Comparison of bayesian networks of different structural complexity for predicting COVID-19 mortality

Abstract

This paper considers a methodology for developing, tuning, and comparing three probabilistic models – Bayesian networks of  different classes with increasing structural complexity – for the task of predicting the mortality of patients with coronavirus disease.  The compared models are the naive Bayes classifier, the tree-augmented naive Bayes classifier, and a structurally constrained  Bayesian network whose structure is recovered from the data by a structure-learning algorithm under expert clinical constraints. To  separate the effect of structural complexity from feature selection, parameter tuning, and threshold choice, a comparison under  matched conditions with successive simplification and enrichment of the model structure is carried out. Particular attention is paid to  distinguishing two aspects of model quality – the discriminative ability and the quality (reliability) of the predicted probabilities – because for a clinical decision-support system it is the trustworthiness of the risk estimate, rather than the mere ranking of patients by  their level of risk, that is decisive. The study is conducted on an open registry of patients with corona virus disease using clinical  features, with evaluation on a sample that reproduces the natural prevalence of fatal outcomes. The paper also considers the transfer  of the predicted probabilities to the natural population through a prior-probability correction, as well as the sensitivity of clinical  decisions to the ratio of error costs and an analysis of the net clinical benefit. The aim of the work is to substantiate the suitability of a  structurally constrained Bayesian network as an interpretable risk-assessment tool in clinical decision-support systems.