BioFact-MoE: Biologically Factorized Mixture of Experts for Vision-Language Prognostic Modeling in Hepatocellular Carcinoma

Researchers have developed BioFact-MoE, a machine learning framework that uses specialized expert networks to separately analyze liver and tumor factors in hepatocellular carcinoma prognosis. The model achieves superior survival prediction accuracy (75%+ AUC at 12-18 months) while providing interpretable biological insights into treatment heterogeneity.

BioFact-MoE addresses a fundamental limitation in medical AI: the tendency of vision-language models to create opaque, entangled representations that blend distinct biological factors. Rather than forcing a single neural pathway to handle both hepatic function and tumor characteristics simultaneously, this framework employs a Mixture of Experts architecture where specialized sub-networks focus on discrete biological domains. This decomposition approach mirrors how clinicians actually reason about HCC prognosis—by evaluating liver reserve and oncologic burden as distinct yet interconnected variables.

The research builds on growing recognition that AI systems in medical domains must balance predictive performance with clinical interpretability. Previous prognostic models often achieved reasonable accuracy without revealing why predictions were made, limiting physician trust and adoption. BioFact-MoE demonstrates that biological factorization can simultaneously improve both metrics: the model outperforms baselines across all time horizons while generating expert weight patterns that correlate with known clinical markers (liver function tests, tumor burden metrics) without explicit supervision during training.

The clinical implications extend beyond statistical improvements. By stratifying patients through pathway-informed gating mechanisms, the framework uncovers treatment-associated survival heterogeneity—identifying subpopulations likely to benefit from different therapeutic approaches. This moves precision medicine from risk quantification toward actionable patient stratification. The validation methodology, including held-out cohort testing and unsupervised correlation with clinical markers, demonstrates scientific rigor appropriate for potential clinical implementation.

Future development should examine whether these biologically-factorized architectures generalize across other heterogeneous cancers and multimodal medical data types, and whether gated expert patterns can directly influence treatment selection in prospective clinical settings.

• →BioFact-MoE separates liver and tumor factors into specialized expert networks, achieving 75%+ AUC on survival prediction across multiple time horizons
• →The factorized approach improves both predictive accuracy and biological interpretability compared to single-pathway vision-language models
• →Gated expert weights reveal clinically meaningful treatment heterogeneity without requiring explicit supervision during model training
• →Unsupervised validation shows hepatic embeddings selectively associate with liver function markers and tumor embeddings with tumor burden metrics
• →The framework demonstrates that architectural choices reflecting biological domain knowledge can enhance both AI performance and clinical trustworthiness