The prognostic tool achieved an area under the receiver operating characteristic curve of 0.846 across 5-fold cross-validation, allowing clinical teams to manage risks proactively before test results arrive.
Predicting COVID-19 Severity From Historical EHR Data
The novel coronavirus pandemic placed an unprecedented burden on healthcare systems globally, pushing scientists to create reliable risk prediction models. Traditional prognostic tools typically rely on post-admission data, including vital signs, lab test results such as lymphocyte counts, C-reactive protein, and creatinine levels, alongside radiologic imaging features. By contrast, the Columbia University Irving Medical Center team built a recurrent neural network that processes a patient’s historical electronic health records to calculate a risk score representing the probability of progressing to severe status—defined as mechanical ventilation, tracheostomy, or death—following infection.
Methodology and Patient Cohort at NewYork-Presbyterian
This institutional repository contains 30 years of comprehensive electronic health record data spanning approximately 6.5 million patients, reflecting a large cohort of individuals treated for COVID-19 in New York City through May 31, 2020. Recurrent neural networks excel at modeling sequential phenomena like speech and language by capturing hidden relationships within sequential clinical data. Similar neural network architectures previously mapped future medical events or heart failure risks using chronological patient records, establishing a technical precedent for this coronavirus application.
Clinical Implications for Hospital Resource Allocation
Proactive risk management at the time of hospital admission provides critical operational advantages for hospital administrators and health policy makers. Because the recurrent neural network model requires no post-diagnosis data such as vital signs or laboratory panels, care teams can identify vulnerable patients immediately upon arrival.
Keep reading