Understanding how expert knowledge may improve artificial intelligence algorithms

This project brings together a new public-private partnership to advance the use of Artificial Intelligence (AI) in healthcare. Traditional AI models, which rely on pre-defined data from specific clinical contexts, often struggle in real-world scenarios where the context isn't clear. Our research aims to explore how healthcare providers handle these open-world situations and develop AI that can do the same.

In healthcare, identifying the right clinical context is crucial for accurate diagnosis and treatment. Traditional AI models, trained on specific datasets, often fail when faced with unfamiliar scenarios. However, healthcare providers use small "bricks of knowledge" to navigate these situations effectively. By investigating this approach, we aim to improve AI's ability to adapt and function in varied clinical contexts.

We will test our hypothesis by comparing three types of AI training datasets: basic data, data enhanced with modular knowledge structures, and a combination of both. This will be done across three clinical use cases in an Intensive Care Unit (ICU) setting, known for its complex and urgent data needs. By comparing the predictive accuracy and interpretability of these models, we seek to understand how to technically define knowledge for AI in healthcare.

Our project's ultimate goal is to create AI that can better mimic the flexible decision-making of healthcare professionals, making medical knowledge more accessible and useful for both practitioners and patients. By enhancing the adaptability and reliability of AI, we aim to significantly impact healthcare delivery and improve patient outcomes. This research is a vital step toward democratizing medical knowledge, ensuring that advanced healthcare is available to all.