Clinical subphenotypes in COVID-19: derivation, validation, prediction, temporal patterns, and interaction with social determinants of health

Chang Su, Chang Su 0002, Yongkang Zhang, James H. Flory, Mark G. Weiner, Rainu Kaushal, Edward J. Schenck, Fei WangNpj Digit. Medicine 2021 · DOI

TL;DR

This paper identifies four distinct clinical subphenotypes of COVID-19 using machine learning on routinely collected patient data. These subphenotypes are predictive of clinical outcomes, show distinct temporal patterns throughout the pandemic, and importantly, reveal how social determinants of health disproportionately impact the most severe subphenotype.

What Problem Does This Paper Address?

The COVID-19 pandemic has revealed a wide spectrum of disease severity and clinical manifestations. Understanding the underlying biological mechanisms driving these differences is crucial for effective treatment, clinical trial design, and public health interventions. However, the heterogeneity of COVID-19 has made it challenging to develop a unified understanding. Previous attempts to identify distinct patient subgroups (subphenotypes) have been limited by cohort size, data availability, and a lack of robust validation. Furthermore, the influence of social determinants of health (SDoH) on these potential subphenotypes and their outcomes has not been thoroughly investigated. This paper aims to address this by deriving and validating COVID-19 subphenotypes, examining their temporal evolution, and exploring their interaction with SDoH.

Methods

The study employed a multi-stage machine learning approach.

  • Subphenotype Derivation: Agglomerative hierarchical clustering was used on presenting clinical data from a development cohort to identify distinct patient subgroups.
  • Validation: Robustness and usability of the derived subphenotypes were assessed through several strategies:
  • Sensitivity Analyses: The development cohort was re-analyzed with variations in data quality control and clustering methods (including Gaussian Mixture Models) to confirm the stability of the 4-cluster structure.
  • Internal Validation: Subphenotypes were re-derived in a separate internal validation cohort using the same clustering method.
  • Predictive Modeling: An XGBoost classifier was trained on the development cohort to predict subphenotype membership. This model was then used to predict subphenotypes in an external validation cohort.
  • Leave-One-Center-Out Analysis: This method was used to assess the stability of the subphenotypes across the five participating medical centers.
  • Temporal Pattern Analysis: Bar charts were used to visualize the proportion of each subphenotype over time, week by week, since the start of the pandemic in NYC.
  • SDoH Impact Analysis: Logistic regression and Cox regression analyses were used to assess the association between SDoH variables and 60-day mortality within each subphenotype. Additionally, patients were stratified into high, middle, and low SDoH groups using clustering, and their mortality rates were compared.

Datasets

The study utilized data from the INSIGHT Clinical Research Network (CRN), a collaboration of five major medical centers in New York City (NYC).

  • Development Cohort: 70% of patients from four medical centers (N=8199).
  • Internal Validation Cohort: 30% of patients from the same four medical centers (N=3519).
  • External Validation Cohort: Patients from the fifth medical center (N=2700).

The data included routinely collected clinical variables at the time of presentation to the emergency department or hospitalization, as well as demographic information, comorbidities, clinical outcomes, medications, and neighborhood-level socioeconomic characteristics (SDoH).

Key Insights and Ideas

  • Distinct COVID-19 Phenotypes: The study successfully identified four clinically meaningful subphenotypes, moving beyond a monolithic view of COVID-19. This suggests that COVID-19 is not a single disease but rather a syndrome with distinct biological underpinnings.
  • Predictive Power of Subphenotypes: The identified subphenotypes were highly predictive of clinical outcomes, particularly 60-day mortality, with Subphenotype IV showing significantly worse outcomes. This highlights their potential for risk stratification.
  • Temporal Dynamics of Subphenotypes: The prevalence of subphenotypes evolved over the course of the pandemic. Mild subphenotypes (I and II) were more common early on, while the severe subphenotype (IV) increased in proportion after the initial peak, suggesting different roles in disease spread and severity over time.
  • SDoH Exacerbates Disparities in Severe Cases: A critical finding is that SDoH disproportionately impacts the most severe subphenotype (IV). While SDoH had a less pronounced effect on milder subphenotypes, disadvantaged socioeconomic status significantly amplified mortality risk in patients with severe COVID-19. This underscores the intersectionality of biological vulnerability and social disadvantage.
  • Usability of Routine Data: The study demonstrates that robust subphenotypes can be derived and validated using readily available clinical data, making this approach applicable in real-world clinical settings and for future patient stratification.

Key Results

  • Four Subphenotypes Identified: The clustering analysis consistently identified four subphenotypes across development, internal, and external validation cohorts.
  • Outcome Prediction: Subphenotype IV was associated with the highest 60-day mortality rate (49.64%), significantly higher than Subphenotypes I (6.94%), II (17.33%), and III (22.68%).
  • Temporal Shift: Subphenotype IV's proportion increased significantly from weeks 6-9 of the pandemic, coinciding with the peak mortality in NYC.
  • SDoH Impact: In Subphenotype IV, disadvantaged SDoH strata were associated with a 55.19% mortality rate, a 5.55% increase compared to the subphenotype level. This effect was much less pronounced in other subphenotypes.
  • Predictive Model Performance: The XGBoost classifier achieved high performance in predicting subphenotype membership, demonstrating the practical utility of the derived subphenotypes.

Limitations and Future Work

  • Data Limitations: The study acknowledges that presenting symptoms and vital signs were not available, which could have provided further insights. The reliance on a snapshot of data at presentation limits the capture of disease progression.
  • Missing Data: While imputation methods were used, missing values can still introduce bias.
  • Generalizability: The study was conducted in NYC. Further validation in diverse geographical areas is needed to confirm the generalizability of these subphenotypes.
  • Longitudinal Data: The authors suggest that future work could benefit from analyzing longitudinal patient data to capture the full complexity of the disease arc and identify more dynamic subphenotypes.
  • Clinical Trial Integration: The identified subphenotypes and predictive models offer a promising avenue for improving clinical trial design by enabling more targeted patient enrollment and stratification.

Summary updated

BibTeX
@article{su2021clinical,
  author = {Chang Su and Chang Su and Yongkang Zhang and James H. Flory and Mark G. Weiner and Rainu Kaushal and Edward J. Schenck and Fei Wang},
  title = {{Clinical subphenotypes in COVID-19: derivation, validation, prediction, temporal patterns, and interaction with social determinants of health}},
  journal = {Npj Digit. Medicine},
  year = {2021},
  doi = {10.1038/s41746-021-00481-w},
}
Figure