Can we have a (real) doctor, please?bookmark_border

Large language models (LLMs) show promise in medical diagnosis, but real-world deployment remains challenging due to high-stakes clinical decisions and imperfect reasoning reliability. As a result, careful inspection of model behavior is essential for assessing whether diagnostic reasoning is reliable and clinically grounded. However, debugging medical LLMs remains difficult. First, developers often lack sufficient medical domain expertise to interpret model errors in clinically meaningful terms. Second, models can fail across a large and diverse set of instances involving different input types, tasks, and reasoning steps, making it challenging for developers to prioritize which errors deserve focused inspection. Third, developers struggle to identify recurring error patterns across cases, as existing debugging practices are largely instance-centric and rely on manual inspection of isolated failures. To address these challenges, we present VeriLLMed, a visual analytics system that integrates external biomedical knowledge to audit and debug medical LLM diagnostic reasoning. VeriLLMed transforms model outputs into comparable reasoning paths, constructs knowledge graph-grounded reference paths, and identifies three recurring classes of diagnosis errors: missing errors, relation errors, and branch errors. Case studies and expert evaluation demonstrate that VeriLLMed helps developers identify clinically implausible reasoning and generate actionable insights that can inform the improvement of medical LLMs.

Patients often struggle to communicate coherent accounts of their health histories during time-constrained clinical encounters. These accounts, which we refer to as health stories, include both clinical events and lived experiences. Existing systems prioritize structured, clinician-centered data and provide limited support for eliciting and communicating patient-generated narratives. We present HealthTale, a patient-centric visualization system designed to elicit health stories from patients and structure them to facilitate communication during initial clinical conversations. Its design arises from a multi-stage qualitative investigation across domain expert discussions (n=3), online narratives (n=20), patient (n=11) and clinician (n=6) interviews, and elicited health stories (n=22), identifying recurring patterns in how individuals construct and communicate their health stories. HealthTale transforms freeform narratives into structured timeline representations, grounded in a data abstraction that models health stories as events grouped by health concern and time. The abstraction captures both clinical and contextual information while accommodating temporally imprecise data and non-linear distributions of events across time. Through evaluation with patients (n=34) and clinicians (n=3), we find that HealthTale supports recall, organization, and self-advocacy, while enabling clinicians to rapidly interpret patient-generated narratives and establish a shared understanding.

Gait disorders are commonly observed among older adults, who frequently experience various issues related to walking. Additionally, researchers and clinicians extensively in- vestigate mobility related to gait in normally and abnormally developing children, athletes, and individuals with orthopedic and neurological disorders. Effective gait analysis enables us to understand the causal mechanisms of mobility and balance control of patients, develop tailored treatment plans to improve mobility, reduce fall risks, and track rehabilitation progress. However, analyzing gait data is a complex task due to the multivariate nature of the data, the large volume of information to be interpreted, and the technical skills required. Existing tools for gait analysis are often limited to specific patient groups (e.g., cerebral palsy), only handle a specific subset of tasks in the entire workflow, and are not openly accessible. To address these shortcomings, we conducted a requirements assessment with gait practitioners (e.g., researchers, clinicians) via survey methods and identified key components of the workflow including (1) data processing and (2) data analysis and visualization. Based on the findings, we designed VIGMA, an open-access visual analytics framework integrated with computational notebooks and a Python library, to meet the identified requirements. Notably, the framework supports analytical capabilities of disease progression and comparisons between multiple patient groups, which are lacking in existing tools. We validated the work through three usage scenarios with experts specializing in gait and mobility rehabilitation. VIGMA is available at github.com/komar41/VIGMA.

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.

Cancer care is a longitudinal, time-sensitive process that depends on the coordinated activities of diverse healthcare professionals (HCPs), whose patterns of collaboration evolve continuously throughout the course of treatment. Prior studies have shown that care team structure is associated with patient outcomes; however, the temporal dynamics of these collaborations and their relationship to evolving patient risk remain insufficiently understood. To address this gap, we introduce a visual analytics framework for modeling and interpreting time-varying care team interactions in oncology. We construct dynamic collaboration networks from electronic health record (EHR)-mediated communication data and employ a temporal graph neural network (TGNN) to estimate patient outcome risk over time. Building on these predictions, we examine how changes in care team composition and specialist participation are associated with shifts in estimated risk across different treatment stages. To support practitioner-driven analysis, we design a multi-view interactive system that integrates TGNN-based risk modeling with coordinated visualizations of care team networks, feature-level explanations derived from a surrogate logistic regression model, and exploratory what-if analysis. Rather than inferring causality, the system enables analysts to investigate how risk estimates co-evolve with collaboration patterns and to identify the care team characteristics most strongly associated with these temporal changes. We demonstrate the utility of the system through expert-guided analysis scenarios on a cohort of 505 cancer patients spanning breast, colorectal, and lung cancers, and evaluate its effectiveness in uncovering clinically meaningful collaboration patterns. This work contributes a practical approach for improving the interpretability of temporal graph-based models in longitudinal, team-based healthcare settings, with broader applicability to domains involving complex, evolving collaborative processes.

Public health organizations regularly produce and publish data visualizations to raise awareness of critical issues, influence decision-making processes, and promote overall well-being. However, the design practices shaping these visualizations in real-world settings remain largely unexamined, limiting the research community's ability to evaluate their effectiveness, accessibility, and alignment with communication goals. To address this gap, we construct and analyze a large-scale corpus of over 4,000 real-world data visualizations drawn from more than two dozen websites associated with U.S. and international public health organizations. We evaluate salient design characteristics like chart type, visualization accessibility, use of embellishments like iconography, and design flaws. This work contributes to understanding real-world decisions in designing data visualizations and supports public health officials in improving data visualization-related communications. Visualizations in our finalized corpus and the labeled dataset can be found at https://washuvis.github.io/ChartingPublicHealth/.

Narrative visualization embeds data in visual stories to make medical information more relatable for non-experts. Despite the growing use of character elements in health communication, evidence on whether individual characters support affective responses remains inconclusive. Physiological evidence independent of verbal self-report is especially scarce, although such measures should complement participants’ self-reported experiences. We present a mixed-methods study using electrodermal activity (EDA), eye tracking, and questionnaires to compare two medical data stories: an individual, character-based version and a general, population-level version without an individual protagonist. We examine how this framing influences physiological arousal, visual attention, viewing behavior, and self-reported affective response. Story-level EDA comparisons provide only limited support for stronger arousal in the character-based story. Stronger evidence comes from eye-tracking-based peak classification, where character illustrations were robustly associated with EDA peaks, and from questionnaire responses showing more negative empathy-related emotions for the individual story. The general story elicited more awe and joy, suggesting that individual and population-level framings may support different emotional qualities. We also observed a preliminary story-order effect: participants who first saw the individual story showed higher peak-based physiological arousal, although this effect cannot be fully disentangled from fatigue, novelty, or learning effects. By combining physiological, gaze-based, questionnaire, and lightweight qualitative evidence, our work advances time-resolved assessment of narrative medical visualization and highlights the need to interpret arousal, attention, curiosity, and self-reported emotions together.