VIVA: Virtual Healthcare Interactions Using Visual Analytics, With Controllability Through Configuration

TL;DR

The paper tackles enabling rapid, data-driven analysis of HLBC's virtual healthcare usage to optimize triage and preserve health-system capacity during a pandemic. It introduces VIVA, a visual analytics tool built on data, task, and workflow abstractions and the Scan, Act, Adapt framework, augmented by the Controllability Through Configuration model to bridge programming and end-user interactivity. The authors validate the approach through three stakeholder-driven case studies and trace an architectural evolution from baseline to fully interactive configurations. They argue that the Controllability Through Configuration model and the constrained, attribute-focused analysis workflow offer a generalizable blueprint for rapid visualization design in regulated domains.

Abstract

At the beginning of the COVID-19 pandemic, HealthLink BC (HLBC) rapidly integrated physicians into the triage process of their virtual healthcare service to improve patient outcomes and satisfaction with this service and preserve health care system capacity. We present the design and implementation of a visual analytics tool, VIVA (Virtual healthcare Interactions using Visual Analytics), to support HLBC in analysing various forms of usage data from the service. We abstract HLBC's data and data analysis tasks, which we use to inform our design of VIVA. We also present the interactive workflow abstraction of Scan, Act, Adapt. We validate VIVA's design through three case studies with stakeholder domain experts. We also propose the Controllability Through Configuration model to conduct and analyze design studies, and discuss architectural evolution of VIVA through that lens. It articulates configuration, both that specified by a developer or technical power user and that constructed automatically through log data from previous interactive sessions, as a bridge between the rigidity of hardwired programming and the time-consuming implementation of full end-user interactivity. Availability: Supplemental materials at https://osf.io/wv38n

Figures (12)

• Figure 1: The HLBC patient call context with different actors (orange) along a patient encounter process, leading to data that is stored in four different data collections (blue). The Refer numbers are relative to the previous pipeline stage, and the Call End numbers are absolute for the whole collection.
• Figure 2: Data sources. Each source characterized by total size (for Apr 2020 through Mar 2021), monthly increase, attribute types (categorical, ordered, quantitative), and high-level analysis goals.
• Figure 3: VIVA full interface, with Case Study 1 data. (a) The full interface of VIVA in Rollup mode after customizing an attribute: the new ClientAgeRangeBinned attribute shown in the Detail Panel has three levels (Young, Middle, Old) and its thumbnail is added to the current Concern attribute set, just after the original Client_Age_Range entry in the Multiples Panel grid. (b) Customization starts by filtering out the NULL level. (c) Customization continues with merging together the last three levels and Rename to old. (d) Partitioning over time triggers a switch from bar to line chart. (e) Changing to from day to week granularity aggregates data, as shown with min/max bars.
• Figure 4: VIVA schematic, and Case Study 1 continuation. (a) Schematic interface diagram: the Multiples Panel on the left contains a scrollable grid of attribute thumbnails (T) and controls on top. The Details Panel on the right features a large detail view of the selected attribute(s), with controls above and below. The purple box depicts the visible screen area. (b) Stratifying age by gender yields a segmented bar chart. (c) Exploding age by gender reifies the stratification, creating a new Concern with one new derived attribute for each original bar, whose levels are the former segments.
• Figure 5: Case Study 1, Step 4: Stratification with segmented bars vs. Sankey flow diagrams, showing triage dispositions. (a) The initial visual encoding in Stratify mode is segmented bars, with bars for the first attribute selected (nurse triage, RNN) and segments for the second one (physician triage, MDD). (b) Selecting relative percentages reveals structure that is not visible with the default absolute counts. (c) If analysts select Flow, the chart type switches to a Sankey diagram, emphasizing the chronological order of the encounter when the patient is handed off from nurse to physician. (d) Reversing the order of selection instead emphasizes where the later recommendation differs from the previous one. (e) Flow mode allows analysts to select three attributes in the Multiples Panel for simultaneous display. (f) Seeing three attributes side by side in the Detail Panel highlights the progression of clinical judgements, from the generic recommendations (KBD) that are sometimes frequently overridden by nurses (RND), and in turn those are frequently changed by physicians (MDD).