An Explainable AI Approach for the Rapid Diagnosis of COVID-19 Using Ensemble Learning Algorithms.

Area of Science:

• Medical Informatics
• Machine Learning
• Epidemiology

Background:

• Artificial intelligence (AI) offers potential for COVID-19 screening but faces limitations due to its 'black-box' nature.
• Explainable AI (XAI) aims to address this by providing transparent insights into AI decision-making processes.

Purpose of the Study:

• To develop and evaluate an explainable AI approach for screening COVID-19 patients.
• To compare the performance of different ensemble machine learning algorithms in COVID-19 detection.

Main Methods:

• A retrospective study analyzed 1,374 participants (759 COVID-19 patients, 978 controls) using 32 blood test indices.
• Four ensemble learning algorithms were employed: Random Forest (RF), Adaptive Boosting (AdaBoost), Gradient Boosting Decision Tree (GBDT), and Extreme Gradient Boosting (XGBoost).
• Local Interpretable Model-Agnostic Explanations (LIME) were used for feature importance visualization.

Main Results:

• The GBDT model achieved the highest Area Under the Curve (AUC) of 86.4% (95% CI: 0.821-0.907), outperforming RF, AdaBoost, and XGBoost.
• Key predictors identified by the models included lactate dehydrogenase (LDH), white blood cells (WBC), and eosinophil counts (EOT).
• Cumulative feature importance was highest for LDH (0.145), WBC (0.130), and EOT (0.128).

Conclusions:

• Ensemble machine learning, particularly GBDT combined with LIME, demonstrates efficiency in screening COVID-19 patients.
• These explainable AI methods can serve as valuable tools for auxiliary COVID-19 diagnosis.
• Elevated WBC, LDH, and EOT levels are associated with an increased likelihood of COVID-19 infection.