Separating Wheat from Chaff: Joining Biomedical Knowledge and Patient Data for Repurposing Medications

Galia Nordon, Gideon Koren, Varda Shalev, Eric Horvitz, Kira Radinsky

Introduction: The Challenge of Drug Repurposing

The pharmaceutical industry is currently grappling with "Eroom's Law"—the observation that drug discovery is becoming progressively slower and more expensive despite technological advancements. Developing a new drug typically requires 14 years and billions of dollars in investment. Drug repurposing—identifying new therapeutic uses for existing, approved medications—offers a significantly faster and more cost-effective alternative, as these compounds have already cleared human safety trials.

However, identifying viable repurposing candidates is a significant challenge. While Electronic Medical Records (EMR) provide a wealth of patient data, they are rife with spurious correlations. For instance, researchers found that purchasing bandages correlates with successful hypertension treatment. Clearly, bandages do not cure high blood pressure; they serve as a proxy for health-conscious behavior. The central problem is "separating the wheat from the chaff": distinguishing coincidental correlations from genuine biological mechanisms.

Methods: A Dual-Stage Pipeline

The authors propose a Data-Driven Medical Model designed to automate the clinical intuition of a researcher across millions of data points. The pipeline integrates two distinct stages to validate potential drug-disease links.

The proposed data-driven medical model pipeline combining EMR correlation discovery with knowledge graph reasoning.

The system first scans EMR data for patients prescribed a drug for one condition who subsequently showed improvement in an unrelated, "off-label" condition. To mitigate biases related to age or lifestyle, the researchers employ Propensity Score Matching. This involves creating a control group of untreated patients matched by age, weight, BMI, and sex, followed by Chi-squared tests to determine if the drug's effect is statistically significant.

2. Correction by Reasoning (The "Chaff" Filter)

To filter out spurious correlations (the "bandage" effect), the system consults a Biomedical Knowledge Graph constructed from SemMedDB, which contains 90 million semantic triples extracted from PubMed.

  • Path Search: The system looks for biological pathways connecting the drug to the disease (e.g., Statin → Atrial Natriuretic Peptide → Hypertension).
  • Semantic Pruning: To prevent the reasoning from becoming too broad, the authors implement algorithms to prune paths that are excessively long or contain overly general nodes (e.g., "Vitamins" or "United States"). They utilize the Unified Medical Language System (UMLS) to identify and exclude these non-specific semantic types.

Datasets

The research leverages three primary data sources:

  • Maccabi Healthcare Services: A longitudinal dataset of over 1.5 million patients in Israel, encompassing over a decade of clinical records.
  • PubMed / SemMedDB: A repository of 27 million citations used to build the foundational knowledge graph.
  • UMLS (Unified Medical Language System): Used for semantic filtering to ensure the reasoning remains focused on biological mechanisms.

Key Insights: Knowledge as a Validator

A critical insight from this work is the "Bandage Paradox": Big Data, without a biological sanity check, can easily lead to erroneous medical conclusions. Rather than using the knowledge graph solely for discovery, the authors use it as a validator. If a correlation appears in the EMR but lacks a plausible biological pathway in the literature, the candidate is deprioritized. Furthermore, the use of UMLS semantic types allows the system to recognize that a path through a "Hormone" or "Enzyme" is far more valuable than one through a generic "Physical Object."

Key Results

The system was evaluated on two major conditions: hypertension and diabetes.

Precision metrics comparing the proposed algorithm with baseline models across hypertension and diabetes test cases.
  • Hypertension: The system identified Statins (cholesterol medication) and Proton-pump inhibitors (acid reflux medication) as potential treatments. While the Statin link was known to some specialists, the system uniquely identified a specific causal pathway involving the ANP hormone.
  • Diabetes: The system identified Alpha blockers (typically for prostate issues) as a candidate.
  • Precision Performance: In the hypertension test case, the system achieved 100% Precision@2. In contrast, baseline models that relied solely on correlations achieved 0% precision at the same level, as their top results were dominated by "chaff" like bandages and gauze.

Limitations and Future Work

While promising, the system has limitations. EMR data is observational and may contain hidden confounders despite propensity matching. Additionally, the system is limited by the current state of medical literature; it cannot validate drugs that work via entirely undiscovered biological pathways.

Future work involves developing this framework into an interactive tool for medical researchers to visualize reasoning paths and generate hypotheses for clinical trials. Currently, the system is already being utilized by researchers at Maccabi Healthcare to explore new therapeutic avenues.

Summary updated

Abstract

We present a system that jointly harnesses large-scale electronic health records data and a concept graph mined from the medical literature to guide drug repurposing—the process of applying known drugs in new ways to treat diseases. Our study is unique in methods and scope, per the scale of the concept graph and the quantity of data. We harness 10 years of nation-wide medical records of more than 1.5 million people and extract medical knowledge from all of PubMed, the world’s largest corpus of online biomedical literature. We employ links on the concept graph to provide causal signals to prioritize candidate influences between medications and target diseases. We show results of the system on studies of drug repurposing for hypertension and diabetes. In both cases, we present drug families identified by the algorithm which were previously unknown. We verify the results via clinical expert opinion and by prospective clinical trials on hypertension.
BibTeX
@inproceedings{nordon2019separating,
  author = {Galia Nordon and Gideon Koren and Varda Shalev and Eric Horvitz and Kira Radinsky},
  title = {{Separating Wheat from Chaff: Joining Biomedical Knowledge and Patient Data for Repurposing Medications}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33019565},
}
Figure