Automated Risk Assessment of COVID-19 Patients at Diagnosis Using Electronic Healthcare Records

摘要

COVID-19 causes significant morbidity and mortality and early intervention is key to minimizing deadly complications. Available treatments, such as monoclonal antibody therapy, may limit complications, but only when given soon after symptom onset. Unfortunately, these treatments are often expensive, in limited supply, require administration within a hospital setting, and should be given before the onset of severe symptoms. These challenges have created the need for early triage of patients likely to develop life-threatening complications. To meet this need, we developed an automated patient risk assessment model using a real-world hospital system dataset with over 17,000 COVID-positive patients. Specifically, for each COVID-positive patient, we generate a separate risk score for each of four clinical outcomes including death within 30 days, mechanical ventilator use, ICU admission, and any catastrophic event (a superset of dangerous outcomes). We hypothesized that a deep learning binary classification approach can generate these four risk scores from electronic healthcare records data at the time of diagnosis. Our approach achieves significant performance on the four tasks with an area under receiver operating curve (AUROC) for any catastrophic outcome, death within 30 days, ventilator use, and ICU admission of 86.7%, 88.2%, 86.2%, and 87.8%, respectively. In addition, we visualize the sensitivity and specificity of these risk scores to allow clinicians to customize their usage within different clinical outcomes. We believe this work fulfills a clear clinical need for early detection of objective clinical outcomes and can be used for early screening for treatment intervention.