An explainable machine learning model for COVID-19 severity prognosis at hospital admission

Abstract

The coronavirus disease −2019 (COVID-19) pandemic has resulted in serious healthcare challenges. Due to its high transmissibility and hospitalization rates, COVID-19 has led to many deaths and imposed a considerable burden on healthcare systems worldwide. The development of prognostic approaches supporting clinical decisions for hospitalized patients can contribute to better management of the pandemic. We deploy several Artificial Intelligence (AI) techniques to derive COVID-19 severity classification prognosis models for unvaccinated patients hospitalized with mild symptoms using immunological biomarkers. The risk levels are precisely defined, targeting patients with uncertain prognostic trajectories. Forty molecular biomarkers were evaluated for their ability to predict the course of the illness. Seven biomarkers, including IL-6, IL-10, CCL2, LDH, IFNα, ferritin, and anti-SARS-CoV-2 N protein IgA antibody, emerge as the most significant early predictors for the prospective development of severe disease. After applying feature selection, we settled for two complete sets of five and three biomarkers to generate appropriate classification models. A Random Forest model with five biomarkers appears to be the most effective, with an accuracy of 0.92 for the external set. Yet, a Decision Tree model with just three biomarkers, and an accuracy of 0.84 for the external set, provides marginally lower yet robust performance and an explainable structure that broadly reflects our current understanding of disease severity. These findings suggest that the severity is influenced by a few key pathological processes. Therefore, a three-biomarker model that utilizes IL-6, IFNα, and anti-SARS-CoV-2 N protein IgA antibody levels may enhance clinical decision-making and patient triage at hospitalization, contributing to the successful management of the disease.

Ähnliche Arbeiten