An LLM-based multi-agent system for patient discharge assessment

Hospital discharge planning plays a central role in patient flow, quality of care, and hospital efficiency. However, discharge assessment is often fragmented across teams and performed under tight time constraints. Existing machine-learning models trained on electronic health record (EHR) data have been applied to forecast near-term discharge events, but they primarily estimate timing or probability and provide limited structured clinical interpretability. To address these limitations, we propose an LLM-based multi-agent clinical decision support system for discharge readiness assessment. The architecture decomposes discharge reasoning into five clinically distinct specialist agents, each producing structured, evidence-linked claims grounded in specific EHR fields. A deterministic, schema-based claim verification layer cross-checks the asserted field values against the patient record prior to arbitration, and a reliability-aware arbiter synthesizes verified assessments into a final readiness determination. This design replaces a single scalar score with multidimensional, auditable reasoning while explicitly separating generative reasoning from value-level data checking. Evaluated on a retrospective MIMIC-IV cohort across four LLM backbones, binary prediction of proxy safe early discharge under the SAFE_D24 definition ranged from 0.490 to 0.620 F1, with PPV from 0.521 to 0.628. GPT-4.1 achieved the strongest performance (0.620 F1, 75.8% specificity, 61.3% sensitivity). Ablation analysis demonstrates measurable performance gains from deterministic verification. Beyond discrimination metrics, the system produces traceable barrier identification, claim-level verification statistics, and configurable human-review triggers, supporting traceable and auditable discharge assessment with explicit safety and operational trade-offs.