Prediction of COVID-19 severity

Early and effective detection of severe infection cases during a pandemic can significantly help patient prognosis and resource allocation. However, data may be slow to become available, and questions about the applicability of global datasets to local communities abound. Will global trends apply to a particular community with a particular population demographic profile, e.g., age, comorbidity prevalence, socio-economic status, etc.?

To tackle this issue, we develop machine learning that tools that (1) can accurately predict COVID-19 severity at the time of RT-PCR testing, before any blood tests, imaging, or presentation at hospitals; and (2) can be applied to small datasets reflective of a local population to best train predictions for a specific healthcare unit’s population. We frame the prediction problem as both an imbalanced classification, as severe cases are far less likely than mild cases, and as an anomaly detection problem. With a small dataset with few features, our tools can accurately predict hospitalization, ICU, and death, and can be easily re-trained for other emerging pandemics.