A Design Study Process Model for Medical Visualization

AI summary

Overview

Research area: Human-Computer Interaction — specifically visualization design-study methodology, applied to medical visualization and visual analytics.

Technical level: Intermediate. The paper is a process-model / methodology contribution rather than an algorithm or system paper, but it assumes familiarity with visualization design-study vocabulary (design study, task abstraction, evaluation methods such as controlled studies and case studies).

Scope: The paper proposes a design study process model tailored to medical visualization, derived from a review of 78 medical-related visualization papers, and applies it to four use cases (one guided design and three reanalyses; only two of the reanalyses appear in the truncated content).

What This Paper Is About

Existing design study process models (such as Munzner's four-level nested model, Sedlmair et al.'s nine-stage framework, and the design study "Lite" methodology) are generic, aiming at a wide range of application areas rather than medicine specifically. The authors argue that this generality makes it hard for researchers to handle the complexity of medical problems and limits the efficiency of interdisciplinary collaboration between visualization researchers and medical experts. The paper therefore builds a medical-specific process model that adds explicit handling of three factors — stakeholders, stages, and subgroup analysis — and provides step-by-step guidance from collaborator selection through evaluation and promotion.

Key Contributions

1. A complete design study process model for medical visualization covering the whole process from collaborator selection to final evaluation, organized into eight steps: (1) select collaborators and identify domain problems; (2) identify stakeholders; (3) locate the analysis goals; (4) dismantle the analysis goals; (5) design visualization; (6) implement prototype; (7) evaluate; (8) promote. The model introduces three factors the authors identify as critical for medical visualization: stakeholders, stages, and subgroup analysis.

2. Identification of three features of medical visualization research through a literature review — a medical problem may involve multiple stakeholders; analysis and resolution may need to be carried out in stages; and some problems focus on differences between subgroups.

3. A refined task and stage taxonomy that distinguishes design stages according to analytic logic versus cognitive habits, and classifies tasks as inferential or descriptive, with inferential tasks further split into hypothesis-based (when multiple subgroups are involved) and hypothesis-free. This taxonomy is expressed per target user group (professionals, semi-professionals, non-professionals).

4. Practical recommendations and demonstrated application, including recommendations that researchers conduct thorough discussions with collaborators, determine stakeholders and target users before designing, and prioritize a controlled study for evaluation while using a rigorous method such as a pilot study to help calculate the minimum sample size of the controlled study. The model is applied to guide one design and to reanalyze three medical visualization works, and the authors report reflections on the model and delineate it from existing models.

Main Findings

• Literature review scale and funnel: The search of IEEE TVCG, ACM CHI, Computers & Graphics, and Computer Graphics Forum spanning 2020 to 2025 yielded 325, 11, 67, and 20 articles respectively; adding all papers published in the Eurographics Symposium on Visual Computing for Biology and Medicine (VCBM) between 2020 and 2024 (74 articles) gave a combined total of 497 articles from five sources. Title and keyword screening refined this to 175 articles; abstract screening excluded a further 97, leaving 78 articles for full-text assessment. Title and keyword screening excluded 322 non-eligible articles.

• Three features are common in the reviewed literature: Among the 78 reviewed papers, 62 involved multiple stakeholders, 33 involved multiple stages, and 38 involved multiple subgroups.

• Stakeholder and target user are not the same thing: Of the 62 works involving multiple stakeholders, the authors found that relevant tools or systems are designed for a specific role — the target user — and not always for all stakeholders, especially when multiple stakeholders have inconsistent levels of expertise. They map three combinations to three application types: professionals as target users (with various-background professionals as stakeholders) for insight analysis and decision support; semi-professionals such as students (with professionals and semi-professionals as stakeholders) for medical education; and non-professionals such as patients (with professionals and non-professionals as stakeholders) for information communication.

• Data and user coverage of the reviewed works: Reviewed works span data from molecular to individual to population levels, including genes, proteins, CT, MRI, electronic health records, patient self-recorded data, prescription data, and population mobility data. Users are mainly medical professionals such as surgeons, radiologists, (bio)medical experts, microscopists, and pathologists, with some works involving non-medical professionals such as patients and the general public, and semi-professionals such as medical students.

• Stage differentiation depends on target users: Stages for professionals follow analytic logic, where later-stage analysis depends on earlier-stage analysis. For semi-professionals, passive learning can be staged by cognitive habits (for example, from easy to difficult, or from diagnosis to treatment to prognosis), while active learning involves a data analysis process and can be staged by analytic logic. For non-professionals such as patients or the general public, staging by cognitive habits is beneficial.

• Task type depends on stages and subgroups: Tasks within a stage are defined as descriptive when stages are distinguished by cognitive habits (passive student learning, or information communication to non-professionals), and as inferential when stages are distinguished by analytic logic (active student learning, or professional insight analysis and decision support). Inferential tasks are further distinguished as hypothesis-based if different subgroups are involved and hypothesis-free otherwise; a hypothesis-based approach is recommended when subgroups are involved.

• Evaluation methods and evidence levels: Commonly used comprehensive evaluation approaches are controlled studies (described as highly recognized in medicine, quantitative or qualitative and usually quantitative), case studies (less common, usually qualitative), and user evaluations (less common, usually qualitative). Each can combine methods such as dashboard comparison, insight-based evaluation, heuristic evaluation, eye-tracking, interviews and focus groups, and standardized questionnaires. A single evaluation method is usually not sufficient, so multiple methods need to be combined.

• Design study tends to have a higher success rate than human-centered design in interdisciplinary settings: 63% vs. 25%, as cited from prior work.

• Use case — multi-outcome causal graphs: The model guided a prior work on visual analysis of multi-outcome causal graphs. In collaboration with a clinical expert, the domain problem was defined as analyzing interactions between multiple disease outcomes and their influencing factors. Two relevant datasets were selected, stakeholders were identified as clinical and public health experts, and their goals were decomposed into two distinct analysis stages and six specific tasks. Evaluation used quantitative metrics on benchmark data, a qualitative case study (N = 1), and expert user evaluation (N = 3). The authors note that if a solution only focused on public health experts it might weaken detailed comparison among diseases needed for clinical decision-making, while over-leaning toward clinical experts might neglect overall multimorbidity patterns.

• Use case — PROACT: Reanalyzing PROACT, a tool for communicating health risks to prostate cancer patients, the authors note that urologists were selected as collaborators, urologists and prostate cancer patients were naturally determined as stakeholders, and two rounds of discussions (mainly with urologists) determined two analysis goals and an initial six-page prototype combining visualization techniques such as pie charts and bar charts, a narrative structure, and clinical prediction models. Patients (N = 6) and doctors (N = 2) evaluated the initial prototype via semi-structured interview, producing an iterative redesign and a revised ten-page prototype; a second evaluation used 6 new patients and the same 2 urologists. The authors' critique is that although both doctors and patients were ultimately involved, the initial design focused primarily on doctors' perspective and needs, with patient feedback considered only at the evaluation stage.

• Not reported: Specific numeric results of the quantitative benchmark evaluation in the multi-outcome causal graph study are not given in the truncated content, and the third reanalyzed work is not present in the truncated content.

Methodology in Plain English

The authors used a two-part approach.

First, they conducted a structured literature review. They jointly agreed on inclusion and exclusion criteria, cross-validated the screening process, and resolved coding discrepancies through discussion. They excluded review/survey papers; application papers related to mixed reality, virtual reality, and augmented reality; non-medical-related visualization; scientific visualization (such as work focusing on image segmentation and calibration); studies involving non-human species; and guidelines-related papers. They searched five key publication venues with the query ((“biomedical” OR “clinical” OR “disease” OR “health” OR “healthcare” OR “medicine” OR “medical”) AND (“visualization” OR “visual analytics”)), narrowing 497 articles down to 78 for full-text assessment with a primary focus on each paper's methodological contributions. From close reading, they identified recurring structural features — multiple stakeholders, multiple stages, multiple subgroups — and counted how often each appeared.

Second, they used those features, together with their own interdisciplinary research experience and insights from earlier design study process models and practical guidelines, to formulate the model. Each of the eight steps comes with referable guidance. The new elements introduced by the model are highlighted in the model figure and summarized in two tables: one mapping stakeholders to target users to application types, and one mapping target users to stages, subgroups, and task types.

Third, they demonstrated the model in practice by applying it to guide the design of one visual analysis method and by retrospectively reanalyzing existing medical visualization works through the lens of the model.

Why This Matters

The paper addresses a real methodological gap: medical visualization projects must reconcile several audiences with different expertise, proceed through analyses that depend on one another, and frequently center on subgroup differences such as healthy versus pathological cohorts, yet generic process models do not spell out how to handle these. By making the medical specifics explicit and operational, the model aims to make visualization design targeted, generalizable, and operational, and to reduce the friction of visualization–medicine interdisciplinary collaboration.

Real-world applications:

• Clinical decision support tools for professionals, including staged analyses such as acute ischemic stroke analysis pipelines driven by different stages of an analysis pipeline.
• Patient-facing and public-facing risk communication, as in the prostate cancer risk communication tool PROACT, where narrative structure and clinical prediction models are combined with visualizations.
• Medical education tools for students, staged either by cognitive habits (from easy to difficult) or by analytic logic.
• Public health and multimorbidity analysis, where population-level pattern insight must be balanced against individual clinical case comparison, as in the multi-outcome causal graph work.

Industry relevance: The model is directly relevant to healthcare IT and clinical software teams building decision-support dashboards; to companies developing patient education and shared decision-making products; to health data science groups working with electronic health records, prescription data, or population mobility data; and to medical device or imaging software vendors whose tools must serve radiologists, pathologists, and other specialists with different needs and expertise levels.

Future Directions

• Complete and balance the use-case set. The introduction states four use cases of medical visualization works, and the abstract describes guiding one visual analysis design and reanalyzing three works, but only two reanalyses appear in the truncated content; the remaining case and its lessons are an open item for readers to check in the full paper.
• Detailed comparison with existing models. The paper points to a comparative analysis of its model and existing models in Section 6.2; the specifics of how it differs step by step from the four-level nested model, the nine-stage framework, the design study "Lite" methodology, and the design activity framework are not detailed in the truncated content.
• Validation of the proposed task taxonomy. Whether the inferential/descriptive and hypothesis-based/hypothesis-free distinctions reliably improve outcomes compared with existing task classifications is a question the paper's recommendation-based framing leaves for empirical testing.
• Evaluation rigor in medical visualization. The authors recommend prioritizing controlled studies and using a pilot study to calculate the minimum sample size, which raises the open question of how practical these highly recognized but resource-intensive evaluations are for typical medical visualization design studies.

Target Audience

Visualization researchers and PhD students conducting problem-driven, interdisciplinary design studies with medical collaborators; medical informatics and health data science researchers who commission or co-design visualization tools; HCI methodologists interested in process models and task taxonomies; and practitioners in clinical software, patient education, or public health analytics who need a structured way to move from a medical problem to an evaluated visualization solution.

Authors’ abstract

We introduce a design study process model for medical visualization based on the analysis of existing medical visualization and visual analysis works, and our own interdisciplinary research experience. With a literature review of related works covering various data types and applications, we identify features of medical visualization and visual analysis research and formulate our model thereafter. Compared to previous design study process models, our new model emphasizes: distinguishing between different stakeholders and target users before initiating specific designs, distinguishing design stages according to analytic logic or cognitive habits, and classifying task types as inferential or descriptive, and further hypothesis-based or hypothesis-free based on whether they involve multiple subgroups. In addition, our model refines previous models according to the characteristics of medical problems and provides referable guidance for each step. These improvements make the visualization design targeted, generalizable, and operational, which can adapt to the complexity and diversity of medical problems. We apply this model to guide the design of a visual analysis method and reanalyze three medical visualization-related works. These examples suggest that the new process model can provide a systematic theoretical framework and practical guidance for interdisciplinary medical visualization research. We give recommendations that future researchers can refer to, report on reflections on the model, and delineate it from existing models.