Research map / Large language models / Agents, code and math

Teaching LLMs to Recommend and Defer in Underrepresented Epilepsy Care

NeurIPS 2026 · Large language models › Agents, code and math · approach: Agent instruction-following benchmarks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Manana, a non-parametric prompt-learning framework (single and multi-agent memory variants) for predicting anti-seizure medication regimens from longitudinal Ugandan clinic notes, plus Bayesian prompt averaging for confidence-based deferral.

Problem

LLM decision support for longitudinal anti-seizure medication prescribing in low-resource settings must adapt to local prescribing practice and know when to defer to specialists, but public benchmarks are dominated by high-income settings.

Approach

Learns an auditable prompt memory from prescription errors via a Predictor-Inspector-Architect loop with quorum-gated updates across patients, and treats the learned prompt trajectory as a Bayesian model average to produce prescription probabilities and deferral signals.

Key results

• On Cohort B, Manana-Multi reaches 99% top-1 precision on the most confident 25% of cases and 95% on the most confident 50%.
• Manana-Multi achieves the highest monotherapy EM@3 on Cohort A (89.7/95.3/97.0 across V1-V3), outperforming DSPy (GEPA) and TextGrad.
• BPA improves Top-1 accuracy over majority vote on the same Manana-Multi trajectory (70% vs 66% on Cohort A; 75% vs 73% on Cohort B) while Top-3 is unchanged.

Surprising findings

• A previous-regimen copy baseline reaches 85.1% EM@3 on Cohort A later visits, a strong continuation floor that most methods do not exceed on Cohort B.
• DSPy with GEPA beats Manana-Single on several monotherapy and polytherapy cells, so Manana is not uniformly best.

Limitations

• BPA gives a retrospective confidence score; deferral thresholds were chosen on held-out data and are not deployment-ready.
• Endpoint is agreement with physician-selected regimen, not clinical optimality or seizure outcome.
• Results rely on a single base model in main experiments and a small 50-patient training set.

Experimental setup

Datasets: Ugandan pediatric epilepsy Cohort A (332 patients, 1,040 visits) Ugandan pediatric epilepsy Cohort B (367 patients, 1,509 visits) MIMIC-IV (reproducibility check)

Benchmarks: Ugandan Cohort A held-out (EM@3) Ugandan Cohort B transfer (EM@3)

Metrics: EM@3 Top-1 accuracy Jaccard similarity precision at coverage confidence gap

Baselines: EpiPick Naive Bayes XGBoost Base Prompt Single-agent prompt DSPy (GEPA) ExpeL TextGrad Consilium

Builds on: TextGrad ExpeL DSPy GEPA

Closest papers by meaning

• Small Models are LLM Knowledge Triggers for Medical Tabular Prediction (ICLR 2025 · similarity 0.84)
• SAFER: A Calibrated Risk-Aware Multimodal Recommendation Model for Dynamic Treatment Regimes (ICML 2025 · similarity 0.82)
• Evaluating LLMs Across Multi-Cognitive Levels: From Medical Knowledge Mastery to Scenario-Based Problem Solving (ICML 2025 · similarity 0.82)
• Toward Better EHR Reasoning in LLMs: Reinforcement Learning with Expert Attention Guidance (AAAI 2026 · similarity 0.82)
• LLM-Driven Treatment Effect Estimation Under Inference Time Text Confounding (NeurIPS 2025 · similarity 0.82)
• CLR-voyance : Reinforcing Open-Ended Reasoning for Inpatient Clinical Decision Support with Outcome-Aware Rubrics (NeurIPS 2026 · similarity 0.81)