LLMs Learn to Recommend and Defer in Epilepsy Care

A new research paper addresses the challenge of deploying Large Language Models (LLMs) for clinical decision support in settings with limited specialist access, specifically focusing on pediatric epilepsy care in Uganda. The core problem is that LLMs, while capable of parsing medical notes, often struggle to align with local prescribing norms and lack a mechanism to identify when they should defer to human experts. To overcome this, the researchers introduced MANANA, a novel non-parametric prompt-learning framework. MANANA learns from a small dataset of patient records and observed prescription errors, converting these into "auditable prompt memories." This approach allows the LLM to adapt to specific local prescribing guidelines, significantly outperforming traditional machine learning models and direct LLM prompting. Furthermore, the framework incorporates Bayesian prompt averaging to generate prescription likelihoods and a crucial uncertainty-based deferral signal. This enables the system to confidently handle a significant portion of cases (e.g., 95% precision for half of cases) while flagging complex or low-confidence situations for specialist review, thereby enhancing safety and utility in critical medical applications.