TRIXMED: Triage-Routed Mixture of Experts Framework for Interpretable Drug Recommendation

Rongyu Lin, Weiwei Chen, Shichao Pei, Shengzhi LI, Mengyuan Shi

Keywords: Drug Recommendation, Mixture of Experts, Patient Heterogeneity

Abstract: Drug recommendation is a critical task in intelligent healthcare systems that significantly impacts patient outcomes. While large language models (LLMs) have advanced the field through sophisticated semantic understanding, current approaches face two fundamental challenges: (1) they fail to adequately address patient heterogeneity, treating diverse populations with a one-size-fits-all model; (2) their black-box nature undermines clinical trust and adoption. We introduce TRIXMED (Triage-Routed Interpretable eXpert Medicine), a novel framework that integrates Mixture of Experts (MoE) architecture with routing mechanisms that mimic clinical triage processes for personalized drug recommendation. TRIXMED addresses patient heterogeneity by introducing specialized experts that handle distinct patient subgroups, while ensuring interpretability through a clustering-based routing strategy that automatically directs patients to the most appropriate expert based on their clinical profile. Our approach employs a unique warm-up training phase followed by feature extraction and patient stratification, enabling transparent expert routing based on patient characteristics. Extensive experiments on the MIMIC-III datasets show that TRIXMED surpasses the SOTA model, achieving relative improvements of 27.4\% in Jaccard index, and 15.61\% in F1-score, respectively. TRIXMED represents a significant advancement in bridging the gap between AI-powered recommendations and clinical practice through its combination of heterogeneity handling and transparent decision-making.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 18416