Vital Measurements of Hospitalized COVID-19 Patients as a Predictor of Long COVID: An EHR-based Cohort Study from the RECOVER Program in N3C

Sihang Jiang, Johanna Loomba, Suchetha Sharma, Donald E. Brown

Abstract

Various symptoms can persist in the post-acute sequelae of SARS-CoV-2 infection (PASC), commonly known as Long COVID. These heterogeneous symptoms significantly hinder recovery from the pandemic. While a precise definition of Long COVID is still being refined, predicting its occurrence is vital for understanding the disease's characteristics. This study analyzes vital measurements—including oxygen saturation, heart rate, and blood pressure—recorded during the first seven days of hospitalization. We investigate whether patterns in these basic physiological functions can serve as early predictors for Long COVID in hospitalized patients.

Introduction

While the acute phase of COVID-19 is well-documented, Long COVID remains an elusive condition. Predicting who will develop lingering symptoms like profound fatigue or breathlessness is challenging, as many early indicators rely on subjective reporting. This research addresses a gap in the literature by leveraging objective, frequently collected vital signs during hospitalization. By analyzing the variability and patterns of these vitals, we aim to establish an early warning system for PASC.

Methods and Data

The study utilized a retrospective cohort design based on the National COVID Cohort Collaborative (N3C) database, one of the largest clinical data resources in the US. The initial pool included over 5.2 million COVID-positive patients. Two subcohorts were defined: Subcohort A (16,468 patients) for primary XGBoost analysis and Subcohort B (5,304 patients) for time-series modeling.

Distribution of vital sign readings across the patient cohort, showing the frequency of measurements for oxygen saturation, heart rate, and blood pressure.

Researchers engineered 139 features from four vital signs: SpO2, Heart Rate, Systolic BP, and Diastolic BP. These features captured daily averages, summary statistics (min, max, median, skewness, kurtosis), and daily variability. Because Long COVID is often under-coded, a machine-learning-based "computable phenotype" was used to identify cases (likelihood score $\ge 0.75$).

Demographic distribution of age at COVID diagnosis and the length of hospitalization stay for the study cohort.

Results

The study found that vital measurements are powerful predictors of Long COVID. An XGBoost model combining vitals with social determinants of health (SDOH) and pre-existing conditions achieved a Mean AUC of 0.822 and an F1 score of 0.752. Notably, vital measurements alone (AUC 0.787) outperformed models based solely on patient history and SDOH (AUC 0.729).

Key findings include:

  • Variability is Key: The stability of vitals—captured through standard deviation and skewness—was more informative than simple snapshots.
  • Summary Statistics vs. Time-Series: XGBoost models trained on aggregate summary statistics outperformed deep learning models (CNN/LSTM) designed for raw time-series data. This suggests that the overall physiological "stress" over the first week is highly predictive.
  • Top Predictors: The duration of systolic blood pressure monitoring, heart rate trends within the first three days, and a history of chronic lung disease were among the most significant features.

Discussion and Conclusion

This research demonstrates that objective physiological markers from early hospitalization can effectively predict Long COVID. However, limitations exist, including data sparsity for consistent time-series modeling and the reliance on a computable phenotype for labeling. Future work should integrate more diverse data types, such as laboratory results and medications, to further refine these predictive capabilities and explore the underlying biological mechanisms.

Summary updated

BibTeX
@inproceedings{jiang2022vital,
  author = {Sihang Jiang and Johanna Loomba and Suchetha Sharma and Donald E. Brown},
  title = {{Vital Measurements of Hospitalized COVID-19 Patients as a Predictor of Long COVID: An EHR-based Cohort Study from the RECOVER Program in N3C}},
  booktitle = {BIBM},
  year = {2022},
  doi = {10.1109/bibm55620.2022.9995311},
}
Figure