VBridge: Connecting the Dots Between Features and Data to Explain Healthcare Models

Extending the Nested Model for User-Centric XAI: A Design Study on GNN-based Drug Repurposing

“…Interactive visualizations have been widely used as a medium for explanation [14,23,82], since they excel at communication and summarization of complex information.…”

“…Some recent studies take into account the needs of domain users for the development of XAI visualization tools [14,33,41]. These studies contribute novel visualization designs and coordinated views to help domain users make sense of complicated data and generate domain-meaningful insights.…”

“…Most existing AI visu-alization tools are developed for AI developers and AI practitioners rather than domain users who have limited expertise in AI [82]. Studies that target domain users [14,72] often concentrate on proposing novel visualization designs and coordinated views to make sense of complicated data. A specific explanation is usually selected before the design study based on its popularity in the ML community without considering how the domain characteristics and user needs may influence the selection and visualization of explanations.…”

Explainable AI for clinical and remote health applications: a survey on tabular and time series data

“…Given the limited number of research works found in the literature, studies targeting tabular and time series data are grouped together in Table 8. Cheng et al (2021) developed VBridge, a novel visual analytics tool to address three key challenges related to XAI adoption in healthcare, namely clinicians' unfamiliarity with ML features, lack of contextual information, and need for cohort-level evidence. To this aim, the system first provides a hierarchical display of attribution-based feature explanations, by grouping the most relevant features semantically for a better analysis.…”

An Interactive Interpretability System for Breast Cancer Screening with Deep Learning

“…Although this framework provides an efficient human-in-the-loop paradigm to understand medical CNNs, it is not practical in real clinical settings as the framework requires radiologists to finish labeling neurons of a model before even using it, which adds a huge time overhead. Meanwhile, numerous visual analytics interfaces [5,9] have been proposed for machine learning models in healthcare applications such as electronic medical records to improve user understanding and support clinical workflow. Given the huge potential of interactive data visualizations in promoting human understanding of complex deep learning models, it's appealing to find a visual analytics solution to integrate state-of-the-art interpretability techniques into clinical workflows with deep learning components to promote Human-AI collaboration and maximize the utility of powerful medical AI models.…”