Theses in Process

Theses in Process

Interactive Interfaces for medical Treatment Effect Prediction Explanations

Type:
Bachelor Thesis Business Information Systems
    Status:
    in process
    Tutor:

    Abstract

    The application of deep learning can lead to increased prediction performance in a wide variety of use cases. However, in high stakes decision contexts like for example medicine, an increase in performance on its own can be insufficient to foster trust towards a predictive model. If lack of trust has the effect of the high performance decision support system around the deep learning model not being used, nothing is gained. Developing an interface with the intention of convincing users of its plausibility on the other hand risks overtrust and an abandonment of critical thinking, especially among less experienced professionals. XAI methods are a way to support appropriate levels of trust by giving users tools to detect faulty reasoning by an otherwise inscrutable deep learning model. We have developed visualizations as explanations of treatment effect predictions. During our evaluation of these visualizations, several experts voiced their interest in interactive components to explore the model's reasoning and underlying data to develop a better understanding. 

    To build on our visualizations a structured literature review on interactivity in decision support systems is conducted in this Bachelor's thesis. On the basis of this SLR the student develops an interactive explanation interface for a medical treatment effect prediction case and evaluates it in a series of expert interviews.