Multi-agent AI orchestration for temporal-aware extraction of social determinants of health from unstructured clinical records in cancer populations

Purpose: Social determinants of health (SDOH) influence cancer outcomes, yet extraction of these variables from unstructured clinical documents in electronic health records remains challenging due to their temporal complexity and narrative dispersion across multiple document types. Single large language models (LLMs) frequently hallucinate, misidentify temporal context, and cannot reliably distinguish between current barriers and historical mentions. We developed a multi-agent artificial intelligence (AI) orchestration framework that employs specialized agents with defined roles to extract temporally-aware SDOH variables for cancer research and clinical decision support. Experimental Design: We implemented coordinated agents with different roles, focused on specific social determinant domains, directing task allocation, verifying extracted information, and assembling temporally-ordered patient timelines. We assembled two retrospective cohorts including cancer patients with HIV (n=100) and peri-operative pain management (n=524), encompassing 81k words/patient (≈8.1M words in 3,460 documents) and 49k words/patient (≈25.8M words in 15,484 documents), respectively. The framework processed clinical data, nursing communications, social work assessments, pathology reports, and imaging reports. Results: We successfully extracted temporal SDOH variables with >95% confidence across both cohorts, constructing comprehensive patient timelines that distinguished between current and historical barriers. The multi-agent architecture addressed single-model limitations by cross-validating extractions, grounding findings in specific document evidence, and maintaining temporal accuracy. For HIV cohort, the system processed 10.8M model tokens, automatically identifying financial difficulties, transportation barriers, and housing instability with dates and supporting documentation. The peri-operative pain cohort required 34.3M tokens, revealing temporal patterns in social support availability and medication access challenges. Beyond SDOH variables, system automatically extracted complementary clinical information, including diagnoses and care transitions, enabling researchers to contextualize SDOH within broader clinical narratives. Conclusions: We demonstrated that orchestrated multi-agent AI can reliably extract temporally-aware SDOH from real-world clinical documentation, addressing critical limitations of single LLM approaches. By providing structured, time-stamped extraction with evidence provenance, the framework enables practical clinical applications including patient triage for social work intervention, standardized disparities reporting, and research cohort characterization, making it suitable for deployment across diverse cancer populations.