Machine learning tools for medical data analysis are approaching human-level performance in controlled settings in many applications. However, major hurdles still obstruct the wide adoption of AI in clinical practice. In real-world settings, in particular deep learning algorithms are famously brittle and are known to often fail silently and catastrophically, and AI systems have the potential to cause harm to the patient. For example, a failure to detect hypoglycemia in an AI-assisted artificial pancreas system has serious consequences.

Diabetes care will increasingly rely on diverse complex data sources to address unmet patient needs, and Artificial Intelligence (AI) will play a crucial role in interpreting their complex and interdependent relationships. Our long-term goal is to develop certifiably safe, reliable and effective data science tools to improve patient-specific treatment systems. Toward this goal, we are establishing the necessary data processing and analysis frameworks.

Methodological interests in trustworthy AI:

• Interpretable machine learning
• Performance generalisation and performance prediction
• Robustness
• Disentangled representation learning

Application areas:

• Interpretation of Continuous Glucose Monitoring (CGM) data, including for example CGM forecasting and risk factor prediction

• Medical image analysis including for example image classification, segmentation, disease progression modelling