Predicting the Risk of Clinical Deterioration in Patients with Severe COVID-19 Infection Using Machine Learning

Abstract

Background: Only part of the severe patients with coronavirus disease 2019 (COVID-19) may deteriorate to become critically ill and require aggressive treatments and intensive care. We aim to develop an early-warning system to predict the risk of clinical deterioration in an earlier stage to distinguish the patients with or without potential risks in advance by a systematic, data-driven, and analytical approach in severe patients with COVID-19. Methods: This study was designed as a single-centre, retrospective, observational study for the development and validation of a clinical prediction model. We retrospectively analysed 1,364 COVID-19 patients under severe conditions from 22nd Jan. 2020 to 15th Mar. 2020 in Wuhan Tongji hospital (Wuhan, China). We developed a computational modelling framework including data pre-processing, feature generation and ranking, and LightGBM (Light Gradient Boosting), an ensemble machine learning approach, to construct an early-warning system. We thoroughly examined the top features given by the explainable tool from A total of 226 features and attempted to obtain near-optimal performance with much fewer features. To train the model, all samples are divided into a training set and a test set according to a 4:1 ratio. Findings: Nine physiological indicators from the ranking list were used, namely, the first time SARS-CoV-2 nucleic acid detection result, alanine transaminase, lymphocytes, lactate dehydrogenase, antibodies of IgG, C reactive protein, total protein, myoglobin, and monocytes. The prediction model has been validated with AUC 0.9754, ACC 0.9524, recall 0.8571 and specificity 0.9732 using the nine available clinical markers. Interpretation: The prediction model will be used for identifying patients who are at risk of critical disease in early stage of the disease. Due to the high specificity, patients with low risk of critical disease according to the prediction can be classified and managed appropriately to reduce the investment of medical resources.Funding Statement: The study was conceived, developed, and executed by the authors with no funding and external or commercial support.Declaration of Interests: The authors declare that they have no conflict of interest.Ethics Approval Statement: The Ethics Commission of Wuhan Tongji hospital approved the study.