ATHENA Automates EHR Model Design with Agentic NAS.

Designing optimal Transformer-based models for clinical predictions from Electronic Health Records (EHRs) typically involves extensive manual tuning, with ideal architectures varying across tasks and healthcare providers. While Neural Architecture Search (NAS) can automate this, conventional methods are computationally intensive for complex Transformer models. Existing LLM-guided NAS approaches often conduct searches independently, failing to reuse valuable architectural insights across different hospitals. ATHENA (Agentic Transfer across Hospitals for EHR Neural Architecture Search) addresses these challenges. It employs a weight-sharing supernet, pretrained once per hospital, allowing candidate architectures to be quickly evaluated as inherited subnetworks. Crucially, ATHENA incorporates a two-layer cross-hospital architecture prior. The first layer retrieves high-performing examples from other sites based on task descriptions, while the second uses SHAP-based meta-regression to estimate component effects. These priors, combined with validation feedback from the target hospital, guide a multi-agent LLM search. ATHENA has demonstrated superior or comparable performance to baseline NAS methods in clinical prediction tasks across two health systems, significantly reducing manual tuning and showing more consistent architecture selection. The code is publicly available, offering a practical solution for EHR modeling.