Understanding Structured Health Data through Interaction-Aware Mixture-of-Experts

2026-07-13

AI Summary

This work addresses the underutilization of multi-level structured health records and the limited interpretability of existing models in post-stroke muscle tone prediction. To this end, the authors propose an Interaction-Aware Mixture-of-Experts (MoE) model that integrates multi-view representations of structured health data. By incorporating a routing attribution mechanism, the model not only captures interaction effects among distinct data views but also quantifies their differential contributions to predictive decisions. Experimental results demonstrate that, while yielding modest gains in predictive performance, the proposed approach substantially enhances model transparency. The study systematically reveals the critical role of view construction in shaping interpretability, thereby offering clinically actionable insights and a more trustworthy foundation for decision support in stroke rehabilitation.

Abstract

We study interaction-aware mixture-of-experts for post-stroke rigidity prediction using multi-level views of structured health records. Despite minimal performance gains, routing attribution reveals systematic importance differences across views, underscoring view construction as key to interpretability.

Problem

• structured health data
• post-stroke rigidity
• mixture-of-experts
• view interpretability
• interaction-aware modeling

Innovation

• Interaction-Aware Mixture-of-Experts
• structured health data
• view construction
• routing attribution
• post-stroke rigidity prediction