Development of a prognostic model for mortality in COVID-19 infection using machine learning

Abstract

Coronavirus disease 2019 (COVID-19) is a novel disease resulting from infection with severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), which has quickly risen since the beginning of 2020 to become a global pandemic. As a result of the rapid growth of COVID-19, hospitals are tasked with managing an increasing volume of these cases with neither a known effective therapy, an existing vaccine, nor well-established guidelines for clinical management. The need for actionable knowledge amidst the COVID-19 pandemic is dire and yet, given the urgency of this illness and the speed with which the healthcare workforce must devise useful policies for its management, there is insufficient time to await the conclusions of detailed, controlled, prospective clinical research. Thus, we present a retrospective study evaluating laboratory data and mortality from patients with positive RT-PCR assay results for SARS-CoV-2. The objective of this study is to identify prognostic serum biomarkers in patients at greatest risk of mortality. To this end, we develop a machine learning model using five serum chemistry laboratory parameters (c-reactive protein, blood urea nitrogen, serum calcium, serum albumin, and lactic acid) from 398 patients (43 expired and 355 non-expired) for the prediction of death up to 48 h prior to patient expiration. The resulting support vector machine model achieved 91% sensitivity and 91% specificity (AUC 0.93) for predicting patient expiration status on held-out testing data. Finally, we examine the impact of each feature and feature combination in light of different model predictions, highlighting important patterns of laboratory values that impact outcomes in SARS-CoV-2 infection.

Fig. 1

Learned regression coefficients for each of 26 laboratory values provided to a logistic regression model. Features are each present on the x-axis with their corresponding regression coefficient on the y-axis. Features are ranked according to the absolute value of their corresponding coefficient. The last-ranked features protein, total and aspartate aminotransferase have regression coefficients of zero.

Fig. 2

Receiver operating characteristic (ROC) curve depicting the performance of a trained support vector machine classifier using the top 5 highest-weighted laboratory values. Normalized confusion matrix depicting the support vector machine's prediction of patient expiration versus a patient's true expired status.

Fig. 3

Feature importance for each of five core laboratory parameters for our trained SVM model. As the influence of laboratory parameters will be different in the setting of different patients, this additive force diagram depicts the changing influence of these parameters throughout the test set. Large blue and pink background areas represent true-positive and true-negative predictions with false-positive and false-negative predictions overlaid.

Fig. 4

Relationships between feature and Shapley values. The relationship between the value (color) and Shapley value (x-axis) is plotted for each of five laboratory parameters used to train the SVM model.