An Explainable AI Model for ICU admission prediction of COVID­19 patients

Abstract

Machine Learning is a field that is widely used in all aspects of our lives. However, despite of the benefits of its use, its function and the process that every model follows to produce a result can not be easily comprehended by a human. Especially in medical applications and self driving cars, it is very important for the user to understand the steps the model takes to reach the solution and the importance of the features of the model, in order to avoid mistakes and improve the model’s functionality. In this thesis, we created an explainable AI model that can predict the ICU admission of COVID­19 patients, based on their symptoms and lab results. An important feature of our model is the interpretation and justification of the prediction to the user. This would allow for example a doctor that uses the program to know in advance which patients are more likely to be admitted to the ICU and monitor them and also assess the validity of such prediction. For the creation of the program, we first trained models with a variety of different algorithms in Python, using COVID­19 patients’ data from the Sirio Libanes hospital in Sao Paolo, Brazil. Then, we took the model with the highest accuracy, which in this case was used an adaboost algorithm with a random forest weak learner and transferred it in the R language, where we used the InTrees library to create a sum of the most important rules, which we can use to get a good ICU admission prediction. Finally, with the above rules, we created a Prolog program, using the Gorgias framework, which takes as an input some important patient lab results and returns a prediction on whether the patient will be admitted and the key symptoms based on which the program produced that result. The framework for the user’s communication with Gorgias was written in Java.