The authors introduce LongMedBench, a real-world electronic health record (EHR) benchmark designed to evaluate long-horizon clinical decision-making in medical agents. Unlike prior evaluations that focus on short-context knowledge QA, this benchmark addresses the longitudinal nature of medical care by aggregating evidence across repeated visits and evolving treatments.

  • Constructed via a pipeline integrating MIMIC-IV admission records and clinical notes into time-series event streams and long-context memory datasets.
  • Comprises 335 patients with an average of 19.72 inpatient visits per patient and 44.91 medical events per visit.
  • Features an evaluation taxonomy with three suites: fact-based QA, temporal reasoning, and long-horizon decision-making.
  • Experiments reveal that while recent LLMs handle explicit timestamps well, they struggle with implicit time inference.
  • RAG and agent memory systems improve information retrieval performance, but decision-making performance remains highly dependent on the model's immediate context.

This benchmark enables a more realistic assessment of how agents understand and leverage historical patient information over extended horizons.