A machine learning-based predictive model for outcome of COVID-19 in kidney transplant recipients

Purpose: Health systems need tools to deal with COVID-19, especially for high-risk population,such as transplant recipients. Predictive models are necessary to improve management of patients and optimize resources. Methods: A retrospective study of hospitalized transplant patients due to COVID-19 was evaluated(March 3-April 24,2020). Admission data were integrated to develop a prediction model to evaluate a composite-event defined as Intensive Care Unit admission or intensification treatment with antiinflamatory agents. Predictions were made using a Data Envelopment Analysis(DEA)-Artificial Neural Network(ANN) hybrid, whose accuracy relative to several alternative configurations has been validated through a battery of clustering techniques. Results: Of 1006 recipients with a planned or an unscheduled visit during the observation period, thirty-eight were admitted due to COVID-19. Twenty-five patients(63.2%) exhibited poor clinical course(mortality rate:13.2%), within a mean of 12 days of admission stay. Cough as a presenting symptom(P=0.000), pneumonia(P=0.011), and levels of LDH(P=0.031) were admission factors associated with poor outcomes. The prediction hybrid model working with a set of 17 input variables displays an accuracy of 96.3%, outperforming any competing model, such as logistic regression(65.5%) and Random forest(denoted by Bagged Trees,44.8%). Moreover, the prediction model allows us to categorize the evolution of patients through the values at hospital admission. Conclusions: The prediction model based in Data Envelopment Analysis-Artificial Neural Network hybrid forecasts the progression towards severe COVID-19 disease with an accuracy of 96.3%, and may help to guide COVID-19 management by identification of key predictors that permit a sustainable distribution of resources in a patient-centered model. Improving efficiency and patient parformance in the AAN with DEA, we can get high accurancy even with no-big cohorts. (Table Presented).