POSA310 Artificial Intelligence Applied on Administrative Big Data to Predict the Severity of SARS-COV-2 Infection

ABSTRACT

Objectives: To estimate the prognostic factors underlying severity of Sars-Cov-2 infection using a machine learning approach. Methods: The analysis is based on administrative databases of Italian Entities. Patients who were hospitalized with COVID-19 diagnosis (ICD-9 078.89) after 1st January 2020 were included into the dataset together with 13 relevant features representing age, sex and clinical history of each patient. Each record was labelled as 0 (hospitalized patients) or 1 (patients in intensive care or deceased). KerasTuner was used to define the architecture of the Neural Network achieving good accuracy score. To identify prognostic factors underlying severity of Sars Cov-2 infection, feature’s importance was evaluated starting from a Random Forest Classifier. Results: The preliminary dataset built contains 10.448 records from 9.346 hospitalized patients. The selected neural network is made of 13 input nodes, each one representing a feature, 1024 nodes in the hidden layer, processing information that comes from the input layer, and 2 nodes in the output layer, each one representing a label to define patient’s condition. The neural network obtained was able to achieve 64% of accuracy on the testing set. The condition of approximately 2 out of 3 patients was correctly predicted just by analysing their features. The feature’s importance computed from the Random Forest Classifier indicated that patient’s age is the primary prognostic factor underlying severity of Sars Cov-2 infection. The combination of the other features slightly improved model’s performance. Conclusions: The preliminary analysis shows that age is a prognostic factor of fundamental importance in defining the severity of Sars Cov-2 infection. The model obtained could be used to predict disease progression in patients most at risk by analysing their information in the databases. The model will be further improved through a process of feature selection to increase its accuracy and to allow the identification of other prognostic factors.