DILA (Dictionary Label Attention) addresses a critical challenge in healthcare AI: creating medical coding systems that are both accurate and interpretable.

• Disentangles complex embeddings into a sparse, interpretable space
• Provides mechanistic interpretability rather than just local explanations
• Maintains competitive performance while adding transparency
• Enables medical professionals to understand and trust AI predictions

This research matters because medical coding requires high precision with clear reasoning behind each prediction. DILA's approach makes AI systems more trustworthy for clinical decision support while maintaining high accuracy.

DILA: Dictionary Label Attention for Mechanistic Interpretability in High-dimensional Multi-label Medical Coding Prediction