Predicting advanced fibrosis using non-invasive clinical tests and modern machine learning methods in TARGET-NASH.

Challenge

Standard FIB-4 and NFS cutoffs for advanced fibrosis were developed using conventional statistical methods, but it was unclear whether modern machine learning approaches applied to their constituent variables could achieve meaningfully better sensitivity-specificity tradeoffs in a real-world NAFLD cohort.

Solution

TARGET-NASH biopsy and NIT data were used to train and validate logistic regression, lasso, boosted tree, and neural network models using the individual components of FIB-4 and NFS, comparing model performance across standard and profit-matrix-weighted classification approaches.

Impact

Demonstrating that machine learning models applied to standard NIT variables achieve improved performance tradeoffs—especially when weighted toward identifying advanced fibrosis—provides the methodological foundation for developing next-generation non-invasive staging algorithms for NAFLD trials.