Ben Wilson

Trust and AI in Clinical Decision Support

Abstract

Abstract Machine learning (ML) and artificial intelligence (AI) techniques are increasingly visible in the domain of clinical decision support. Significant results have been obtained in the interpretation of diagnostic imaging using computer vision techniques - notably in the specialties of ophthalmology, dermatology, cellular pathology, cardiology, oncology and respiratory medicine. Risk prediction and prognosis tools are also beginning to emerge. However, there is limited research on how well and how far these techniques can be integrated into real clinical workflows. This project set out to study one of the key success factors in translating research into effective practice - that of the trust clinical decision-makers place in the intelligent systems they are able to access.

The project made use of initial clinical user engagement through workshops, interaction studies and interviews to develop a broad online user study that measured how a user’s trust varies with different system characteristics. The intensive workshops and interaction work were used to inform design elements for the main study. In the main online study, clinical decision-makers were asked to evaluate seven different hypothetical systems in three different clinical contexts.

By holding the clinical decision context constant for a set of system characteristics in the online study, and by randomising the other presentational variables, the experimental methodology provides assurance that the results are indeed able to indicate a real signal if it exists. In the course of preparing the online study, we demonstrated a successful approach to clinical engagement that constitutes a framework for further development of participatory design work. We provide a scalable web-based assessment tool with a proven ability to capture sustained clinical input.

We co-created a clinically-informed subset of AI system characteristics that can be used to explore key components of the complex trust relationship users will experience when engaging with intelligent systems in clinical decision contexts. We show some initial results that show promise for future work where more data can be collected. This work suggests that there is value and opportunity in further exploring the characteristics in AI systems that engender trust in clinical decision-makers.