Early and fair COVID-19 outcome risk assessment using robust feature selection

Area of Science:

• Artificial Intelligence in Medicine
• Computational Epidemiology
• Health Informatics

Background:

• Personalized medicine is crucial for optimizing COVID-19 patient care, especially for high-risk individuals needing early intervention.
• Accurate prediction of COVID-19 clinical outcomes is essential for timely treatment and preventing severe complications.
• Existing machine learning models often lack equitable performance across diverse demographic groups, highlighting a need for inclusive solutions.

Purpose of the Study:

• To develop a robust machine learning model for predicting patient-specific risk of death or mechanical ventilation in COVID-19 positive individuals.
• To ensure the model demonstrates equitable performance across various demographic groups, including gender and race.
• To enhance clinical trust through interpretable predictions, including patient clustering and feature importance analysis.

Main Methods:

• Generation of a machine learning model using features available at diagnosis to predict severe COVID-19 outcomes.
• Evaluation of model performance across different patient demographics to ensure fairness and equity.
• Implementation of interpretable AI techniques, including patient clustering and feature importance (patient-level and global).
• Utilized deep learning for patient clustering, comparing its performance against traditional methods using mutual information.

Main Results:

• Achieved an 89.38% area under the receiver operating characteristic curve (AUROC) for predicting severe outcomes.
• Identified dementia as a significant predictor of worse patient outcomes through robust feature ranking.
• Demonstrated superior performance of deep-learning-based clustering in differentiating patient severity compared to traditional clustering methods.
• Developed an application for automated, fair patient risk assessment with minimal data entry.

Conclusions:

• The developed machine learning model offers a data-driven approach to optimize COVID-19 patient management through accurate risk prediction.
• The model's equitable performance across demographics addresses a critical gap in current AI healthcare solutions.
• Interpretability features enhance clinical trust and understanding of the model's predictions, facilitating adoption in real-world settings.
• The study provides a foundation for fairer, more effective AI-driven healthcare decision-making in infectious disease management.