Building the graph of medicine from millions of clinical narratives

Samuel G. Finlayson, Paea LePendu, Nigam H. ShahScientific 2014 · DOI

TL;DR

This paper introduces a novel dataset of co-occurrence matrices derived from millions of clinical narratives, effectively building a "graph of medicine." This resource quantifies the relationships between medical concepts, enabling more robust statistical analyses and a deeper understanding of comorbidity, drug interactions, and disease patterns.

What Problem Does This Paper Address?

Electronic Health Records (EHRs) are a treasure trove of patient information, but extracting meaningful insights has been challenging due to data access restrictions and the complexity of unstructured text. Existing methods for understanding relationships between medical entities (like drugs and diseases) often rely on manual knowledge bases, which are time-consuming to build and quickly become outdated. This paper addresses the need for a scalable, data-driven approach to quantify these inter-relationships, enabling a more accurate and dynamic understanding of clinical practice.

Methods

The core of this work is a sophisticated pipeline for processing millions of clinical notes.

  • Text Annotation: Clinical notes are first processed to identify and extract medical concepts. This involves a multi-step process including mapping terms to a comprehensive set of ontologies (like SNOMED-CT and MedDRA), applying syntactic and semantic rules to clean the lexicon, and using NegEx to identify negated mentions.
  • Note Binning: To capture temporal relationships, notes are grouped into "bins" based on their proximity in time for each patient. Various bin widths (e.g., 1-day, 7-day, 30-day, up to "infinite" day bins representing the entire patient record) are used.
  • Co-frequency Counting: A parallelized annotation, hashing, and counting pipeline is employed to compute pairwise co-occurrence frequencies of medical terms and concepts within these bins. This is a computationally intensive step, but the use of distributed processing (Spark) makes it feasible.
  • Data Structuring: The output is a set of co-occurrence matrices, which are essentially large tables quantifying how often pairs of medical concepts appear together in the clinical narratives.

Datasets

The primary dataset used is from the Stanford Translational Research Integrated Database Environment (STRIDE). This dataset comprises:

  • 20 million unstructured clinical notes
  • Collected from 1.2 million patients
  • Spanning a 19-year period

The authors also mention using a gold standard corpus from the 2008 i2b2 Obesity Challenge for validating the accuracy of their text processing pipeline.

Key Insights and Ideas

  • Quantifying the "Graph of Medicine": The paper moves beyond qualitative descriptions of medical relationships to provide quantitative measures of co-occurrence. This allows for a more precise understanding of how drugs, diseases, and procedures are linked in real-world clinical practice.
  • Scalability and Dynamic Updates: By leveraging EHR data and automated processing, this approach overcomes the limitations of manual knowledge bases. The resulting dataset can be updated as new data becomes available, reflecting the evolving nature of medicine.
  • Unlocking New Analytical Possibilities: The co-occurrence matrices serve as a foundation for a wide range of statistical analyses, including:
  • Bayesian conditional probabilities: Directly calculating the likelihood of one concept given another.
  • Association rules: Identifying patterns like "if a patient has X, they are likely to have Y."
    • Test statistics: Computing relative risks and odds ratios for comorbidity and drug interactions.
  • Detecting Subtle Patterns: The sheer scale of the data allows for the detection of subtle co-occurrence patterns that might be missed in smaller studies or manual analyses.

Key Results

  • The processed dataset contains 3 million terms mapped onto 1 million clinical concepts.
  • The text processing pipeline achieved 74% sensitivity and 96% specificity in identifying clinical conditions, as validated against a gold standard corpus.
  • The co-frequency counting process, while computationally demanding, was successfully parallelized. For example, a single pass over 220 billion pairs took approximately 9.5 hours on a 20-core machine.
  • The resulting dataset can be used to quantitatively assess comorbidity, drug-drug, and drug-disease patterns, with applications in clinical, epidemiological, and financial domains.

Limitations and Future Work

  • Data Access Restrictions: The authors acknowledge that privacy and security concerns limit the sharing of raw EHR data, necessitating the provision of processed co-occurrence counts instead.
  • Ambiguous Terms: While efforts were made to resolve ambiguity, some terms may still have multiple interpretations, which could affect the accuracy of concept mapping.
  • Future Work: The authors suggest that this dataset can be leveraged for:
  • Building drug-drug, drug-disease, and disease-disease networks.
  • Improving feature selection for high-dimensional propensity scoring.
  • Assessing data assumptions in statistical methods like regression analysis.
  • Enhancing enrichment analyses of patient cohorts by accounting for dependencies between annotations.
  • Developing more robust clinical decision support systems.

Summary updated

Abstract

Electronic health records (EHR) represent a rich and relatively untapped resource for characterizing the true nature of clinical practice and for quantifying the degree of inter-relatedness of medical entities such as drugs, diseases, procedures and devices. We provide a unique set of co-occurrence matrices, quantifying the pairwise mentions of 3 million terms mapped onto 1 million clinical concepts, calculated from the raw text of 20 million clinical notes spanning 19 years of data. Co-frequencies were computed by means of a parallelized annotation, hashing, and counting pipeline that was applied over clinical notes from Stanford Hospitals and Clinics. The co-occurrence matrix quantifies the relatedness among medical concepts which can serve as the basis for many statistical tests, and can be used to directly compute Bayesian conditional probabilities, association rules, as well as a range of test statistics such as relative risks and odds ratios. This dataset can be leveraged to quantitatively assess comorbidity, drug-drug, and drug-disease patterns for a range of clinical, epidemiological, and financial applications.
BibTeX
@inproceedings{finlayson2014building,
  author = {Samuel G. Finlayson and Paea LePendu and Nigam H. Shah},
  title = {{Building the graph of medicine from millions of clinical narratives}},
  booktitle = {Scientific},
  year = {2014},
  doi = {10.1038/sdata.2014.32},
}
Figure