Machine learning algorithm improves the detection of NASH (NAS-based) and at-risk NASH: A development and validation...

Area of Science:

• Hepatology
• Machine Learning
• Biomarkers

Background:

• Non-alcoholic steatohepatitis (NASH) detection is challenging.
• At-risk NASH (NASH with fibrosis stage ≥ 2) is crucial for drug development.
• Predictive models can aid in staging and grading NAFLD patients.

Purpose of the Study:

• To develop and validate machine learning models for detecting NASH and at-risk NASH.
• To compare the performance of clinical data versus extended data (clinical + biomarkers).
• To assess models for steatosis, inflammation, ballooning, and fibrosis staging.

Main Methods:

• Supervised machine learning (gradient boosting machine) applied to a NAFLD cohort (n=966).
• Models developed using clinical data and biomarkers.
• Data split into training/validation sets (75/25).
• Direct and composite models constructed for NASH and at-risk NASH.

Main Results:

• Clinical models for steatosis, inflammation, and ballooning showed high AUCs (0.94, 0.79, 0.72).
• Composite NASH model achieved AUC 0.71; composite at-risk NASH model achieved AUC 0.83.
• Fibrosis models (significant and advanced) performed well, with AUCs up to 0.86 for advanced fibrosis.
• Adding biomarkers did not improve NASH or at-risk NASH model accuracy but slightly improved fibrosis detection.

Conclusions:

• Independent machine learning models using clinical predictors can improve NASH and at-risk NASH detection.
• Composite models demonstrate superior performance.
• Biomarkers offer limited benefit for NASH/at-risk NASH prediction but aid fibrosis assessment.