GUI: Guided trajectory construction policy transfer and groupwise preference learning for liver disease prediction

Abstract

Liver disease prediction from electronic health records is clinically important but remains challenging because clinical evidence is heterogeneous, weakly discriminative, and often requires guideline-consistent interpretation. Existing approaches primarily optimize outcome-level classification accuracy, whereas the underlying diagnostic reasoning process is often implicit, unstable, and difficult to control. To address these challenges, we propose GUI, a guided trajectory-based framework for liver disease prediction. GUI formulates diagnostic inference as a process of constructing, transferring, and optimizing case-specific diagnostic trajectories. Specifically, the GUI first constructs diagnostic paths under medical knowledge and guideline constraints. It then transfers structured reasoning policies from expressive teacher models to compact predictors. Finally, it refines diagnostic decisions through groupwise preference learning over competing trajectories within each case. This design enables diagnostic hypotheses to be explicitly generated, compared, and optimized during prediction, rather than being learned only from final labels. Experiments on a liver disease cohort derived from MIMIC-IV demonstrate that GUI consistently outperforms general large language models, domain-specific medical models, and agent-based baselines across multiple evaluation metrics. GUI also shows more stable and interpretable diagnostic behavior. These results indicate that guided trajectory modeling provides an effective and controllable approach for mechanism-oriented liver disease prediction.