Beyond Retrieval Failure: Representational Gaps in Clinical AI for Women’s Health

In low-resource settings, clinical AI is often developed using standardized terminologies and datasets created elsewhere, while representative women’s health narratives are rarely publicly available. This makes it especially important to determine whether apparent model failures arise from the model itself or from limitations in the available clinical representation. We investigate this distinction using 3,361 women’s wellness narratives and a clinician-annotated reference set mapped against SNOMED CT and ICD-11. Only 14.3% of clinician-reviewed narratives were exactly represented, while 38.1% were partially represented and 47.5% were unmapped. We further evaluated three generative language models and the BioLORD-2023-C biomedical retrieval model. Although convergent generative-model judgments identified clinician-confirmed representational gaps with high precision, model abstention varied substantially, and embedding similarity did not reliably distinguish exact, partial, and unmapped narratives. Error analysis revealed force-fit cases in which computationally plausible concepts discarded relational, epistemic, metaphorical, or patient-specific meaning. These findings show that clinical NLP evaluation must distinguish whether a model can retrieve an adequate concept when one exists from whether the target terminology can represent the patient’s meaning in the first place. This distinction is particularly important where locally representative data, specialist oversight, and opportunities to adapt imported clinical AI systems are limited.