Tree-based machine learning model integrating clinical table data and chest x-ray images to predict prognosis of covid-19 patients

Abstract

Background: Since the peak number of COVID-19 mortality in Japan tends to increase with each outbreak, it is important to identify patients at high risk of death at an early stage. In a previous study, a neural network model was simultaneously trained with clinical and imaging information to predict the prognosis of COVID-19 patients. On the other hand, it has been reported that tree-based models outperform neural networks for table data. We aimed to construct and evaluate a tree-based prognostic model that treats chest X-ray images in the same way as other clinical information by reducing them to one dimension using neural networks. Methods: This study used the Stony Brook University COVID-19 dataset of 1313 inpatients. In order to predict mortality of all patients, we developed a model (hereafter referred to as “proposed model”) with LightGBM trained with the clinical table data combined with the result of prediction of a model with DenseNet trained with chest radio graphs. We also developed a model with LightGBM trained with the clinical data alone and a model with DenseNet trained with images. Results: The f1 scores for the proposed model, the table data-only model, and the image-only model were 0.59, 0.58, and 0.25, respectively. The ROC-AUC for the proposed model, the table data-only model, and the image-only model were 0.91, 0.93, and 0.63, respectively. Image prediction had the highest variable importance in the proposed model. Conclusion: the proposed model achieved high prediction accuracy. It was suggested that clinical table data contributed more to than chest X-ray images. On the other hand, images may improve prediction accuracy when combined with clinical table data.