JEDA: Query-Free Clinical Order Search from Ambient Dialogues

TL;DR

JEDA addresses the problem of retrieving canonical clinical orders directly from mixed directive and ambient clinical language without intermediate query rewriting. It introduces a domain-initialized bi-encoder, initialized from PubMedBERT, trained with a duplicate-safe contrastive objective and constrained LLM-guided supervision to align both explicit commands and ambient dialogue with order concepts. The method supports a query-free mode using a short ambient window, achieving strong top-k recall, sharper inter-order separation, and robust performance across diverse query formulations, including encounter-scoped constraints. Practically, JEDA provides a fast, interpretable, LLM-free retrieval layer that can operate in real time and be paired with rule-based reranking for clinical deployment, highlighting the value of domain-aligned reasoning supervision over sheer model scale.

Abstract

Clinical conversations mix explicit directives (order a chest X-ray) with implicit reasoning (the cough worsened overnight, we should check for pneumonia). Many systems rely on LLM rewriting, adding latency, instability, and opacity that hinder real-time ordering. We present JEDA (Joint Embedding for Direct and Ambient clinical orders), a domain-initialized bi-encoder that retrieves canonical orders directly and, in a query-free mode, encodes a short rolling window of ambient dialogue to trigger retrieval. Initialized from PubMedBERT and fine-tuned with a duplicate-safe contrastive objective, JEDA aligns heterogeneous expressions of intent to shared order concepts. Training uses constrained LLM guidance to tie each signed order to complementary formulations (command only, context only, command+context, context+reasoning), producing clearer inter-order separation, tighter query extendash order coupling, and stronger generalization. The query-free mode is noise-resilient, reducing sensitivity to disfluencies and ASR errors by conditioning on a short window rather than a single utterance. Deployed in practice, JEDA yields large gains and substantially outperforms its base encoder and recent open embedders (Linq Embed Mistral, SFR Embedding, GTE Qwen, BGE large, Embedding Gemma). The result is a fast, interpretable, LLM-free retrieval layer that links ambient context to actionable clinical orders in real time.

Figures (7)

• Figure 1: Training record construction. For each signed order, the pipeline selects supporting transcript subclusters, synthesizes a command, preserves verbatim context, generates a concise reasoning, and maps to a canonical concept.
• Figure 2: Contrastive fine tuning with duplicate safe Multiple Negatives Ranking. Queries and documents are encoded with a tied Bi-encoder, duplicate positives are masked, and the MNR loss is applied.
• Figure 3: Query to Order UMAP projections. JEDA produces tighter clusters around the correct order embeddings with clearer inter order separation compared to PubMedBERT.
• Figure 4: Query to Query UMAP projections. JEDA exhibits tighter same order groupings and clearer boundaries.
• Figure 5: Variant level Query to Order cosine similarity distributions. JEDA shows higher central tendency across variants.