{"id":20424,"date":"2026-07-31T10:32:52","date_gmt":"2026-07-31T10:32:52","guid":{"rendered":"https:\/\/woncaeurope2026.org\/sessions\/socio-demographic-gaps-in-llm-guided-pain-management\/"},"modified":"2026-07-31T10:32:52","modified_gmt":"2026-07-31T10:32:52","slug":"socio-demographic-gaps-in-llm-guided-pain-management","status":"publish","type":"wsa_session","link":"https:\/\/woncaeurope2026.org\/fr\/sessions\/socio-demographic-gaps-in-llm-guided-pain-management\/","title":{"rendered":"Socio-Demographic Gaps in LLM-Guided Pain Management"},"content":{"rendered":"<p>Large language models (LLMs) are increasingly used in clinical decision support. Their recommendations may vary by socio-demographic descriptors, which can influence pain care. These variations may create safety concerns, especially when the model assigns high risk while also recommending high-risk treatments.To measure whether LLM-generated pain-management recommendations differ across socio-demographic groups, and to examine whether models apply inconsistent reasoning when assigning opioid recommendations and risk assessments.We evaluated ten widely used LLMs using 1,000 acute-pain clinical vignettes. Each vignette had two versions (cancer and non-cancer pain) and was iterated across 34 demographic categories, including a control version with no demographic information. In total, the models produced 3.4 million responses. For each response, we extracted recommendations on opioid use (yes\/no and dose level), anxiety treatment, perceived psychological stress, risk scores, and need for monitoring. We used logistic and linear mixed-effects models to test differences between demographic groups.Across all models, historically marginalized groups\u2014especially individuals described as Black, LGBTQIA+, or unhoused\u2014received more frequent and stronger opioid recommendations in both cancer and non-cancer pain scenarios. In cancer scenarios, some groups received opioid recommendations in more than 90% of cases. At the same time, these same groups were often labeled as \u201chigher risk\u201d and assigned elevated monitoring or concern scores. The direction of recommendations and the risk scores did not align. Conversely, low-income and unemployed groups often received higher risk scores but fewer opioid recommendations. Patterns in anxiety treatment and psychological stress ratings also varied across groups, despite identical clinical information.LLMs showed systematic and conflicting patterns: they sometimes recommended high-risk analgesics while simultaneously labeling the same patient profile as high risk for misuse or harm. This inconsistency suggests a reasoning problem rather than a simple bias in one direction. These patterns do not reflect guideline-based practice and may lead to unequal or unsafe care if used in clinical settings.LLM-generated pain-management recommendations vary across socio-demographic groups and sometimes contain conflicting reasoning. These patterns pose potential safety risks. Routine evaluation of LLM behavior and alignment with established pain-care guidelines is needed before clinical deployment.<\/p>\n","protected":false},"template":"","class_list":["post-20424","wsa_session","type-wsa_session","status-publish","hentry","description-off"],"_links":{"self":[{"href":"https:\/\/woncaeurope2026.org\/fr\/wp-json\/wp\/v2\/wsa_session\/20424","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/woncaeurope2026.org\/fr\/wp-json\/wp\/v2\/wsa_session"}],"about":[{"href":"https:\/\/woncaeurope2026.org\/fr\/wp-json\/wp\/v2\/types\/wsa_session"}],"wp:attachment":[{"href":"https:\/\/woncaeurope2026.org\/fr\/wp-json\/wp\/v2\/media?parent=20424"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}