Teaching LLMs to Recommend and Defer in Underrepresented Epilepsy Care

This paper studies LLM decision support for longitudinal pediatric epilepsy care in Uganda, where specialist expertise and prescribing resources are limited. It introduces MANANA, a non-parametric prompt-learning framework that derives auditable prompt memories from prescription errors observed in a small patient-level training set. Across two independently collected Ugandan cohorts, MANANA outperforms classical machine-learning, direct-prompting, and prompt-optimization baselines. Bayesian prompt averaging turns the learned prompt trajectory into prescription likelihoods and an uncertainty-based deferral signal. On an independent held-out cohort, visit-level top-3 prescription accuracy improves by 4–8 percentage points over prompt-optimization baselines; the system reaches 95% precision on the most confident half of cases and 99% on the most confident quarter.