Predicting coronavirus disease 2019 severity using explainable artificial intelligence techniques

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Epidemiology

Background:

• Predicting coronavirus disease 2019 (COVID-19) severity is crucial for patient management.
• Conventional statistical methods struggle with the complex interactions of factors influencing COVID-19 severity.
• Explainable machine learning offers a promising approach for developing accurate predictive models.

Purpose of the Study:

• To establish a simple, accurate, and explainable machine learning model for predicting COVID-19 severity.
• To identify key clinical features that contribute to COVID-19 severity prediction.
• To validate the model's performance on an independent patient cohort.

Main Methods:

• Utilized a dataset of 3,301 adult patients diagnosed with COVID-19.
• Employed pointwise linear and logistic regression to extract 41 potential predictive features.
• Applied reinforcement learning to develop a parsimonious predictive model.
• Evaluated model performance using the area under the receiver operating characteristic curve (AUC).

Main Results:

• A predictive model using four features—serum albumin, lactate dehydrogenase, age, and neutrophil count—achieved an AUC of ≥0.905.
• The model demonstrated high predictive accuracy in both discovery (AUC=0.906) and validation (AUC=0.861) cohorts.
• Identified key predictors for COVID-19 severity, including serum albumin, lactate dehydrogenase, age, and neutrophil count.

Conclusions:

• Developed a simple and accurate explainable machine learning model for COVID-19 severity prediction.
• The model, utilizing four key features, shows potential for aiding clinical decision-making.
• Findings may assist in patient stratification and the selection of appropriate therapeutic interventions for COVID-19.