Keywords: Explainability, Decision Support, Tabular
Need: AI systems can achieve performance on narrow tasks comparable to human experts, but still make unexpected errors—particularly in cases where humans perform well. Many such models lack clear, human-understandable measures of uncertainty, and cannot reliably detect out-of-domain inputs that should be deferred to alternative processes.
In NHS settings, this raises challenges for safe deployment and trust. This project will explore techniques for human-AI collaboration, where models and clinicians work together to improve performance, robustness, and efficiency. The focus will be on “arbitrator” settings—where a model, human, or combined process decides how to act—examining shared understanding of uncertainty, methods for transparent decision-making, and ways to fully audit outputs.
Current Knowledge/Examples & Possible Techniques/Approaches:
Recent research has proposed frameworks for human-AI collaboration in decision-making. For example, Dvijotham et al. (2023) explore complementarity-driven deferral systems, where AI defers to clinicians in uncertain cases (Nature Medicine). Open-source projects such as DeepMind’s CoDoC demonstrate calibration and deferral strategies, while evidential deep learning offers principled uncertainty estimation. More complex scenarios, such as active acquisition for multimodal temporal data, are discussed in Kossen et al. (2023), highlighting the challenges of dynamic, data-driven decision-making.
Related Previous Internship Projects: n/a first year of this topic
Enables Future Work: Supports wide deployment of AI applications in healthcare workflows
Outcome/Learning Objectives:
Datasets: Open healthcare datasets that support the exploration of human-algorithm interactions
Desired skill set: When applying please highlight any experience around work with human-computer interaction, agent settings, AI ethics, model explainability, python coding experience and software development (including any coding in the open), and any other data science experience you feel relevant.
Return to list of all available projects.