Adaptive-Personalised Federated Deep Learning for Privacy-Aware NAFLD Screening

Non-alcoholic fatty liver disease (NAFLD) moves nearly a quarter of global adult population, yet current diagnostic pathways still rely on resource-intensive ultrasonography or invasive biopsy. This study introduces an Adaptive-Personalised Federated Deep-Learning (A-P-FedDL) framework that enables collaborative, privacy-preserving prediction of NAFLD from routine clinical and laboratory variables collected at geographically dispersed hospitals. The method builds on a lightweight four-layer convolutional neural network trained under the FedAvg protocol and augments it with two novel components: client-similarity weighting, which dynamically scales each participant’s model update by the statistical distance between local and global feature distributions, and adaptive local-epoch scheduling that lengthens or shortens on-device training depending on convergence speed. Experiments were conducted on a cohort of 577 subjects (377 positive, 200 negative) from New Taipei City Municipal Hospital using stratified five-fold cross-validation repeated three times. A-P-FedDL achieved 94.2 % accuracy, 93.1 % sensitivity, 95.3 % specificity, an F1-score of 0.934, and an AUROC of 0.973, outperforming vanilla FedAvg-CNN by 2.7 percentage points and the best centralised baseline by 3.5 points. The framework also converged in 60 rounds and reduced perclient communication to 6.6 MB, representing a 41 % bandwidth saving. A Wilcoxon test confirmed statistical significance (p = 0.003). These findings demonstrate that personalised aggregation and adaptive training schedules can simultaneously enhance predictive performance and communication efficiency, paving the way for scalable deployment of edge-enabled liver-health screening in primarycare networks. Further, the modular design allows straightforward extension to additional biochemical markers and imaging features, supporting precision hepatology initiatives.