A Retrospective Longitudinal Study of COVID-19 as Seen by a Large Urban Hospital in Chicago

Haotian Chen, Yogatheesan Varatharajah, Sarah Stewart de Ramirez, Paul M. Arnold, Casey Frankenberger, Bala Hota, Ravishankar K. IyermedRxiv 2020 · DOI

Abstract

The rapid spread of COVID-19 has placed unprecedented demand on medical resources, particularly intensive care unit (ICU) beds and mechanical ventilators. Given the disease's diverse clinical manifestations, identifying patients at high risk for severe progression is critical for effective resource management. This study presents a retrospective analysis of 10,123 COVID-19 patients treated at Rush University Medical Center in Chicago. We developed a staging scheme based on clinical milestones and a prognostic model to predict mechanical ventilation requirements upon hospital admission. Our findings identify a compact set of biomarkers—albumin, oxygen saturation (SpO2), and white blood cell count—that predict ventilator demand with an AUC of 0.823.

Introduction

The COVID-19 pandemic challenged healthcare systems to differentiate between patients who would recover with minimal intervention and those who would rapidly deteriorate. This "remarkable diversity" in disease progression made early clinical decision-making difficult. This research aims to provide a clinical roadmap by staging disease progression through discrete milestones (Hospitalization → ICU → Ventilation → Death) and developing an early warning system to predict ventilation needs using data available immediately upon hospital admission.

Methods and Data

Dataset

The study utilized Electronic Medical Record (EMR) data from Rush University Medical Center, covering 10,123 COVID-19 positive patients between March 17 and August 7, 2020. Of this total cohort, 1,788 were hospitalized. The model development focused on a subset of 301 hospitalized patients who had complete laboratory and vital sign data on their day of admission. Notably, the Chicago-based cohort reflected a diverse urban population with a median age of 40, younger than many cohorts reported early in the pandemic.

Analytic Approach

The researchers employed a three-stage methodology:

  1. Temporal Analysis: Tracking the 15-day trajectory of laboratory values to correlate trends with clinical outcomes.
  2. Feature Selection: Utilizing LASSO (Least Absolute Shrinkage and Selection Operator) regression and Decision Trees (Gini Index) to identify the most predictive variables from a wide array of clinical data.
  3. Prognostic Modeling: Developing a Logistic Regression model to predict future ventilation needs. Logistic regression was specifically chosen over more complex "black box" algorithms to ensure the model's interpretability for clinicians.

Key Insights

The Role of Albumin

One of the most significant findings was the discriminative power of albumin, a protein produced by the liver. While most patients experienced an initial decline in albumin levels, survivors typically showed recovery after day seven. In contrast, patients who eventually died exhibited a continuous, progressive decline in albumin levels.

A Compact Predictor Set

While the study initially evaluated 18 variables, the researchers discovered that a "compact" set of just three biomarkers—albumin, SpO2, and white blood cell count—performed as well as, or slightly better than, more complex models. This simplification is vital for clinical utility, allowing for rapid assessment in high-pressure emergency department environments.

Comorbidity and Risk

The study observed a "climbing" prevalence of specific comorbidities as patients progressed toward more severe stages. Chronic kidney disease and chronic ischemic heart disease were identified as the most lethal underlying conditions within this cohort.

Results

The final prognostic model demonstrated strong discriminative performance:

  • Predictive Accuracy: The model achieved an AUC of 0.823 (95% CI: 0.765–0.880).
  • Clinical Utility: At an optimal cutpoint of 0.35, the model yielded a sensitivity of 0.70 and a specificity of 0.82.
  • Validation: In the study cohort, the model correctly identified 67 out of 96 patients who eventually required mechanical ventilation.

Additionally, the data highlighted significant demographic disparities, with African American (32.92%) and Hispanic/Latino (50.88%) patients representing the largest segments of the infected population in the Chicago area.

Discussion and Limitations

This study provides a data-driven framework for predicting COVID-19 severity, yet several limitations remain. As a single-center study, the findings reflect the specific practice patterns and demographics of one Chicago hospital system. Furthermore, the current model does not incorporate radiographic data (X-rays or CT scans), which are standard in assessing pulmonary involvement. Future work will focus on prospective validation to test the model's efficacy as a real-time early warning system in live clinical environments.

Summary updated

Abstract

Abstract The rapid spread of the novel coronavirus disease 2019 (COVID-19) has created high demand for medical resources, including personnel, intensive care unit beds, and ventilators. As thousands of patients are hospitalized, the disease has shown remarkable diversity in its manifestation; many patients with mild to no symptoms recover from the disease requiring minimal care, but some patients with severe disease progression require mechanical ventilation support in intensive care units (ICU) with an increased risk of death. Studying the characteristics of patients in these various strata can help us understand the varied progression of this disease, enable earlier interventions for at-risk patients, and help manage medical resources more efficiently. This paper presents a retrospective analysis of 10,123 COVID-19 patients treated at the Rush University Medical Center in Chicago, including their demographics, symptoms, comorbidities, laboratory values, vital signs, and clinical history. Specifically, we present a staging scheme based on discrete clinical events (i.e., admission to the hospital, admission to the ICU, mechanical ventilation, and death), and investigate the temporal trend of clinical variables and the effect of comorbidities in each of those stages. We then developed a prognostic model to predict ventilation demands at an individual patient level by analyzing baseline clinical variables, which entails (1) a least absolute shrinkage and selection operator (LASSO) regression and a decision tree model to identify predictors for mechanical ventilation; and (2) a logistic regression model based on these risk factors to predict which patients will eventually need ventilatory support. Our results indicate that the prognostic model achieves an AUC of 0.823 (95% CI: 0.765–0.880) in identifying patients who will eventually require mechanical ventilation.
BibTeX
@misc{chen2020retrospective,
  author = {Haotian Chen and Yogatheesan Varatharajah and Sarah Stewart de Ramirez and Paul M. Arnold and Casey Frankenberger and Bala Hota and Ravishankar K. Iyer},
  title = {{A Retrospective Longitudinal Study of COVID-19 as Seen by a Large Urban Hospital in Chicago}},
  year = {2020},
  doi = {10.1101/2020.11.29.20240606},
  howpublished = {medRxiv},
}
Figure