A Bayesian Survival Analysis on Long COVID and non Long COVID patients: A Cohort Study Using National COVID Cohort Collaborative (N3C) Data

Sihang Jiang, Johanna Loomba, Andrea Zhou, Suchetha Sharma, Saurav Sengupta, Jiebei Liu, Donald D. BrownmedRxiv 2024 · DOI

Abstract

Since the outbreak of the COVID-19 pandemic in 2020, research has increasingly focused on the long-term effects of infection. In October 2021, the CDC introduced the ICD-10-CM code 'U09.9' for 'Post COVID-19 condition, unspecified,' formally recognizing Long COVID as a chronic illness. Leveraging the National COVID Cohort Collaborative (N3C)—which aggregates harmonized electronic health records (EHR) across the United States—this study performs a Bayesian survival analysis to compare outcomes between Long COVID patients and COVID-positive controls.

Introduction: Addressing the Long-Term Impact of COVID-19

While the acute phase of COVID-19 is well-documented, the "long tail" of the disease remains a significant clinical and public health challenge. The introduction of the U09.9 diagnostic code provided a standardized way to track these patients, yet large-scale survival analyses remain scarce.

Studying Long COVID presents unique statistical hurdles, primarily due to data imbalance (only a small fraction of patients receive a formal diagnosis) and selection bias (Long COVID patients often have distinct baseline health profiles). This research addresses these issues by using a statistically rigorous Bayesian framework to compare survival lengths while controlling for confounding factors like age, pre-existing comorbidities, and initial infection severity.

Methods: Bayesian Survival Modeling

The researchers employed a Bayesian Survival Analysis approach, allowing for the incorporation of prior knowledge and the generation of probabilistic distributions rather than simple point estimates.

Cohort Matching

To manage the scale of the N3C dataset, the team implemented 1-on-1 matching. Each patient diagnosed with Long COVID was paired with a control patient based on age, clinical site, COVID index date, frequency of clinical visits, and the Charlson Comorbidity Index (CCI). This process resulted in a balanced final cohort of 119,466 patients.

Table showing the baseline characteristics of the matched COVID positive control group and Long COVID group.

Statistical Approach

  • Parametric Modeling: A Log-normal model was used to describe survival time, assuming the logarithm of survival follows a normal distribution.
  • Estimation: Parameters were estimated using Markov Chain Monte Carlo (MCMC) sampling via the PyMC library.
  • Covariates: The model evaluated three primary factors: Long COVID status (U09.9 code), obesity (BMI $\ge$ 30), and "mild" initial infection (defined as no ER visit or hospitalization during the initial infection).

Datasets: The N3C Repository

The study draws from the National COVID Cohort Collaborative (N3C), one of the largest collections of clinical COVID-19 data in the US, containing over 7.3 million COVID-positive patients and 20 billion rows of EHR data.

Table detailing the inclusion criteria and patient counts for the Long COVID group, starting from the initial N3C pool.

The final analysis focused on survival tracked from the date of the U09.9 diagnosis (or matched equivalent) until death or the study's conclusion on April 1, 2024.

Key Insights and Results

The study demonstrates that the implementation of a specific ICD-10 code for Long COVID has made large-scale EHR research significantly more feasible, allowing researchers to move beyond small clinical surveys to massive cohort studies.

  • Long COVID Impact: The coefficient for Long COVID ($\beta_1$) was -0.009, indicating a statistically significant association with shorter survival times compared to matched controls.
  • Initial Severity: A "mild" initial infection ($\beta_3 = 0.094$) was the strongest predictor of longer survival, suggesting that acute-phase trauma (hospitalization/ER visits) continues to impact recovery trajectories long after the initial infection.
  • Obesity: Interestingly, obesity ($\beta_2 = 0.005$) was not found to be a statistically significant predictor in this specific survival model, as its 95% high-density interval crossed zero ([-0.004, 0.014]).

Discussion and Limitations

While the study provides robust evidence of the survival impact of Long COVID, several limitations exist:

  1. Censoring: The model did not explicitly account for right-censoring, treating the study end-date as a fixed boundary.
  2. Distributional Fit: The survival data exhibited left-skewness; future iterations might benefit from exploring Weibull or Gamma distributions to better fit this skewness.
  3. Under-reporting: The reliance on the U09.9 code may lead to "contamination" of the control group if undiagnosed Long COVID patients are included, potentially diluting the observed effect size.

Despite these factors, the research underscores that Long COVID represents a significant mortality risk factor that persists long after the acute phase of the illness.

Summary updated

Abstract

Since the outbreak of COVID-19 pandemic in 2020, numerous researches and studies have focused on the long-term effects of COVID infection. The Centers for Disease Control (CDC) implemented an additional code into the International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM) for reporting 'Post COVID-19 condition, unspecified (U09.9)' effective on October 1st 2021, representing that Long COVID is a real illness with potential chronic conditions. The National COVID Cohort Collaborative (N3C) provides researchers with abundant electronic health records (EHR) data by aggregating and harmonizing EHR data across different clinical organizations in the United States, making it convenient to build up a survival analysis on Long COVID patients and non Long COVID patients among large amounts of COVID positive patients.
BibTeX
@misc{jiang2024bayesian,
  author = {Sihang Jiang and Johanna Loomba and Andrea Zhou and Suchetha Sharma and Saurav Sengupta and Jiebei Liu and Donald D. Brown},
  title = {{A Bayesian Survival Analysis on Long COVID and non Long COVID patients: A Cohort Study Using National COVID Cohort Collaborative (N3C) Data}},
  year = {2024},
  doi = {10.1101/2024.06.25.24309478},
  howpublished = {medRxiv},
}
Figure