The ongoing COVID-19 pandemic has strained hospitals around the world, and a startling number of American hospitals remain near or at full capacity [1]. At the start of the pandemic, every hospital lacked the data to make important predictions: which demographics or comorbidities were at highest risk of hospitalization, which hospital patients needed immediate care, and when a hospital would overflow. Since then, epidemiologists and data scientists have developed mathematical models of COVID-19 hospitalization, allowing healthcare stakeholders to make data-driven predictions. These models, which draw from statistics, machine learning, and artificial intelligence, address the problem of COVID-19 hospitalization from many perspectives.  

One perspective concerns the risk of hospital admission based on existing data for individual patients. A study of over 4,000 patients in Ohio and Florida identified a number of risk factors, including smoking, diabetes, and hypertension [2]. Similar studies in Israel [3], China [4], and the northeastern United States [5] collectively identified a wider number of risk factors, such as age, race, certain blood test results, and blood oxygen levels. Hospitalization prediction models are useful for characterizing the likely needs of COVID-19 patients. For example, given that diabetes is a risk factor for COVID-19, hospitals should anticipate an influx of patients with diabetes and procure more insulin. Similarly, knowing that elderly and infant patients are the age groups most susceptible to hospitalization, hospitals should obtain supplies catered to those patients.  

Another perspective involves triage – identifying which admitted patients need immediate care. In the COVID-19 literature, many triage models predict which admitted patients are most likely to succumb to critical illness—whether ICU admission or ventilation—or death. For example, a study of Mexican patients analyzed several predictive models for ICU admission and death, finding that simple models perform on par with complex models [6]. A parallel approach, taken by researchers at NYU, is to predict which patients are most likely to have favorable outcomes; hospitals can discharge those patients and focus on the rest [7]. In practice, hospitals should implement both models, using each model to verify the other’s predictions and reduce the possibility of errors.  

Yet another perspective examines the probability of hospital overflow. These models require a population-level view of the COVID-19 pandemic, which is often highly tailored to a specific country. Given that COVID-19 emerged just over one year ago, researchers have had few opportunities to gather the necessary country-wide data. That said, one research study was able to make near-future predictions of COVID-19 hospitalization rates, pulling data from China, South Korea, Italy, and the Netherlands [8]. Ultimately, more research is needed for making longer-term predictions, as well as for making predictions in other countries.  

AI predictions of COVID-19 hospitalization may help hospitals make rigorous plans for individual patients and the hospital as a whole. Nevertheless, every model has limitations. Only one model underwent an important test: deployment and scrutiny in a clinical setting [7]. Without rigorous clinical testing, any model should be viewed with skepticism. Every model was developed using data from specific countries [3,4,6,8] or from specific hospital systems [2,5,7]. Applying a model from New York City to the entire United States, for example, might lead to misleading results. More generally, mathematical models are only as sound as their underlying assumptions. Two models with different assumptions can compute drastically different results, even if they study the same populations and employ the same methodologies [9]. With more clinical testing, larger datasets, and a finer understanding of COVID-19, future research will yield more accurate and more applicable hospitalization models.  

References 

[1] Conlen M., et al. How Full Are Hospital I.C.U.s Near You? The New York Times. March 15, 2021. https://www.nytimes.com/interactive/2020/us/covid-hospitals-near-you.html.  

[2] Jehi L., et al. Development and Validation of a Model for Individualized Prediction of Hospitalization Risk in 4,536 Patients with COVID-19. PLOS One 2020; 15: 8. DOI:10.1371/journal.pone.0237419.  

[3] Roimi M., et al. Development and Validation of a Machine Learning Model Predicting Illness Trajectory and Hospital Utilization of COVID-19 Patients: A Nationwide Study. Journal of the American Medical Informatics Association 2021; 0: 0. DOI:10.1093/jamia/ocab005.  

[4] Xie J., et al. Development and External Validation of a Prognostic Multivariable Model on Admission for Hospitalized Patients with COVID-19. [Preprint]. The Lancet. D-20-04664.  

[5] Hajifathalian K., et al. Development and External Validation of a Prediction Risk Model for Short-term Mortality among Hospitalized U.S. COVID-19 Patients: A Proposal for the COVID-AID Risk Tool. PLOS One 2020; 15: 9. DOI:10.1371/journal.pone.0239536.   

[6] Wollenstein-Betech S., et al. Personalized Predictive Models for Symptomatic COVID-19 Patients using Basic Preconditions: Hospitalizations, Mortality, and the Need for an ICU or Ventilator. International Journal of Medical Informatics 2020; 142. DOI:10.1016/j.ijmedinf.2020.104258.  

[7] Razavian N., et al. A Validated, Real-time Prediction Model for Favorable Outcomes in Hospitalized COVID-19 Patients. NPJ Digital Medicine 2020; 3: 130. DOI:10.1038/s41746-020-00343-x.  

[8] van Wees J.-D., et al. Forecasting Hospitalization and ICU Rates of the COVID-19 Outbreak: An Efficient SEIR Model. [Preprint]. Bulletin of the World Health Organization 2020. DOI:10.2471/BLT.20.256743.  

[9] Holmdahl I., and Buckee C. Wrong but Useful — What Covid-19 Epidemiologic Models Can and Cannot Tell Us. The New England Journal of Medicine 2020; 383. DOI:10.1056/NEJMp2016822.