Federated Deep Learning for Privacy-Preserving Multi-Site Liver Fibrosis and Steatosis Prediction: A Comprehensive Study on the BEHSOF Tabular Clinical Dataset

Abstract

A crucial clinical challenge in the treatment of non-alcoholic fatty liver disease (NAFLD), chronic hepatitis, and associated hepatic conditions is the non-invasive evaluation of the degree of liver fibrosis and steatosis. The current gold standard, liver biopsy, is invasive, expensive, and prone to sampling variability, which has prompted the creation of reliable machine-learning surrogates using regularly obtained clinical laboratory data. However, institutional data-sharing limitations, patient privacy laws (GDPR, HIPAA), and the statistical heterogeneity of multi-site cohorts limit the use of centralized deep learning models in clinical settings. This paper introduces Fed-TabNet, a federated learning framework that uses the publicly available BEHSOF Figshare dataset to train a tabular neural network (a three-layer multilayer perceptron (MLP) with 128→64→1 architecture) across two geographically separated clinical sites (TAL and BEH). No raw data leaves the original institution. Ten communication rounds with two local epochs each are coordinated by the FedAvg aggregation algorithm. On the held-out test set, the suggested system outperforms all benchmarked centralized and federated baselines with 99.0% classification accuracy, AUROC = 0.992, AUPRC = 0.978, Brier Score = 0.031, and Expected Calibration Error (ECE) = 0.011. We present a thorough evaluation pipeline that includes confusion matrices, ROC/PR curves, calibration analysis, decision curve analysis (DCA), and feature importance attribution. The framework offers a clinically implementable, privacy-preserving method for multi-site risk assessment of hepatic disease.