A simple method to predict future COVID-19 hospitalization

By Michael Greenwood, M.Sc.Reviewed by Danielle Ellis, B.Sc.Oct 11 2021

Given the widespread and impactful nature of the coronavirus disease 2019 (COVID-19) pandemic, many models have been developed to predict various disease metrics such as the number of expected infections and resulting hospitalizations, allowing healthcare services to prepare for such outbreaks.

Study: COVIDNearTerm: A Simple Method to Forecast COVID-19 Hospitalizations. Image Credit: ffikretow/ Shutterstock

This news article was a review of a preliminary scientific report that had not undergone peer-review at the time of publication. Since its initial publication, the scientific report has now been peer reviewed and accepted for publication in a Scientific Journal. Links to the preliminary and peer-reviewed reports are available in the Sources section at the bottom of this article. View Sources

In a paper recently uploaded to the preprint server medRxiv*, a simple model with the ability to forecast probable COVID-19 hospitalization numbers in the short term is described, requiring less input information and generating more accurate predictions than other models.

Background

Early models developed in the beginning stages of the COVID-19 pandemic were generally focused on assessing the impact of non-pharmaceutical interventions. In contrast, the model developed here, termed COVIDNearTerm, is intended to forecast COVID-19 hospital numbers two to four weeks in the future.

Other short-term prediction models have been developed that rely on the observed rate of transition of the virus between groups or other values that are strongly prone to fluctuation and thus require a significant quantity of data collection in an unprecedented situation.

The authors set out to make a prediction model that avoided many of these sources of variability, depending instead on hospital-based data such as admission trends in recent weeks to establish the probability of COVID-19 outbreak.

The authors note that most severe severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infections result in hospitalization three to ten days following the appearance of first symptoms. Thus a short-term method of prediction is necessary to accommodate sudden surges in infection.

Developing a predictive COVID-19 model

The number of individuals hospitalized with COVID-19, like other infectious diseases, follows an exponential growth and decline pattern. The group began by representing this mathematically, where today's hospitalizations are a multiple of yesterdays plus a random error margin.

A set of training data was utilized to determine the factor of multiplication and data weighted so that the most recent observations are more heavily considered. Ultimately, this allowed the group to estimate future hospitalization rates based on previous daily intake, where the rate of increase or decline is representative of an oncoming or past outbreak.

To test the validity of the COVIDNearTerm model, the group compared it with the CalCAT model previously utilized in the early stages of the pandemic. Data from six Bay Area counties between May 4^th, 2020, and June 14^th, 2020, was utilized as a training set to develop predictions two, three, and four weeks later, in addition to a long-term prediction of hospital conditions on May 1^st, 2021.

The prediction obtained from the model was then compared to the actual numbers on these dates, allowing the model's accuracy to be established. 14-day predictions had an average percentage error of 16-36% across all six counties and 34-54% at the four-week time point. The smallest difference between predicted and actual numbers was always in the most populous county, with the largest sample size, while the largest errors were always in the least populated county.

Errors within the CalCAT model were notably larger than provided by the COVIDNearTerm model. For example, in Santa Clara County, the former generated errors of 31%, 42%, and 58% at one, two, and three weeks, respectively. In comparison, the latter produced only 16%, 23%, and 34% errors at the respective time point.

The group also compared several other short-term prediction models using the same data set, finding that other methods produced more accurate results in some cases. However, out of 18 comparisons, COVIDNearTerm was the most accurate in ten cases, demonstrating that hospital admission data can be used to model COVID-19 cases in the short term at least as effective as other data sources such as widespread conventional testing to determine community transmission rates.

Conclusion

The authors point out that they had access to the most updated and recent hospitalization data during their predictions, which other models may not have had. Several variables that influence hospitalization rates have been entirely ignored in this model.

The group highlights the difficulty in implementing these additional variables in a generally applicable manner to all hospitals. Further, this model relies on the continuity of COVID-19 cases to predict the rate of exponential increase or decline in the future and thus is unsuitable for predicting the course of a sudden outbreak.

This news article was a review of a preliminary scientific report that had not undergone peer-review at the time of publication. Since its initial publication, the scientific report has now been peer reviewed and accepted for publication in a Scientific Journal. Links to the preliminary and peer-reviewed reports are available in the Sources section at the bottom of this article. View Sources