Enriching Electronic Health Records with Semantic Features: Leveraging Multi-Modality Embeddings.

The exponential growth of Electronic Health Records (EHRs) offers rich clinical narratives for
healthcare improvement, but their unstructured nature makes information retrieval challenging.
While models like BERT excel in classification tasks, their potential for EHR-based retrieval—
especially in combining structured and unstructured data—remains underutilized.
We introduce ClinicalNarr, a transformer-based model for clinical information retrieval. Our
initial research examined four language representation settings, finding that concept-only em-
beddings achieve superior performance (BERTScore F1: 0.699) by emphasizing key medical
concepts while filtering noise. Building on these insights, ClinicalNarr leverages MIMIC-III
data, combining structured ICD codes with unstructured narratives to improve retrieval. When
tested on MedNLI, it achieved 90.5% accuracy, outperforming previous models, with strong cor-
relations with physician judgments (r=0.69) and medical coders (r=0.76).
Comparative analysis reveals HNSW’s superiority over IVF for semantic retrieval (NDCG@10:
0.49 vs. 0.43). ClinicalNarr outperformed existing models, achieving three times better hit
rate. This advantage expanded with our novel ontology-augmented evaluation methodology,
which increased performance by 13%—reaching a 52% Hit Rate compared to competing models
(15-20%).
This study establishes a benchmark for clinical information retrieval using MIMIC-III, sig-
nificantly advancing clinical decision support capabilities by improving semantic alignment in
healthcare information systems