Prediction Of Criticality In COVID-19:A Machine-learning-based Model Study

Background: The novel coronavirus pneumonia is a serious threat to human life safety and the global health system.However,the current risk assessment of critical illness of new crown pneumonia is not precise enough.An object of this paper is to investigate to develop a machine learning(ML)model to distinguish mild from critical illness in Corona Virus Disease 2019,which could serve as a future predictor of the occurrence of critically ill patients and facilitate the rational allocation of health care resources.Methods: The medical records of patients with confirmed COVID-19 diagnosis were retrospectively analyzed,including gender,age,comorbidities and laboratory indices.All patients were randomly divided into training and testing cohorts according to 7:3.We used six ML models to select features in the training cohort,and used subject operating characteristic(ROC)curves to evaluate the performance of the models.The test dataset was used to validate the accuracy of the model.In addition,766 patients from another hospital were used as an external validation cohort to further validate the generalization of the optimal ML model.Results: The LR model constructed based on clinical features,comorbidities and laboratory indicators performed best with AUC values of 0.864(95% CI 0.825,0.904) and 0.794(95% CI 0.725,0.875)in the training and validation sets,respectively.In addition,the LR model correctly distinguished between mild and critical cases of Covid-19 with an AUC of 0.785(95% CI 0.761,0.848)and high sensitivity(0.750)in the external independent validation data.The calibration was good for accurate matching to the diagnosis,both in the training,test and external validation cohorts.Decision curve analysis demonstrated the clinical usefulness of column line plots.Conclusions: The current study may contribute to predict the criticality in COVID-19.Our ML model summarizes the clinical features and constructs an early warning system for criticalization of Covid-19.Besides,ML method should facilitate clinical monitoring and treatment ofcriticalized patients.The column line diagram proposed in this study can be conveniently used to facilitate individualized prediction of critical illness in Covid-19.