Semi-Supervised Fatty Liver Classification Using Attention-Based Graph Neural Network Models

Fatty liver disease is a common condition linked to metabolic syndrome, cardiovascular diseases, and liver cirrhosis, and timely, accurate diagnosis is crucial. In clinical studies, incorporating deep learning models often faces the challenge of scarce labeled data. This study investigates the effectiveness of graph-based deep learning models with attention mechanisms to predict fatty liver disease even with limited labeled data.

We utilized a dataset of 7,953 individuals, focusing on clinical variables obtained during health check-ups. Graph Neural Networks (GNNs) with attention mechanisms were assessed for predicting fatty liver disease in a semi-supervised learning setting. GNNExplainer was employed for feature importance analysis, and subgroup analysis was conducted to identify clusters with distinct risk factors.

Our findings indicate that attention-based GNNs significantly outperformed conventional models in predicting fatty liver disease under semi-supervised settings, with statistically significant improvements in area under the curves (AUCs) (all P < 0.05) compared to logistic regression across most labeling scenarios. With only 10 labeled samples per class, the Graph Attention Network (GAT) and Simplified Graph Transformer with Graph Attention achieved AUCs of 0.7049 ± 0.0570 and 0.7184 ± 0.0395, respectively, and achieved AUCs of 0.7893 ± 0.0171 when 100 labeled samples were used. Feature importance analysis identified HbA1c (relative importance score = 1.0), body fat amount (0.998), and glucose (0.6934) as the most influential predictors. Subgroup analysis revealed two distinct patient clusters-one characterized by metabolic risk factors and the other by demographic and lifestyle factors-emphasizing the potential for individualized risk stratification in fatty liver disease.

Attention-based GNNs demonstrated strong predictive performance for fatty liver disease using a small number of labeled samples. This methodological approach illustrates how graph-based learning can leverage relational structures in routine clinical data to support data-efficient, individualized risk assessment in label-constrained settings.