Search papers, labs, and topics across Lattice.
This study investigates the use of interaction-aware mixture-of-experts models to predict post-stroke rigidity by leveraging multi-level views of structured health records. Although the performance improvements were modest, the analysis of routing attribution highlighted significant differences in the importance of various data views, emphasizing the critical role of view construction in enhancing model interpretability. These findings suggest that understanding how different views contribute to predictions can inform better model design and clinical decision-making.
Routing attribution reveals that the choice of data views can dramatically influence model interpretability, even when performance gains are minimal.
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.
