MExplore: Visual Analytics for Illness Scripts

• MExplore is an interactive visual analytics system that converts unstructured medical texts into structured illness scripts.
• It employs a fine-tuned BERT-based pipeline for named-entity extraction, achieving over 88% precision, recall, and F1 scores.
• The system provides multi-level visualizations that reduce cognitive load and guide users through diagnostic schema construction.

Searching arXiv for the specified paper and closely related work to ground the article. arXiv search query: (Pang et al., 16 Jul 2025) MExplore is an interactive visual analytics system for medical expertise acquisition from unstructured medical texts. It is designed to support the construction of “illness-script” based reasoning from narrative documents such as case reports and clinical notes, which contain rich clinical detail but also impose high cognitive load, confidentiality constraints, and substantial variability in writing style. The system combines a fine-tuned BERT-based named-entity extraction pipeline with a four-panel, multi-level visual analytics framework that supports progressive, interactive exploration of medical knowledge. Its central mechanism is to extract nine categories of medical entities and organize them as interpretable “knowledge units” for identifying core entities, revealing their interconnections, and assembling comprehensive illness scripts (Pang et al., 16 Jul 2025).

1. Conceptual basis and design objectives

MExplore is motivated by the need to help medical learners build illness scripts from the vast, unstructured texts that permeate clinical practice and the biomedical literature. In the formulation used by the system, illness scripts are the canonical cognitive structures through which clinicians link enabling conditions (EC), pathophysiological faults (FT), and clinical consequences (CQ). The system therefore treats medical expertise acquisition not as passive reading, but as a guided process of extracting and organizing clinically salient knowledge units.

Three obstacles motivate the design. First, the sheer volume of narrative documents and their variable writing styles impose high cognitive load. Second, data confidentiality and privacy concerns limit the use of large third-party LLMs. Third, existing tools tend to present either raw text or static summaries without guiding learners through the process of constructing diagnostic schemas. MExplore addresses these gaps by combining entity extraction with coordinated visual exploration so that learners can move from high-level browsing to focused analysis.

The system’s stated goals are to reduce extraneous cognitive load by filtering nonessential text, provide multi-granular overviews and drill-down paths, and maintain data privacy by employing a lightweight, locally deployable BERT variant rather than off-the-shelf LLMs that may hallucinate or leak sensitive information. A plausible implication is that MExplore is positioned as a human-in-the-loop analytic environment rather than an automatic diagnostic assistant.

2. Extraction pipeline and representation learning

MExplore comprises two principal subsystems: an automated extraction and relationship-computation pipeline, and a multi-view visual analytics front end. The extraction subsystem is built around a fine-tuned Mac-BERT named-entity recognition model (Pang et al., 16 Jul 2025).

Mac-BERT serves as the backbone, and a token-classification head is appended as a single linear layer mapping the hidden representation $h_i \in \mathbb{R}^d$ to logits over the nine entity classes plus “O” for non-entity. Training uses the standard cross-entropy loss

$$L = - \frac{1}{N} \sum_i \sum_k y_{i,k} \log p_{i,k},$$

where $y_{i,k} \in \{0,1\}$ indicates the ground-truth label for token $i$ and class $k$, and $p_{i,k}$ is the softmax probability. The model is fine-tuned on the publicly available CBLUE benchmark, which contains 47194 paragraphs and nine entity types, using typical hyperparameters of learning rate $\approx 2 \times 10^{-5}$, batch size $\approx 32$, and 3 epochs. On held-out CBLUE validation data, the model achieves Precision, Recall, and $F_1$ scores above 88%, with the usual definitions

$$\mathrm{Precision} = \frac{TP}{TP+FP}, \qquad \mathrm{Recall} = \frac{TP}{TP+FN}, \qquad F_1 = \frac{2 \cdot \mathrm{Precision} \cdot \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}.$$

Each medical document $$L = - \frac{1}{N} \sum_i \sum_k y_{i,k} \log p_{i,k},$$0 is split into paragraphs $$L = - \frac{1}{N} \sum_i \sum_k y_{i,k} \log p_{i,k},$$1. Mac-BERT extracts entity mentions and associated embeddings for each paragraph, forming Medical-Entity Sets (MES). Paragraph vectors are then concatenated per document and passed through SimCSE to compute pairwise document similarity. Topic inference is performed with BERTopic by averaging the MES embeddings of each document as its topic embedding. This pipeline turns unstructured narratives into an entity-centric representation that can be visualized across multiple scales.

3. Entity ontology and illness-script units

The system recognizes nine standardized medical entity types, each treated as a knowledge unit for downstream exploration.

Abbreviation	Entity type
dis	specific diseases or diagnoses
sym	clinical symptoms and signs
dru	pharmacological agents and medications
equ	instruments and diagnostic equipment
pro	surgical or therapeutic procedures
bod	anatomical structures
ite	laboratory tests and imaging studies
mic	pathogenic microorganisms
dep	clinical departments or specialties

Formally, each paragraph $$L = - \frac{1}{N} \sum_i \sum_k y_{i,k} \log p_{i,k},$$2 produces a set

$$L = - \frac{1}{N} \sum_i \sum_k y_{i,k} \log p_{i,k},$$3

where each entity $$L = - \frac{1}{N} \sum_i \sum_k y_{i,k} \log p_{i,k},$$4 and $$L = - \frac{1}{N} \sum_i \sum_k y_{i,k} \log p_{i,k},$$5. MESs become nodes in later visual structures, especially the radial tree in the association analysis view, with leaf children corresponding to individual entities.

This entity ontology is not merely a tagging schema. It provides the building blocks through which learners can organize enabling conditions, pathophysiological faults, and clinical consequences into illness scripts. The paper’s framing suggests that the MES abstraction is intended to bridge low-level textual mention extraction and higher-level clinical reasoning.

4. Multi-level visual analytics framework

The visual front end implements a “cosmos-→star map→planet core” metaphor across four coordinated views, enabling progressive exploration from a corpus overview down to detailed entity relationships (Pang et al., 16 Jul 2025).

The MD Space View provides a cosmic overview. Each document is rendered as a 3D node, described as a “celestial body,” sized by text length and colored by topic. A force-directed layout uses similarity edges, with gravity proportional to the SimCSE score, so that related documents cluster spatially. Users can search keywords, adjust a similarity threshold slider, rotate and zoom the 3D view, and lasso-select subsets for deeper analysis.

The MP Star Map decomposes selected documents into paragraphs, represented as stars. Intra-document edges $$L = - \frac{1}{N} \sum_i \sum_k y_{i,k} \log p_{i,k},$$6 link paragraphs from the same file, while similarity edges $$L = - \frac{1}{N} \sum_i \sum_k y_{i,k} \log p_{i,k},$$7 with SimCSE $$L = - \frac{1}{N} \sum_i \sum_k y_{i,k} \log p_{i,k},$$8 connect cross-document semantically related paragraphs. The KaFFPa graph-partitioning algorithm is applied to create subgraphs that likely share a common theme. For each star $$L = - \frac{1}{N} \sum_i \sum_k y_{i,k} \log p_{i,k},$$9, position is determined by three forces:

$y_{i,k} \in \{0,1\}$0

$y_{i,k} \in \{0,1\}$1

with $y_{i,k} \in \{0,1\}$2 denoting the nine poles for entity types, together with a collision force $y_{i,k} \in \{0,1\}$3 to prevent overlap. The combined force is

$y_{i,k} \in \{0,1\}$4

Stars in the same subgraph share color, brightness gradients encode paragraph counts, and border luminance indicates the number of entities in each paragraph. Hover and click operations reveal tooltips and allow subgraph selection.

The Association Analysis View organizes the MESs of a selected paragraph-subgraph into a radial tree. Non-leaf nodes represent MESs, and leaf nodes are individual entities colored by type. MESs sharing common entities are grouped under shared branches through a custom tree-construction algorithm. This makes co-occurrence and inclusion relationships explicit and supports the tracing of associations that are central to illness-script formation.

The Focused Sectional View offers a drill-down on any selected MES or individual medical entity. The innermost ring shows entity-type proportions $y_{i,k} \in \{0,1\}$5 for class $y_{i,k} \in \{0,1\}$6. The surrounding “mantle” is a radial area chart in which each axis corresponds to a co-occurring MES, divided by entity types along the polar axes, while areas between axes link same-type entities across MESs. The paper states that this dual encoding, verbal plus imaginal, reinforces memory traces per multimedia-learning theory and reveals the target’s contextual signature.

Across all panels, users can zoom, pan, rotate, search, adjust thresholds, highlight elements, and link selections synchronously. The intended progression is from hundreds of documents to the detailed relationships among a handful of entities.

5. Evaluation: case studies, user study, and expert interviews

MExplore was evaluated through three case studies, a user study, and interviews with domain experts. In the reported case-study workflow, two senior clinicians, identified as E3 and E5, used the system to develop an illness script for medication-related osteonecrosis of the jaw (MRONJ) (Pang et al., 16 Jul 2025). Their work proceeded through three successive tasks: identification of areas of interest via MD Space and MP Star Map to find oral-surgery-related clusters; association analysis to isolate key EC factors such as bisphosphonate use and cancer history together with FT and CQ factors; and focused analysis of zoledronic acid, tooth extraction, and imaging features. The experts reported that the entity-based visual representations reduced cognitive overload and provided trustable contexts, outperforming ad-hoc search.

The user study involved twenty second-year medical students, evenly split by sex, who were unfamiliar with oral candidiasis, meningitis, and herpes zoster. Participants were randomized into an MExplore group, labeled MEX, and a traditional-resource group, labeled OTH. After a 20-minute system tutorial, each participant had 30 minutes to construct illness scripts identifying EC, FT, and CQ from anonymized case texts. Scoring awarded one point per correct information unit, with a maximum of 100%, and a surprise recall test was administered two weeks later.

The reported results were as follows. Script accuracy for MEX averaged 87–93% versus 78–88% for OTH, corresponding to a 3–6% absolute gain. Retention was 81–86% for MEX versus 69–75% for OTH, corresponding to a 6–13% gain. Completion time was 10–13 minutes for MEX versus 12–15 minutes for OTH, or 1.5–3 minutes faster. The paper explicitly notes that these differences were not formally tested for significance, but reports that they indicate consistent improvements in both efficiency and long-term recall.

The expert interview component involved ten domain experts, E1–E10, completing an 11-item Likert questionnaire. Median ratings were in the “agree” to “strongly agree” range, highlighting ease of learning, intuitive visual design, and effectiveness in extracting core knowledge units, supporting structured cascading exploration, revealing interconnections, and enabling focused analysis.

6. Significance, limitations, and broader implications

MExplore demonstrates an entity-driven approach for bridging unstructured clinical narratives and the structured cognitive frameworks that novice learners must build. Its central contribution is a carefully engineered combination of fine-tuned NER, contrastive embedding, topic modeling, and multi-view visual design, culminating in what the paper describes as an effective human-computer partnership for medical expertise acquisition (Pang et al., 16 Jul 2025).

A key distinction from alternative approaches is the retention of human-in-the-loop control. Rather than outsourcing reasoning to opaque LLMs, the system requires learners to engage actively with extracted entities and their relationships while constructing their own illness scripts. This suggests a pedagogical emphasis on guided knowledge organization rather than answer generation.

The paper also identifies several limitations. The current implementation focuses on two Chinese datasets, CBLUE and West China Stomatology. Occasional latency arises when rendering very large subgraphs. Future work is described as expanding to additional institutions, optimizing force-layout algorithms for real-time performance, and exploring semi-supervised model adaptation to new entity classes.

The broader implications are framed cautiously. The modular pipeline and abstract metaphors are described as lending themselves to adaptation in other domains, including legal texts and patent analysis. A plausible implication is that the architecture is intended as a general pattern for transforming unstructured domain narratives into entity-centric exploratory environments, with medical expertise acquisition as the initial demonstration.