Search papers, labs, and topics across Lattice.
This paper introduces MamaBench, a novel counterfactual benchmark designed to assess the robustness of large language models (LLMs) in maternal and child health diagnostics by evaluating their ability to distinguish between clinically similar cases. The study employs a new retrieval method, Evidence-Anchored RAG (EA-RAG), which enhances model performance by focusing on evidence coverage rather than aggregate similarity. Results show that while base accuracy can be misleadingly high, EA-RAG significantly reduces the Bias Trap Rate (BTR) by 5.5 percentage points, highlighting the ongoing challenges in achieving counterfactual robustness in clinical AI applications.
Base accuracy can mislead clinical AI assessments, with MamaBench revealing a 16-28 percentage point gap in robust accuracy across leading LLMs.
Large language models achieve strong scores on medical benchmarks, yet these benchmarks evaluate each question in isolation, providing no measure of whether a system can distinguish clinically similar presentations requiring different interventions. We introduce MamaBench, the first counterfactual benchmark for maternal and paediatric AI: 434 expert-authored clinical narratives in 217 pairs across 371 pathologies, evaluated via the Bias Trap Rate (BTR), the conditional probability that a model fails the counterfactual given success on the base case. We propose Evidence-Anchored RAG (EA-RAG), a three-stage retrieval method that replaces aggregate similarity with an evidence coverage objective through clinical parameter extraction, coverage auditing, and contrastive sub-queries. Across eight configurations of four frontier LLMs, base accuracy overstates robust accuracy by 16-28 percentage points in every model. EA-RAG achieves 20.3% BTR and 65.0% robust accuracy on Claude Sonnet 4.6, a 5.5 percentage point BTR reduction without degrading base accuracy. The residual 20% BTR confirms that counterfactual robustness in clinical AI remains an open challenge. Keywords: counterfactual evaluation, clinical AI, maternal healthcare, retrieval-augmented generation, diagnostic robustness
