Artificial Intelligence for Better Diagnostic Data Collection

• Applicant: Dr. Henrik von Kleist
• Subject Area: Methods in Artificial Intelligence and Machine Learning ; Epidemiology and Medical Biometry/Statistics
• Term: since 2026
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 576603640

Machine learning systems are increasingly used to support clinical diagnostics, with sepsis prediction becoming a flagship example. Early detection of sepsis is critical, and many hospitals now deploy real-time risk scores that continuously update as new patient data—such as lab results and vital signs—arrive. However, these models implicitly rely on key measurements being available, without addressing which tests should be performed and when. In clinical practice, deciding which tests to order is itself a challenging task, constrained by time, cost, and uncertainty. This project focuses on Active Feature Acquisition (AFA): AI systems that recommend what to measure, aiming to support timely and cost-efficient decision-making. During my PhD, I developed the theoretical foundations to evaluate AFA systems. I now aim to translate these insights into applied, clinically useful tools—starting with a concrete use case in sepsis prediction (Objective 1). Afterwards, I will extend the framework to the broader LLM-based diagnostic dialogue setting (Objective 2), treating the language model itself as the AFA agent that interactively acquires the most informative symptoms, tests, or images before making a diagnosis. The project will be conducted at the CAUSALab at Harvard University, in close collaboration with Miguel Hernán and James Robins, leading experts in causal inference and sequential decision-making.

• DFG Programme: Fellowship
• International Connection: USA
• Host: Professor Dr. Miguel Hernan