Human-Centered Machine Learning in Medicine

The design of human-centered machine learning tools for medical applications raises two fundamental questions in the interaction between clinicians and machines. First: how to annotate medical data while balancing trade-offs in the type, number, and quality of annotations?  To address this question, we have to account for both human factors, such as the availability of annotators and their expertise, and the effect different types of annotations have on the learning paradigm. The second question is: how to transform algorithmic predictions into actionable and responsible decisions that improve the clinical workflow?

In this talk, I will address these questions in the context of thyroid malignancy prediction in whole slide biopsy scans. I will present a maximum likelihood estimation framework that allows us to analyze how different types of annotations contribute to the training process. I will then present an algorithm for thyroid malignancy prediction devised from our analysis and will showcase its incorporation in the clinic as a screening and ancillary testing tool. I will show results implying that the algorithm significantly reduces both the workload of pathologists and the number of unnecessary surgeries.  

I will conclude the talk by showing how similar questions about the interaction between clinicians and machines arise in other medical problems such as the analysis of medical text and Computed Tomography (CT) volumes. I will show how our framework leads to different methodologies in these cases due to differences between these types of data.