Attention-driven graph-based machine learning for non-invasive diagnosis of NAFLD

Abstract

An estimated 25%–30% of the global population is affected by non-alcoholic fatty liver disease (NAFLD), a silent yet progressive condition that can advance from simple steatosis to severe stages like non-alcoholic steatohepatitis (NASH), fibrosis, and cirrhosis, significantly heightening the risk of liver cancer. Currently, the gold-standard method for staging NAFLD is liver biopsy, an invasive procedure with risks such as bleeding, infection, and sampling error. Due to its high cost and impracticality for routine monitoring, there is a critical need for reliable, non-invasive diagnostic tools capable of effectively identifying NAFLD stages. We developed a graph-based framework in which each patient is represented as a node in a similarity network. Edges are formed via k-nearest neighbors (KNN) on standardized clinical and biochemical features, with missing values imputed by KNN to preserve biologically plausible variability. A two-layer Graph Attention Network (GAT) then learns edge-specific attention weights to focus on the most informative inter-patient relationships. Tested on a proprietary ILBS cohort (n = 622), our model achieved 75.2% accuracy (AUC = 0.768; F1 = 0.752), an 11% absolute improvement over Support Vector Machines and Random Forests, and demonstrated robustness in 10-fold cross-validation and adversarial noise tests. On a separate public dataset (n = 80) spanning lipidomic, glycomic, fatty acid, and hormone panels, it exceeded 99% accuracy (AUC $>$ 0.99). Attention-based explanations further highlighted key patient similarities driving each prediction. These findings suggest that attention-driven graph learning can clearly improve non-invasive NAFLD staging, enabling early detection and supporting personalized disease monitoring in diverse clinical settings.