The Delta variant is driving a surge in hospital admissions and straining ICUs around the world. Under such conditions, accurately anticipating demand for hospital resources can mean the difference between timely treatment and, in some cases, between life and death. When and how are COVID-related admissions rates likely to change, and by how much? And how far in advance can we make such predictions accurately?
In response to this challenge, the UC Davis DataLab has partnered with UC Davis Health to develop improved machine learning and AI models for predicting COVID-19 admissions and overnight stays, and to put them into use by doctors and nurses working with COVID patients. The models have increased the UC Davis Medical Center’s ability to predict COVID-related admissions one day, two days, and even a full week in advance. And, these predictions are also more accurate, with an 8% decrease in the error rate. More accurate predictions of hospital admissions and bed occupancy have real, tangible consequences for hospital operations, such as ensuring adequate staffing levels and equipment availability, that help maintain standards of care during this health crisis. Increasing the prediction timeline also benefits hospital staff, providing workers longer windows for adjusting their work-life schedules.
Data science takes a team: Building on prior models
During the “third wave” of COVID-19 infections (December 2020-January 2021), UC Davis Health and a team at the Clinical and Translational Science Center (CTSC) began developing models for predicting COVID-19 admissions at the UC Davis Medical Center. These models used 66 distinct data features extracted from hospital databases that were identified by clinicians and statisticians as potentially relevant for determining COVID-19 case counts. These data included the total number and positivity rate of COVID-19 tests administered at the hospital, the number of COVID-19 patients receiving high-flow nasal oxygen, and even the day of the week.
This prior work provided an important foundation for DataLab’s modeling effort, which focused on iteratively improving the initial model’s predictions by adapting how the identified data features are used in the analysis. While all 66 of the previously identified data features are important and potentially useful, DataLab was able to determine that some features are more important for predicting COVID-19 hospital admissions. Based on that insight, DataLab developed a modeling workflow that allowed for deprioritization of certain variables. The new model includes only the 20 most powerful variables in terms of prediction, which helped reduce the prediction error by 8% compared to the initial all-feature model.
This improvement is especially evident when looking at surges or lulls in COVID cases. While all models struggle to make predictions during unprecedented times, the DataLab’s modeling efforts were particularly effective at increasing accuracy when COVID hospital admission rates are changing, rather than during times when admission rates are staying fairly consistent day-to-day. And these times of rapid change are exactly when accurate predictions are most helpful for planning, resource allocation and staffing.
See the DataLab COVID Admissions project description for more background on this project and technical details regarding the statistical techniques the DataLab used.
This research was conducted primarily by Dr. Wesley Brooks and Professor Vladimir Filkov at the UC Davis DataLab in partnership with the project lead, Dr. Jason Adams at the UC Davis Medical Center, with input from the wider DataLab and IT Health Informatics teams. UCDH’s IT Health Informatics is working with the team to implement the model into operation, informing medical staff of the predictions daily.
This post was written by Dr. Pamela Reynolds and Jessica Nusbaum. Questions can be directed to email@example.com.