Trust at Hand: A Multi-touch Tabletop Interface for Collaborative Training of Machine Learning Models in the Medical Domain

Abstract

Despite strides in medical AI research, adoption of medical AI models still lags behind, with low trust among medical practitioners. This can also be observed in medical AI research teams, where low levels of collaboration between team members during the process of model creation resulting in low trust in the model. In this thesis I present a prototype large form factor multi-touch interface for labeling and training models that addresses undertrust in medical AI workflows by: (i) Increasing normative trust through frequent interaction between domain experts and the AI model. (ii) Increasing affective trust between domain experts and data scientists through encouraging frequent collaborative interactions. (iii) Collects rich spatial data during labeling through multi-touch tabletop interface, which can be later leveraged by data scientists during model training. User study data shows that the system increased normative and affective trust when compared to traditional AI workflows.