Tabular Population Priors as Structured Retrieval for EHR Medication Set Prediction

Impact 6 arXiv cs.AI

Summary

  • Researchers proposed using tabular population priors as a structured retrieval method for EHR medication set prediction.
  • The approach uses a sparse condition-medication co-occurrence table estimated from training admissions instead of patient-level nearest neighbors.
  • Replacing the patient retriever in PACERAG improved F1 score from 0.303 to 0.338 and Jaccard from 0.218 to 0.237 on MIMIC-IV admissions.
  • This method requires no additional LLM calls and no inference-time patient index while improving task alignment and audibility.

Why it matters

Replacing patient-level nearest neighbor retrieval with structured population priors improves clinical RAG performance for medication prediction without adding LLM overhead or inference-time patient indexing.

Who should care