Severity prediction in COVID-19 patients using clinical markers and explainable artificial intelligence: A stacked ensemble machine learning approach

Abstract

The recent COVID-19 pandemic had wreaked havoc worldwide, causing a massive strain on already-struggling healthcare infrastructure. Vaccines have been rolled out and seem effective in preventing a bad prognosis. However, a small part of the population (elderly and people with comorbidities) continues to succumb to this deadly virus. Due to a lack of available resources, appropriate triaging and treatment planning are vital to improving outcomes for patients with COVID-19. Assessing whether a patient requires the hospital’s Intensive Care Unit (ICU) is very important since these units are not available for every patient. In this research, we automate this assessment with stacked ensemble machine learning models that predict ICU admission based on general patient laboratory data. We have built an explainable decision support model which automatically scores the COVID-19 severity for individual patients. Data from 1925 COVID-19 positive patients, sourced from three top-tier Brazilian hospitals, were used to design the model. Pearson’s correlation and mutual information were utilized for feature selection, and the top 24 features were chosen as input for the model. The final stacked model could provide decision support on whether an admitted COVID-19 patient would require the ICU or not, with an accuracy of 88%. Explainable Artificial Intelligence (EAI) was used to undertake system-level insight discovery and investigate various clinical variables’ impact on decision-making. It was found that the most critical factors were respiratory rate, temperature, blood pressure, lactate dehydrogenase, hemoglobin, and age. Healthcare facilities can use the proposed approach to categorize COVID-19 patients and prevent COVID-19 fatalities.