FedLiver: Federated Learning for Hepatocellular Carcinoma Classification using Open Gene Expression Data
Hepatocellular carcinoma is a leading cause of cancer-related mortality worldwide. However, early detection has proven challenging. Machine learning (ML) methods for HCC classification typically require centralized gathering of sensitive genetic data, limiting effective hospital collaboration due to privacy concerns. FedLiver, a federated learning (FL) platform, enables multiple institutions to jointly train an accurate HCC classification model without exchanging raw patient gene expression data. Using public transcriptome data from The Cancer Genome Atlas (TCGA-LIHC, n=424 samples), the top 500 most variable genes were selected from 59,319 genes. To replicate independent and identically distributed (IID) and non-independent and non-identically distributed (Non-IID) healthcare contexts, data was dispersed among five virtual clients. Hyperparameter optimization enhanced performance while lowering model complexity in a lightweight neural network with a privacy-preserving Federated Averaging (FedAvg) approach. On IID partitions, the framework’s 96.54% accuracy and 0.9933 AUC closely matched centralized performance (97.65% accuracy). Under Non-IID conditions, FedLiver maintained strong performance (94.16% accuracy, 0.9894 AUC). A smaller model architecture with a learning rate of 0.001 was shown to provide the best trade-off between accuracy, speed, and generalization (64,32). These results demonstrate that FL permits collaborative AI for liver cancer diagnosis while protecting privacy and complying with legal requirements such as the General Data Protection Regulation (GDPR) and the Health Insurance Portability and Accountability Act (HIPAA). The code and data pipeline are made publicly available for repeatable research.
Authors:
Muhammad Rezki Rasyak and Bens Pardamean.
ICIMTECH