Hospital Projection

UCLA Health has released to the community the system it used for projecting the expected hospital resources needed for the care of COVID-19 patients, a calculation that includes inpatient beds, ICU beds and ventilators.

The system focuses on projections of up to a few weeks, since such short-term predictions are highly accurate, while still useful to provide guidelines for policymakers. The system has been in place since mid-April, helping UCLA Health model the resource needs during the first COVID-19 patient surge in Los Angeles.

The system is now available for use by other health systems through a web interface hosted at UCLA.

►Press Release available

Full details about the approach are in the following preprint

"Projecting hospital resource utilization during a surge using parametric bootstrapping" August 02, 2020
Jeffrey N Chiang, Ulzee An, Misagh Kordi, Brandon Jew, Clifford Kravit, William J Dunne, Ronald Perez, Neil R Parikh, Drew Weil, Richard F Azar, Robert Cherry, Karen A Grimley, Samuel A Skootsky, Christopher E Saigal, Vladimir Manuel, Eleazar Eskin, Eran Halperin

How to use STOP COVID Hospital Projections

Input to our model are hospital census counts broken down by resource for the current day. By default, the tool assumes that the number of new cases doubles every 4 days for the two week projection period as a “worst-case” scenario. However, previous case counts can be provided in order to estimate the actual doubling rate for your institution.

The tool provides the predicted resource needs of COVID patients for the next two weeks. An example output is shown below. The predictions include three curves. A blue curve showing hospital bed usage for COVID patients, an orange curve showing ICU bed usage for COVID patients and a green curve for ventilator usage for COVID patients. The curves also contain 95% confidence intervals.

The projections for hospital resource usage have two components. The first is the amount of resources that currently hospitalized COVID patients will utilize before they are discharged. The second is the number of patients with COVID who will arrive at the hospital and the future resources that they will utilize. The largest amount of uncertainty in hospital resource predictions is in predicting the number of new patients that will arrive in the future. When considering historical data in UCLA Health, we observed that the majority of the COVID-19 patients had been admitted for at least seven days. For this reason, the near-term projections (e.g., two weeks) are more accurate compared to long term projections.


Figure 2. Hospital usage broken down by length of stay. Most of the COVID patients in the hospital on any given day have already been admitted for several days.

Figure 3. Left. Performance on a simulated population with higher average age. Right. Performance on a simulated population with lower average age. The dashed lines indicate the projected number of admissions and the solid lines indicate actual number of admissions. Orange: Projections without accounting for the difference in mean age. Blue: Accounting for the difference in age by providing the algorithm with the mean age of the patients in the health system.

By default, our model assumes that the average age of incoming patients is around 45 years. However, this may not be the case at your institution. Our model allows the option of providing the average age at your institution in order to generate more accurate predictions. As shown in figure 3, our model is more accurate when the correct average age is provided.

Model Description

The model assumes that some number of patients will test positive for SARS-Cov-2 every day, and this number is governed by a doubling rate. If historical data are provided, the model estimates the two-week doubling time. Our approach to resource usage projection is based on parametric bootstrap. Under this approach, we define a probabilistic generative model which emulates the resource demand of COVID-19 patients. To generate resource usage predictions, we sample the resource usage of the expected individuals to generate a trajectory of resource demand. We repeat this process 1000 times, then take the median required capacity across all trajectories at each day as our prediction.

STOPCOVID19 Hospital Projections in the News:
UCLA Press Release

Hospital Projections Team:

  • Eleazar Eskin, PhD, Professor and Chair of Computational Medicine, Professor of Computer Science and Human Genetics, UCLA
  • Eran Halperin, PhD, Professor of Computational Medicine, Computer Science, Human Genetics, and Anesthesiology, Associate Director of Informatics, Institute for Precision Health, and the Head of AI in Medicine, UCLA
  • Brandon Jew, Bioinformatics Interdepartmental Program, UCLA
  • Jeffrey Chiang, PhD, Department of Computational Medicine, UCLA
  • Misagh Kordi, PhD, Department of Computer Science, UCLA
  • Ulzee An, Department of Computer Science, UCLA