...

Area of Science:

• Biochemistry
• Genetics
• Medical Informatics

Background:

• Neonatal Intrahepatic Cholestasis caused by Citrin Deficiency (NICCD) is an inherited metabolic disorder.
• Standard newborn screening (NBS) for NICCD can yield false negatives due to normal citrulline levels, delaying diagnosis.
• Early detection of NICCD is crucial for timely intervention and improved patient outcomes.

Purpose of the Study:

• To develop and validate an interpretable machine learning (ML) model for predicting false-negative NICCD cases.
• To identify key predictive features for NICCD beyond standard screening parameters.
• To enhance the accuracy and effectiveness of NBS for NICCD.

Main Methods:

• Retrospective analysis of data from 53 NICCD patients and 212 controls.
• Development and evaluation of six ML models, including XGBoost, using metabolite and demographic data.
• Application of SHAP (Shapley Additive exPlanations) for model interpretability and feature importance analysis.

Main Results:

• The XGBoost model achieved high performance (AUC > 0.97, F1 score > 0.83) in predicting false-negative NICCD cases.
• Key predictive features included birth weight, citrulline, glycine, phenylalanine, ornithine, and succinylacetone.
• SHAP analysis provided patient-level insights into model predictions and feature interactions.

Conclusions:

• An interpretable ML model using metabolite and demographic data significantly improves the detection of false-negative NICCD cases.
• This approach facilitates earlier identification and intervention for NICCD patients.
• The developed model and calculator enhance the overall newborn screening system for metabolic disorders.