MedViz: An Agent-based, Visual-guided Research Assistant for Navigating Biomedical Literature

Biomedical researchers face increasing challenges in navigating millions of publications in diverse domains. Traditional search engines typically return articles as ranked text lists, offering little support for global exploration or in-depth analysis. Although recent advances in generative AI and large language models have shown promise in tasks such as summarization, extraction, and question answering, their dialog-based implementations are poorly integrated with literature search workflows. To address this gap, we introduce MedViz, a visual analytics system that integrates multiple AI agents with interactive visualization to support the exploration of the large-scale biomedical literature. MedViz combines a semantic map of millions of articles with agent-driven functions for querying, summarizing, and hypothesis generation, allowing researchers to iteratively refine questions, identify trends, and uncover hidden connections. By bridging intelligent agents with interactive visualization, MedViz transforms biomedical literature search into a dynamic, exploratory process that accelerates knowledge discovery.

💡 Research Summary

MedViz is a novel visual‑analytics system that re‑imagines biomedical literature search as an interactive, space‑based sense‑making activity. The authors begin by highlighting three fundamental shortcomings of current approaches: (1) traditional search engines such as PubMed or Semantic Scholar return ranked lists of articles, which provide little insight into the global structure of a research field; (2) list‑based interfaces do not reveal how topics relate, where dense or sparse research areas lie, or how themes evolve over time; and (3) recent LLM‑driven chatbot tools (e.g., Elicit, Scite) operate on small, opaque subsets of papers, offering text‑only answers without exposing the underlying retrieval process, thereby limiting transparency and iterative refinement.

To address these gaps, MedViz integrates three tightly coupled components: a large‑scale data‑processing pipeline, an interactive WebGL‑based visual interface, and a context‑aware multi‑agent reasoning framework.

Data pipeline – Users supply a PubMed query result, PMID list, or any curated bibliography. The system fetches metadata, abstracts, and citation data from PubMed, CrossRef, and iCite. Articles are embedded into high‑dimensional semantic vectors using a pretrained sentence‑transformers model. Dimensionality reduction (LargeV is or OpenTSNE) projects these vectors into a 2‑D layout, preserving local neighborhoods. Hierarchical clustering identifies coherent topic groups, and an LLM (GPT‑4) automatically generates human‑readable labels for each cluster. Citation and co‑citation relationships are processed and simplified with edge‑bundling, producing a semantic map that simultaneously encodes topical similarity and relational structure.

Visual analytics interface – The 2‑D map is rendered as a point cloud of up to one million articles using WebGL and Three.js. Points can be colored, sized, or filtered by publication year, citation count, keywords, or other metadata. Users can search, highlight, draw selection boxes, or click clusters to define an explicit analytical context. Temporal filtering is achieved by sliding a year range over the same fixed layout, allowing users to observe topic drift without recomputing the embedding. This “space‑based” view lets researchers perceive the overall landscape, spot under‑explored regions, and compare historical versus contemporary research patterns.

Context‑aware multi‑agent reasoning – At the core is a Scholar agent that acts as an orchestrator. When a user poses a natural‑language question, the Scholar agent parses the query together with the current visual context (selected articles, clusters, time slice). It decomposes the request into subtasks and dispatches them to specialized agents:

• Evidence agent retrieves and structures information (e.g., methods, results, conclusions) from the selected papers.
• Analytical agent performs statistical summarization, trend analysis, or meta‑analytic calculations on the extracted data.
• Discovery agent searches for hidden patterns, gaps, or generates hypothesis‑level insights within the selected semantic region.

Crucially, agents do not merely return textual answers; they can issue UI commands back to the front‑end—highlighting clusters, annotating topic labels, adjusting temporal filters, or pinning representative articles. This tight feedback loop creates an iterative sense‑making cycle: question → visual selection → agent analysis → visual feedback, all grounded in a user‑defined subset of the literature.

Applications – The system supports rapid familiarization with unfamiliar fields (e.g., cancer immunotherapy), focused systematic‑review style evidence extraction, comparative temporal analyses of thematic clusters, and hypothesis generation by exposing under‑studied connections. Because the reasoning is anchored in an explicit visual context, users retain full transparency over which papers influence the answer, mitigating the hidden‑bias problem of autonomous LLM search tools.

Technical stack and extensibility – The pipeline relies on pandas, NumPy, sentence_transformers, LargeV is, and Datashader for scalable preprocessing. The back‑end uses FastAPI and OpenAI APIs to host the agents, while the front‑end combines Vue.js, d3.js, ECharts, and Three.js for interactive rendering. The entire system is containerized, enabling both hosted deployment and local installation.

Future directions – The authors plan to integrate knowledge‑graph representations to enrich semantic relations, incorporate multimodal biomedical data (images, genomics, clinical records), and expand analytical capabilities for automated meta‑analysis and hypothesis testing.

In summary, MedViz introduces a new paradigm that fuses large‑scale visual analytics with context‑aware, multi‑agent AI. By visualizing the entire semantic space of a biomedical corpus and grounding LLM‑driven reasoning in user‑selected visual subsets, the system transforms literature search from a static list‑based retrieval task into an exploratory, transparent, and controllable knowledge‑discovery workflow, substantially reducing cognitive load while enhancing reproducibility and insight generation.