Columbia Open Health Data for COVID-19 Research: Database Analysis

Junghwan Lee, Jae Hyun Kim, Cong Liu, George Hripcsak, Karthik Natarajan, Casey Ta, Chunhua WengJournal of Medical Internet Research 2021 · DOI

Abstract

We present COHD-COVID, a publicly accessible database providing useful clinical data for hospitalized patients with COVID-19, hospitalized patients with influenza, and general hospitalized patients. We expect COHD-COVID to provide researchers and clinicians quantitative measures of COVID-19-related clinical features to better understand and combat the pandemic.

Introduction

Clinical research during the COVID-19 pandemic has been significantly hindered by a "data bottleneck." While Electronic Health Records (EHRs) contain vital information regarding disease progression and treatment efficacy, this data is typically protected by strict privacy regulations such as HIPAA. De-identification is often insufficient, as unique combinations of rare conditions can lead to "re-identification" risks. To overcome this, researchers at Columbia University have launched COHD-COVID, a publicly accessible database that provides aggregated, privacy-preserved clinical data from thousands of hospitalized patients.

Methodology

The researchers developed a robust pipeline to transform raw EHR data into a standardized, shareable resource. The process involves four primary stages:

  1. Standardization: Clinical data was converted into the OMOP Common Data Model (CDM). This standard format ensures that the data is interoperable and can be easily compared with datasets from other institutions. To manage data sparsity, the team used hierarchical mapping to group specific clinical concepts (e.g., brand-name drugs) into broader "ancestor" concepts (e.g., active ingredients).
  2. Cohort Selection: Three distinct cohorts were defined for comparative analysis: a COVID-19 cohort (patients hospitalized during the pandemic), an influenza cohort (pre-pandemic), and a general hospitalized cohort (pre-pandemic).
  3. Privacy Protection: This is the core innovation of the project. To prevent re-identification, the team excluded all clinical concepts occurring in 10 or fewer visits. Furthermore, they applied Poisson Randomization to perturb the true counts of clinical events, adding statistical uncertainty that protects individual privacy while maintaining the overall research utility of the data.
  4. Accessibility: The final dataset is hosted via a public API and the Figshare repository, allowing researchers worldwide to query concept prevalence and co-occurrences.
The overall workflow for creating the COHD-COVID database, from raw EHR data to public API.

Dataset Composition

The study utilized the EHR database from NewYork-Presbyterian/Columbia University Irving Medical Center (NYP/CUIMC), a major medical center that served as an early epicenter of the pandemic. The cohorts included:

  • COVID-19 Cohort: 4,127 patients (4,846 visits) between March and August 2020.
  • Influenza Cohort: 3,261 patients (3,860 visits) from 2014–2019.
  • General Cohort: 175,930 patients (314,680 visits) from 2014–2019.

Clinical Insights and Results

The analysis successfully validated known clinical hallmarks of COVID-19. Fever, cough, and dyspnea (shortness of breath) showed significantly higher prevalence in the COVID-19 cohort compared to both the general and influenza baselines.

Comparison of condition prevalence across COVID-19, Influenza, and General cohorts.

Key quantitative findings include:

  • Prevalence Ratios: By calculating the Prevalence Ratio (PR), the authors demonstrated that "Acute lower respiratory tract infection" was 7.3 times more likely in COVID-19 patients than in the general hospitalized population.
  • Drug Utilization: Acetaminophen was the most common medication (0.7912 prevalence), followed by Enoxaparin (0.5803), a blood thinner used to manage COVID-related coagulopathy.
  • Demographic Trends: Senior patients (>65 years) exhibited significantly higher rates of essential hypertension and hypoxemia compared to younger adults, quantifying the increased clinical burden on the elderly.

Discussion and Limitations

While COHD-COVID provides a powerful tool for pandemic research, it is limited by its single-site nature. Clinical practices and patient demographics at NYP/CUIMC may not be representative of all regions. Additionally, the database uses visit-based counts rather than patient-based counts, meaning the metrics reflect clinical events per hospital visit rather than longitudinal patient history. Future work will focus on integrating data from multiple institutions to enhance the database's diversity and predictive accuracy.

Summary updated

Abstract

We present COHD-COVID, a publicly accessible database providing useful clinical data for hospitalized patients with COVID-19, hospitalized patients with influenza, and general hospitalized patients. We expect COHD-COVID to provide researchers and clinicians quantitative measures of COVID-19-related clinical features to better understand and combat the pandemic.
BibTeX
@article{lee2021columbia,
  author = {Junghwan Lee and Jae Hyun Kim and Cong Liu and George Hripcsak and Karthik Natarajan and Casey Ta and Chunhua Weng},
  title = {{Columbia Open Health Data for COVID-19 Research: Database Analysis}},
  journal = {Journal of Medical Internet Research},
  year = {2021},
  doi = {10.2196/31122},
}
Figure