Characterizing chronological accumulation of comorbidities in healthy veterans: a computational approach

Julian C. Hong, Elizabeth R. Hauser, Thomas S. Redding, Kellie J. Sims, Ziad F. Gellad, Meghan C. O’Leary, Terry Hyslop, Ashton Madison, Xuejun Qin, David G. Weiss, A. Jasmine Bullard, Christina D. Williams, Brian A. Sullivan, David A. Lieberman, Dawn ProvenzaleScientific 2021 · DOI

TL;DR

This study uses directed network graphs on electronic health records (EHR) from healthy veterans to map the chronological accumulation of comorbidities. By analyzing the relationships between diagnoses, they identify key "hub" diseases that often precede others and "authority" diseases that frequently follow, offering a computational approach to understand disease trajectories and inform personalized healthcare.

What Problem Does This Paper Address?

Understanding how multiple chronic conditions (comorbidities) develop and interact over time is crucial for effective healthcare. This is particularly challenging because diseases don't occur in isolation; they influence each other, creating complex health trajectories. While previous research has explored disease networks, there's a need for methods that specifically capture the chronological progression of comorbidities, especially in well-defined, longitudinal cohorts. This paper aims to address this by using a computational approach to visualize and quantify these temporal relationships, moving beyond simple co-occurrence to understand cause-and-effect-like pathways.

Methods

The core of this study is the application of directed network graphs to electronic health record (EHR) data. Here's a breakdown of their approach:

  • Data Source: They utilized EHR data from the Veterans Affairs (VA) Corporate Data Warehouse (CDW) for participants in the Veterans Affairs Cooperative Studies Program #380, a prospective longitudinal study of screening colonoscopy.
  • Diagnosis Identification: They extracted International Classification of Diseases, Ninth Edition (ICD-9) diagnosis codes. To focus on significant conditions, they filtered for diagnoses affecting at least 10% of patients and excluded non-diagnostic codes, symptoms, and complications of pregnancy. The first chronological instance of each diagnosis for each patient was identified.
  • Chronological Relationship Quantification: The key metric used is pairwise chronological relative risk (RR). This calculates the ratio of the probability of developing diagnosis 'j' after diagnosis 'i' compared to the probability of developing diagnosis 'j' without having had diagnosis 'i' previously. An RR > 1 indicates that diagnosis 'i' increases the likelihood of diagnosis 'j' occurring subsequently.
  • Network Construction: Using the calculated RRs, they built a directed graph where nodes represent diagnoses and directed edges represent relationships with RR > 1. The igraph software package was used for this.
  • Key Node Identification: They employed Kleinberg hub and authority centrality scores to identify important nodes. Hubs are diagnoses that precede many other important diagnoses, while authorities are diagnoses that are frequently preceded by important diagnoses. PageRank was also used as a measure of diagnostic importance.
  • Community Detection: The Infomap algorithm was used to identify clusters or "communities" of diagnoses that are more densely interconnected within the community than with diagnoses outside it, suggesting groups of related comorbidities.

Datasets

  • Veterans Affairs Cooperative Studies Program #380 (CSP #380) Cohort: This is a prospective longitudinal study of screening and surveillance colonoscopy.
  • VA Corporate Data Warehouse (CDW): This dataset contains electronic health records (EHR) for VA patients, including inpatient and outpatient diagnoses (ICD-9 codes), pharmacy data, and manually curated data. The study specifically used data from October 1999 to December 2015.
  • Participant Selection: The analysis focused on 2,210 healthy veterans who had at least 5 years of linked EHR history within the CDW.

Key Insights and Ideas

  • Visualizing Disease Progression: Directed network graphs provide a powerful visual tool to understand how diseases unfold over time, moving beyond static co-occurrence to dynamic pathways.
  • Identifying "Hub" and "Authority" Diseases: The identification of "hub" diseases (e.g., ischemic heart disease, diabetes) that often precede others, and "authority" diseases (e.g., acute kidney failure) that frequently follow, offers a way to prioritize interventions.
  • Uncovering Novel Hypotheses: While some relationships confirmed known clinical associations (e.g., heart disease and kidney disease), the analysis also revealed less obvious or potentially novel hypotheses about disease progression.
  • Community Structure of Comorbidities: The identification of distinct "communities" of diagnoses suggests that certain groups of conditions tend to cluster together, potentially indicating shared underlying mechanisms or common risk factors.
  • Personalized Risk Stratification: By understanding these chronological relationships, clinicians could potentially flag patients at higher risk for certain subsequent conditions and implement earlier preventative strategies.

Key Results

  • Cohort Characteristics: The study analyzed 2,210 healthy veterans with a median follow-up of 14.4 years. They had a median of 95 distinct ICD-9 diagnoses.
  • Network Structure: The network included 145 distinct ICD three-digit codes. It had an edge density of 11% and a reciprocity of 23%, indicating a limited but significant number of directional relationships.
  • Top Hub Diseases (Preceding): Ischemic heart disease (411.), inflammatory and toxic neuropathy (357.), and diabetes mellitus (250.*) were identified as key preceding diseases.
  • Top Authority Diseases (Subsequent): Acute kidney failure (584.) and hypertensive chronic kidney failure (403.) were identified as key subsequent diagnoses.
  • Community Detection: Four distinct communities of correlated comorbidities were identified, including neuropsychiatric, cerebrovascular, ear disorders, and a general "other" category.
  • Specific Pathway Examples: The study highlighted pathways such as ischemic heart disease leading to renal failure and anemia, and diabetes progressing to other diabetic complications and eventually acute kidney failure.
  • Unexpected Relationships: The analysis identified less clinically anticipated relationships, such as gout and chronic skin ulcers being associated with a high relative risk for subsequent acute kidney failure.

Limitations and Future Work

  • EHR Data Imperfections: The authors acknowledge that EHR data can be imperfect and diagnoses may be systematically misreported, limiting causal interpretations.
  • Generalizability: The findings are specific to this cohort of aging veterans within the VHA system and may not directly extrapolate to other populations.
  • Potential for Missed Events: Limiting the cohort to those with at least 5 years of follow-up might have missed rapidly fatal events.
  • Exclusion of Rare Diagnoses: Diagnoses affecting fewer than 10% of patients were excluded, potentially biasing the analysis against rare but significant conditions.
  • Future Work: The authors suggest applying these algorithms to more heterogeneous and inclusive populations. They also plan to use prospectively collected health status data to audit the quality of EHR-based data and explore how service assignments might impact generalizability. Further work could leverage network analyses to improve healthcare efficiency and prioritize high-impact diagnoses.

Summary updated

Abstract

Understanding patient accumulation of comorbidities can facilitate healthcare strategy and personalized preventative care. We applied a directed network graph to electronic health record (EHR) data and characterized comorbidities in a cohort of healthy veterans undergoing screening colonoscopy. The Veterans Affairs Cooperative Studies Program #380 was a prospective longitudinal study of screening and surveillance colonoscopy. We identified initial instances of three-digit ICD-9 diagnoses for participants with at least 5 years of linked EHR history (October 1999 to December 2015). For diagnoses affecting at least 10% of patients, we calculated pairwise chronological relative risk (RR). iGraph was used to produce directed graphs of comorbidities with RR > 1, as well as summary statistics, key diseases, and communities. A directed graph based on 2210 patients visualized longitudinal development of comorbidities. Top hub (preceding) diseases included ischemic heart disease, inflammatory and toxic neuropathy, and diabetes. Top authority (subsequent) diagnoses were acute kidney failure and hypertensive chronic kidney failure. Four communities of correlated comorbidities were identified. Close analysis of top hub and authority diagnoses demonstrated known relationships, correlated sequelae, and novel hypotheses. Directed network graphs portray chronologic comorbidity relationships. We identified relationships between comorbid diagnoses in this aging veteran cohort. This may direct healthcare prioritization and personalized care.
BibTeX
@inproceedings{hong2021characterizing,
  author = {Julian C. Hong and Elizabeth R. Hauser and Thomas S. Redding and Kellie J. Sims and Ziad F. Gellad and Meghan C. O’Leary and Terry Hyslop and Ashton Madison and Xuejun Qin and David G. Weiss and A. Jasmine Bullard and Christina D. Williams and Brian A. Sullivan and David A. Lieberman and Dawn Provenzale},
  title = {{Characterizing chronological accumulation of comorbidities in healthy veterans: a computational approach}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-85546-2},
}
Figure