Introduction
Acquiring medical expertise is a complex process of building "illness scripts"—mental models that link a patient’s history and enabling conditions to pathophysiological faults and clinical consequences. While Electronic Medical Records (EMRs) and medical literature are rich sources for these scripts, they are largely unstructured and overwhelming for novice learners. This creates a "reading bottleneck," where the cognitive load of identifying salient information in dense text hinders deep comprehension.
MExplore is a visual analytics system designed to bridge this gap. By combining automated entity extraction with a multi-level metaphorical interface, the system provides a structured, interactive pathway for students to explore and synthesize expertise from raw clinical text without the risks of "deceptive expertise" (hallucinations) often associated with fully automated summarization.
The MExplore Framework
The system employs a workflow that transitions from raw data to structured visual exploration through three primary stages:
- Entity Extraction: A fine-tuned Mac-BERT model extracts Medical Entities (MEs) across nine categories, including diseases, symptoms, and pharmacological agents.
- Relationship Computation: The system utilizes SimCSE for document similarity and BERTopic for thematic classification, allowing for the clustering of related clinical cases.
- Multi-level Visual Metaphors: To reduce extraneous cognitive load, the interface maps complex medical data onto familiar spatial schemas.
Visual Design Levels
- MD Space View (Cosmic Metaphor): Documents are represented as celestial bodies in a 3D layout where proximity indicates semantic similarity. This provides a high-level overview of the document landscape.
- MP Star Map (Constellation Metaphor): Paragraph-level subgraphs are clustered into constellations, visualizing how specific medical concepts (stars) connect within a chosen topic.
- Focused Sectional View (Planet Metaphor): A radial "cross-section" of a specific entity or paragraph set, showing the distribution of associated symptoms, treatments, and co-occurring entities.
Datasets and Implementation
MExplore was developed and evaluated using two major datasets:
- CBLUE (Chinese Biomedical Language Understanding Evaluation): A benchmark dataset used for fine-tuning the extraction models.
- MWCSS (Medical text of West China School of Stomatology): A real-world dataset comprising approximately 100,000 unstructured medical documents (800,000 paragraphs) covering specialized clinical diagnoses and treatment plans.
Key Results and Evaluation
A randomized controlled trial with 20 medical students compared MExplore (MEX) against traditional study methods (OTH), such as textbooks and search engines. The results demonstrated significant advantages for the visual analytics approach:
- Higher Accuracy: The MEX group achieved 3–6% higher accuracy in constructing illness scripts across various disease categories.
- Superior Retention: In follow-up testing two weeks later, the MEX group outperformed the control group by 6–13% in retention accuracy, suggesting that active visual exploration leads to more durable mental models.
- Efficiency: Despite the interface's complexity, the MEX group completed tasks 1.5 to 3 minutes faster on average, with significantly less disorientation than those using traditional methods.
Conclusion and Future Work
MExplore demonstrates that "disaggregating" text into component entities through visual metaphors can enhance long-term memory and expertise acquisition. While currently focused on stomatology, the methodology is generalizable to other medical specialties and potentially other fields requiring structured knowledge exploration, such as legal or engineering research. Future work will focus on optimizing algorithms for larger datasets and adding "exploration playback" features to help students review their learning paths.
Summary updated