COVID-19 diagnosis prediction in emergency care patients: a machine learning approach

Andre Filipe de Moraes Batista; Joao Luiz Miraglia; Thiago Henrique Rizzi Donato; Alexandre Dias Porto Chiavegatto Filho

Calibration details for the support vector machines algorithm are presented in Figure 2 , which shows that despite the small sample size, the algorithm is well calibrated throughout the probability distribution, i.e. among patients with a high predicted risk according to the algorithm, a high percentage of them are COVID-19 positive, and vice-versa. The five most important variables for the predictive performance of the algorithm according to the Mean Decrease Accuracy (MDA) were lymphocytes, leukocytes, eosinophils, basophils and hemoglobin, respectively. We also tested the performance of the algorithms using 10-fold cross validation repeated 10 times for each algorithm, instead of using the 70-30 split. This strategy is frequently used in machine learning in health studies especially in cases of relatively small samples such as ours, but it simulates less clearly the results in clinical applications. The predictive performance of the algorithms was better in comparison with the previous strategy, with an AUC of 0.87, Sensitivity of 0.755 and Specificity of 0.829 for the best-performing algorithm, which was again the support vector machines (Annex 1).