Abstract
Massive research efforts have been made in response to the COVID-19 (coronavirus disease-2019) pandemic. Utilization of clinical data can accelerate these research efforts to fight against the pandemic since important characteristics of the patients are often found by examining the clinical data. To provide shareable clinical data to catalyze COVID-19 research, we present Columbia Open Health Data for COVID-19 Research (COHD-COVID), a publicly accessible database providing clinical concept prevalence, clinical concept co-occurrence, and clinical symptom prevalence for hospitalized COVID-19 patients. COHD-COVID also provides data on hospitalized influenza patients and general hospitalized patients as comparator cohorts. The data used in COHD-COVID were obtained from Columbia University Irving Medical Center’s electronic health records. We expect COHD-COVID will provide researchers and clinicians quantitative measures of COVID-19 related clinical features to better understand and fight against the pandemic.
Introduction
During the initial waves of the COVID-19 pandemic, the scientific community faced a critical bottleneck: access to high-quality, granular clinical data. While epidemiological trends were widely reported, the specific clinical interactions—such as which drugs were most frequently prescribed alongside specific diagnoses—remained largely siloed within hospital systems due to HIPAA regulations and privacy concerns.
To address this, researchers developed COHD-COVID (Columbia Open Health Data for COVID-19 Research). This publicly accessible database provides aggregated, privacy-preserving clinical statistics derived from the electronic health records (EHR) of over 4,000 patients at the Columbia University Irving Medical Center (CUIMC). By offering standardized data on clinical concept prevalence and co-occurrence, COHD-COVID enables researchers to compare the clinical features of COVID-19 against influenza and general hospital populations.
Methodology
Data Standardization and Privacy
The researchers transformed raw EHR data into the OMOP Common Data Model (CDM), a standardized format that facilitates cross-institutional research. To ensure the data could be shared publicly without compromising patient privacy, several key steps were taken:
- Hierarchical Aggregation: Clinical concepts were aggregated into higher-level categories (e.g., mapping specific dosages to a general drug ingredient) to account for variations in hospital coding.
- Poisson Randomization: To protect patient identity, the researchers excluded any clinical concept appearing in 10 or fewer visits and applied Poisson randomization to "perturb" the true counts. This adds a small amount of statistical noise that protects privacy while maintaining the data's scientific utility.
- Visit-Based Counting: The database utilizes visit-based counts rather than unique patient counts. This approach reduces bias for cohorts with longer observation windows, ensuring a fairer comparison between acute COVID-19 hospitalizations and chronic patients seen over several years.
Cohort Definitions
The study utilizes three primary cohorts derived from CUIMC’s records:
- COVID-19 Cohort: 4,127 hospitalized patients (March 1 – Sept 1, 2020).
- Influenza Cohort: 3,261 patients hospitalized with flu (2014–2019) to serve as a viral respiratory baseline.
- General Cohort: 175,930 patients (2014–2019) to serve as a baseline for standard hospitalizations.
Key Clinical Insights
The Stroke Connection and Respiratory Distress
One of the most significant findings in the COHD-COVID dataset is the high prevalence ratio of Cerebral infarction (stroke) in COVID-19 patients compared to influenza patients. This data supports early clinical observations that SARS-CoV-2 causes unique thrombotic (clotting) issues not typically seen in seasonal respiratory viruses.
Furthermore, by grouping granular clinical codes into 11 major symptoms (e.g., Dyspnea, Fever, Myalgia), the researchers confirmed that while COVID-19 shares traits with the flu, its prevalence of acute respiratory distress is significantly higher.
Pharmacological Trends
The dataset captures the evolving clinical strategies of the pandemic's first wave. Acetaminophen emerged as the most common drug ingredient, reflecting the primary strategy of managing high fevers. The data also captures the early, widespread use of Hydroxychloroquine, showing high co-occurrence with respiratory infection diagnoses during the initial months of 2020.
Addressing Data Artifacts
During the study period, CUIMC transitioned its EHR system from Allscripts to Epic. To ensure this shift did not introduce "data artifacts" or skew the results, the authors performed rigorous t-tests, confirming that the clinical trends observed were a result of disease pathology rather than changes in documentation software.
Results and Scale
The resulting database is substantial, including 1,066 distinct clinical concepts and nearly 50,000 concept pairs (co-occurrences). Key findings include:
- Prevalence: Fever, Cough, and Dyspnea were the most prevalent conditions among hospitalized COVID-19 patients.
- Comparative Ratios: COVID-19 patients showed significantly higher ratios for Acute Respiratory Distress Syndrome (ARDS) compared to the general hospital population.
Limitations and Future Directions
While COHD-COVID provides a robust resource, it is limited by its single-site nature. The data reflects the specific demographics and clinical practices of New York City during the early pandemic. Additionally, like all EHR-based research, the data is dependent on the accuracy of clinician coding.
Future work aims to expand this model to multiple institutions, creating a multi-site database that can validate these findings on a broader scale. Currently, the data is available via tab-delimited files on Figshare and through a dedicated web API (http://covid.cohd.io).
Summary updated