Profiling short-term longitudinal severity progression and associated genes in COVID-19 patients using EHR and single-cell analysis

Here we propose CovSF, a deep learning model designed to track and forecast short-term severity progression of COVID-19 patients using longitudinal clinical records. The motivation stems from the need for timely medical resource allocation, improved treatment decisions during pandemics, and the understanding of severity progression related immunology. The COVID-19 Severity Forecasting model, CovSF, utilizes 15 clinical features to profile the severity levels of hospital admitted patients and also forecast their severity levels of up to three days ahead. CovSF was trained on a large COVID-19 cohort (n=4,509), achieving an AUROC of 0.92 with 0.85 and 0.89 sensitivity and specificity on an external validation dataset (n=443). The type of oxygen therapy administered was utilized as the target predictive label, which is often used as the severity index. This approach enables the inclusion of a more comprehensive dataset encompassing patients across the full spectrum of severity, rather than restricting the analysis to more narrowly defined outcomes such as ICU admission or mortality. We focused on profiling deteriorating and recovering health conditions, which were validated using patient matched single-cell transcriptomes. Especially, we showed that the immunology significantly differed between the samples during deterioration and recovery, whose severity levels were the same, and thus presenting the importance of longitudinal analysis. We believe that the framework of CovSF can be extended to other respiratory infectious diseases to alleviate the strain of allocating hospital resources, especially in pandemics.