Prediction of COVID-19 severity from clinical and biochemical markers: a single-center study from Saudi Arabia

It is known that the severity of COVID-19 is linked to the prognosis of patients; therefore, an early identification is required for patients who are likely to develop severe or critical COVID-19 disease. The purpose of this study is to propose a statistical method for identifying the severity of COVID-19 disease by using clinical and biochemical laboratory markers.

A total of 48 clinically and laboratory-confirmed cases of COVID-19 were obtained from King Fahad Hospital, Medina (KFHM) between 27th April 2020 to 25th May 2020. The patients' demographics and severity of COVID-19 disease were assessed using 39 clinical and biochemical features. After excluding the demographics, 35 predicting features were included in the analysis (diabetes, chronic disease, viral and bacterial co-infections, PCR cycle number, ICU admission, clot formation, cardiac enzymes elevation, hematology profile, sugar levels in the blood, as well as liver and kidney tests, etc.). Logistic regression, stepwise logistic regression, L-2 logistic regression, L-2 stepwise logistic regression, and L-2 best subset logistic regression were applied to model the features. The consistency index was used with kernel Support-Vector Machines (SVM) for the identification of associated markers.

L-2 best subset logistic regression technique outperformed all other fitted models for modeling COVID-19 disease severity by achieving an accuracy of 88% over the test data. Consistency index over L-2 best subset logistic regression identified 14 associated markers that can best predict the COVID-19 severity among COVID-19 patients.

By combining a variety of laboratory markers with L-2 best subset logistic regression, the current study has proposed a highly accurate and clinically interpretable model of predicting COVID-19 severity.