Electronic Health Records (EHRs) are split into two worlds. The first is the world of structured data: lab results, vital signs, and billing codes. This data is easy for machines to process but often lacks context. The second is the world of unstructured narratives: physician notes, discharge summaries, and pathology reports. These narratives contain the "clinical soul" of a patient’s journey—nuanced observations, diagnostic reasoning, and social context—but they are notoriously difficult for algorithms to parse.
Research in this field focuses on bridging this gap, transforming "messy" clinical prose into structured, actionable insights to improve everything from mortality prediction to clinical trial recruitment.
The Information Gap in Structured Data
Standardized medical codes (like ICD-10) are primarily designed for billing, not for capturing clinical nuance. As Doron Stupp and colleagues at Google note, these codes often prioritize reimbursement over the subtle reasoning found in a doctor’s "Assessment and Plan" section [Stupp et al., 2022].
This gap has real-world consequences. For example, Preethi Raghavan demonstrated that up to 77% of clinical trial eligibility criteria cannot be satisfied using structured data alone; the necessary details are buried in free-text notes [Raghavan et al., 2015]. Similarly, Willie Boag and Marzyeh Ghassemi found that clinical notes contain latent signals about patient rapport and mistrust that are far more predictive of end-of-life outcomes than simple demographic data [Boag et al., 2018].
Mechanisms of Information Fusion
To unlock this information, researchers use multimodal learning, which treats structured and unstructured data as complementary views of the same patient.
- Feature Extraction and Embedding: Early approaches used Named Entity Recognition (NER) to identify medical terms in notes, which were then converted into vector representations. Mengqi Jin and Parminder Bhatia showed that fusing these text embeddings with time-series vitals improves ICU mortality predictions by capturing the "acuity" that numbers alone miss [Jin et al., 2018].
- Topic Modeling: Unsupervised methods like Latent Dirichlet Allocation (LDA) allow researchers to discover hidden themes in notes. Katherine Redfield Chang used this to identify cancer phenotypes that were invisible to standard diagnostic codes [Chang et al., 2016].
- Domain-Specific Language Models: General-purpose models often struggle with medical jargon. Researchers like Colin Magdamo have turned to domain-adapted models like Clinical BERT to cluster complex diseases like Alzheimer’s into more precise subtypes based on co-morbidities mentioned in notes [Magdamo et al., 2024].
Scaling with Large Language Models
The advent of Large Language Models (LLMs) has shifted the focus toward zero-shot extraction—getting structured data without needing thousands of hand-labeled examples.
However, LLMs face a "needle in a haystack" problem when processing long, dense clinical records. Yixiang Qu and colleagues found that simple preprocessing, such as Retrieval-Augmented Generation (RAG), can "supercharge" smaller, privacy-preserving models to outperform massive general-purpose ones [Qu et al., 2024].
To make these models practical for hospital deployment, researchers like Jimeng Sun use hybrid pipelines: a fast, traditional deep learning model flags relevant sentences (e.g., those mentioning social determinants of health), and only those snippets are sent to an expensive LLM for detailed labeling [Sun et al., 2025].
From Text to Interoperable Knowledge
The ultimate goal is to move from "text in a box" to standardized knowledge graphs. Systems like Infherno Frei, 2025 and mCODEGPT [Zhang & Long, 2025] focus on mapping free text directly to global standards like FHIR (Fast Healthcare Interoperability Resources) or oncology ontologies.
By decomposing complex narratives into discrete, visual, or logical aspects—such as the texture and location of a tumor described in a radiology report [Phan et al., 2024]—these systems allow computers to "see" the patient with the same level of detail as a human clinician. As George Hripcsak and Anna Ostropolets argue, organizing data to mimic clinical reasoning is the key to moving beyond manual chart review toward truly automated clinical decision support [Ostropolets et al., 2022].