Forecasting Hospital Admissions Related to COVID-19

COVID-19 Research Area(s): Mathematical Modelling and Operations

We develop a model that forecasts the number of hospital admissions (and subsequently to the ICU) related to COVID-19 on a daily basis.  We use a very rich data set obtained from hospitals in N. America combined with population data described below. The novelty of our approach is that we don’t use an explicit epidemiological model that uses infection spread rate. This is because, due to the lack of tests and the resulting highly restricted testing policies, reported number of infections is not close to being an accurate estimate of the actual infected in a population. Our approach is to combine a growth model with a machine learning model that consider many of the following attributes --- past daily hospital admissions (data from US CMS system + select BC hospitals), number of ED presentations related to COVID like symptoms (same as above), number of daily self diagnostic assessments that returns positive on the CDC website, social media and google searches for symptoms, weather, population density, demographic data (age, sex, comorbidities etc.), access to health care and other relevant variables related to Social distancing. Finally, the main measure that is of interest to health care system at this point (and one that will continue to be for the next many months) is the number of patients seeking intensive/acute resources at the hospital, which is precisely our objective.