Multimodal graph attention network for COVID-19 outcome prediction

Area of Science:

• Medical Imaging and Artificial Intelligence
• Computational Biology and Bioinformatics
• Infectious Disease Modeling

Background:

• Predicting individual COVID-19 disease progression is challenging due to unknown patient- and disease-specific factors.
• Early prognosis of COVID-19 patient outcomes, such as intensive care unit (ICU) admission, is crucial for resource allocation and treatment planning.
• Current methods often rely on acute indicators, limiting the ability for timely and accurate predictions.

Purpose of the Study:

• To develop a holistic, multimodal graph-based approach for predicting COVID-19 patient outcomes.
• To integrate diverse data modalities, including imaging (CT scans) and non-imaging (clinical data), for a comprehensive patient representation.
• To enhance early prognosis of critical events like ICU admission, ventilation, and mortality.

Main Methods:

• A multimodal similarity metric was used to construct a population graph, enabling patient clustering.
• Radiomic features were extracted from chest CT scans using a segmentation neural network, serving as a latent image feature encoder.
• A Graph Attention Network (GAN) integrated radiomic and clinical data (vital signs, demographics, lab results) for end-to-end outcome prediction.

Main Results:

• The multimodal graph-based approach demonstrated superior performance compared to single-modality and non-graph baselines.
• Patient clustering within the graph provided insights into patient relationships and disease progression patterns.
• The Graph Attention Network effectively predicted key COVID-19 patient outcomes, including ICU admission, ventilation requirement, and mortality.

Conclusions:

• The proposed multimodal graph-based approach offers a robust framework for predicting COVID-19 patient outcomes.
• Integrating imaging and clinical data via graph networks enhances predictive accuracy and provides deeper understanding of patient cohorts.
• This methodology holds potential for improving clinical decision-making and resource management in emerging infectious diseases.