Multi-objective optimization determines when, which and how to fuse deep networks: an application to predict COVID-19 outcomes

Combining multiple data types improves COVID-19 predictions: The study demonstrates that fusing chest X-ray images with clinical information produces better predictions of severe COVID-19 outcomes than using either data type alone, achieving state-of-the-art results on the AIforCOVID dataset.

Strategic fusion matters more than just adding data together: The research addresses a critical question for responsible AI—not all ways of combining different data sources are equally effective. The authors use multi-objective optimization to systematically determine the best approach, timing, and selection of which data types to combine, rather than simply merging everything together.

Explainability builds trust in medical AI predictions: The paper employs explainable AI (XAI) algorithms to reveal which medical features the model relies on and how different data types rank in importance. This transparency helps clinicians understand why the model makes specific predictions about patient outcomes, which is essential for responsible deployment in healthcare.

The method shows robustness beyond initial testing: The developed approach maintains strong performance when validated on external data, suggesting it's not just overfitted to the training dataset—an important consideration for ensuring AI systems work reliably in real-world clinical settings.

Diversity in AI models improves outcomes: By optimizing for both predictive performance and diversity among the neural networks being combined, the approach demonstrates that using varied models together produces more reliable results than relying on a single best-performing model.