Naïve Bayes for Health-Status Predictive Monitoring in COVID-19: Leveraging Drugs and Diagnoses

The COVID-19 outbreak, originated in China in late December 2019, resulted in around 15 million deaths by 2021 due to its rapid transmission and challenging early diagnosis. Since then, scientists have leveraged machine learning potential to enhance decision support systems for managing the disease. This paper proposes a predictive model based on the Naïve Bayes Classifier for monitoring the health-status of COVID-19 patients admitted to Spanish hospitals during the pandemic’s first two waves. Three events have been considered: favourable evolution, intensive care unit admission, and exitus. The model predicts reasonably good the probability of the patient’s event, even as early as seven days prior to the event day. Various input feature combinations among demographic, diagnostic and pharmacological variables, were explored. Best predictive capability is obtained when all features are jointly used (accuracy rate values close to 0.8 for every event-type). Predictive capability empowers clinicians to anticipate patient’s health-status progression and support them to make informed treatment decisions, potentially preventing fatal outcomes.