We used a user-centered design process [27], as illustrated in Figure 1. Initially, we collaborated with domain experts to analyze requirements and define specific visual analytics tasks. Guided by these tasks, we established the hierarchy, structural organization, and computational metrics for the textual data. We subsequently designed and developed a multilevel visual analytics framework tailored to these data specifications. Throughout the development lifecycle, we engaged in close collaboration with experts and users, conducting multiple rounds of testing and validation. The feedback and recommendations from these sessions drove iterative redesigns, ensuring continuous system optimization.

The system must automatically extract MEs from raw medical texts and categorize them. This transforms unstructured data into discrete, manageable knowledge units (R1), which serve as the foundational elements for visualization.

The system should organize extracted data into a hierarchical structure. It must provide cascading views that allow users to transition seamlessly from a high-level thematic overview to coarse-grained paragraph relations, and finally to fine-grained entity details. This capability facilitates incremental exploration and understanding of complex medical knowledge (R2).

The system needs to construct and visualize a clear mapping of relationships, specifically the inclusion of MEs within paragraphs and the co-occurrence of elements. This visualization should clarify how knowledge units are structurally connected (R3).

The system must provide interaction mechanisms to select a specific ME or MES as a focal point. Upon selection, the visualization should be dynamically reconfigured to display the distribution, composition, and associated context centered around the target knowledge unit (R4).

In this paper, we analyze and process 2 realistic datasets. The first dataset is the Chinese Biomedical Language Understanding Evaluation (CBLUE) provided by the Key Laboratory of Computational Linguistics (Peking University, China) [41]. CBLUE is a comprehensive benchmark compiled from authoritative medical textbooks and clinical practice records. The dataset contains approximately 96,000 MDs, covering a wide range of more than 500 common diseases. To ensure high quality, the MEs within the dataset were meticulously annotated by medical experts. These MEs are categorized into 9 classes on the basis of Chinese ME annotation standards [41]: disease (dis), clinical symptoms (sym), drugs (dru), medical equipment (equ), medical procedures (pro), body (bod), medical examination items (ite), microorganisms (mic), and department (dep).

The second dataset, the Medical Text of West China School (Hospital) of Stomatology (MWCSS), was collected between 2023 and 2025 and provided by the West China School (Hospital) of Stomatology (Sichuan University, China). This dataset contains a large volume of unstructured text data from clinical practice. It includes approximately 100,000 MDs, covering various facets of medical processes, such as clinical diagnoses, treatment processes, specialist examinations, auxiliary examinations, treatment plans, interventions, and drug instructions.

The accuracy of ME extraction is critical to the effectiveness of data processing and visual analytics. Therefore, we systematically assessed the performance and computational resource demands of several models for ME extraction.

Due to strict data confidentiality protocols, which preclude the transmission of data to external servers, models must be deployable locally. In light of the practical limitations of resource-constrained environments, and based on the latest benchmark list from FlagEval [42], we selected high-performing, open-source LLMs for evaluation: DeepSeek-70B and Qwen-32B. Additionally, we included MacBERT, an enhanced BERT model with a novel masked language modeling correction pretraining task [43], which has been identified as the top-performing BERT-based model according to the CBLUE benchmark, for comparison.

Fine-tuning and inference were performed via two NVIDIA A10 GPUs or a single NVIDIA RTX 4060 Ti GPU for MacBERT. The fine-tuning parameters were as follows: learning rate of 3e-5, 2 epochs, batch size of 16, and warm-up ratio of 0.1. For DeepSeek-70B and Qwen-32B, both models required two A10 GPUs and were instruction-tuned via 4-bit quantized models with QLoRA (Quantized Low-Rank Adaptation) [44] (LoRA [Low-Rank Adaptation] settings: r=16, α=64, dropout=0.05). All the fine-tuning parameters and LLM prompt words were selected based on benchmark tests [41] and relevant comparative experiments [45] and have been validated. The computational resource consumption for each model is summarized in Table 1.

The results highlight a significant disparity in resource demand. While LLMs (DeepSeek-70B and Qwen-32B) require substantial fine-tuning and inference time on dual A10 GPUs, MacBERT demonstrates remarkable efficiency, completing the same tasks in a fraction of the time. Notably, MacBERT’s inference speed on consumer-grade hardware (a single RTX 4060 Ti) suggests that it is highly suitable for local deployment in practical medical environments where high-end computational clusters may not be available.

Table 2 summarizes the performance metrics for each model. Fine-tuning consistently improved performance across all the models. Notably, the fine-tuned MacBERT model achieves the highest F₁-score (0.605), surpassing both LLMs, and its precision approaches human annotation levels [41].

This analysis demonstrates that MacBERT not only delivers the best performance after fine-tuning but also requires significantly fewer computational resources than the LLMs. MacBERT, therefore, emerges as the optimal choice for ME extraction tasks. This conclusion aligns with the literature [46], which emphasizes that while LLMs excel in generative tasks, BERT-based models retain distinct advantages in NER and other specialized domains [47] and sentiment analysis [48].

Beyond common medical conditions, the system’s robustness concerning rare diseases and less common terminology warrants further discussion. Although our evaluation used a dataset of common diseases—where MacBERT demonstrated strong performance—recent research [49] suggests that the performance landscape may shift when encountering “long-tail” medical data. Specifically, while fine-tuned BERT-based models generally maintain overall superiority, LLMs can exhibit superior capabilities in identifying rare disease entities in few-shot settings. Therefore, for specialized tasks focusing on rare diseases where relevant samples and sufficient computational resources are available, LLMs may be used to achieve higher precision. Nevertheless, within our current visual analytics framework centered on general-purpose unstructured medical text, MacBERT provides the most reliable performance-to-cost ratio.

For NER requiring deep semantic insight, language-specific pretrained models often outperform general multilingual models [50]. In this study, we use MacBERT—the top-performing model on the CBLUE dataset—to maximize extraction accuracy for Chinese corpora. Notably, the underlying pipeline remains highly adaptable; the base encoder can be seamlessly replaced (eg, substituting BioBERT [51] for MacBERT) to process English or other language datasets.

Figure 2 shows the pipeline of data processing. The process can be described as follows.

Building upon the constructed dataset, we developed MExplore, a multilevel visual analytics system designed to facilitate progressive exploration and cascading visual analysis (T2). The system adopts a 3-tiered metaphor—cosmic space, star map, and planet cross-section—which inspires the design of the MD space view (Figure 3A), the MP star map (Figure 3B), and the focused sectional view (Figure 3C). We use a hierarchical layout that systematically guides users from macrolevel document retrieval to microlevel detail exploration.

The analytical workflow follows a top-down trajectory (Figure 3), starting from the MD space view, where users filter and define a subset of MDs as the basis for all subsequent analysis based on cluster and spatial distributions. Once the analytical scope is established, the system transitions to the MP star map, enabling users to identify structural patterns and semantic clusters within MP subgraphs. From there, users can incorporate specific subgraphs into the association analysis view to reveal intricate relationships between MESs and MEs. For the most granular inquiry, users can select individual nodes within this view to generate a corresponding focused sectional view, facilitating a deep dive into the target knowledge unit’s context.

By structuring the analysis as an interactive, stepwise process, MExplore allows users to progressively increase analytical granularity without losing global contextual coherence [58]. Furthermore, the design leverages grounded metaphors to map complex information onto pre-existing cognitive schemas [59], effectively mitigating the visual complexity encountered by the user and enhancing memory encoding for knowledge retention [60]. Ultimately, this multitiered approach facilitates a continuous analytical loop.

As shown in Figure 4A, upon entering keywords, the system executes a full-text vector search to retrieve relevant MDs. The retrieved MDs are mapped into a 3D force-directed graph where each MD is represented as a planet—its size encodes text length, and its position is determined by semantic similarity. Interdocument links visualize this similarity. Users can adjust a threshold slider to filter connections by similarity, aggregating MDs into topic-based clusters represented as nebulae labeled with topic titles. This layout emphasizes core structural relationships and allows users to follow a progressive approach starting from these nebulae, clicking a specific nebula to enter Focus Mode to explore individual MDs in detail. This interaction mechanism leverages spatial proximity to facilitate the efficient identification of areas of interest [61].

The colors of the nebulae and their constituent planets represent their respective topics. This color encoding is discarded upon transitioning to detailed analysis views, preventing semantic interference between the MD topic and ME attributes.

To facilitate comprehensive discovery, a guide for unexplored nebulae is displayed in the lower-left corner, allowing for thorough exploration and rapid focusing by clicking. Simultaneously, the panel below ranks MDs by vector similarity scores to support instant focused viewing and prevent users from overlooking important MDs. Throughout this process, users iteratively select MDs to build an analytical collection, which serves as the foundation for all subsequent analysis views.

Relying solely on full-text embedding similarity can introduce bias, as document representations may disproportionately reflect longer sections while overshadowing concise but critical information. For instance, detailed medical histories in outpatient records can overshadow concise diagnosis sections during similarity computation.

To mitigate this, MDs are decomposed into MPs, each modeled as a graph vertex. Intradocument edges E_d connect the MPs within the same MD. Cross-document MP pairs (Figure 5A) exceeding the similarity threshold set in Figure 4A are linked by similarity edges E_s. The KaFFPa algorithm [62] then partitions the graph, segregating semantically distinct MPs into subgraphs (Figure 5B).

Each resulting subgraph is visualized as a constellation in the MP star map (Figure 4B), the MPs are depicted as stars, and E_d is rendered as a constellation line, whereas E_s is omitted to reduce clutter but is retained as latent factors influencing the layout, the similarity-based gravitational force F_similarity, and the structural gravitational force F_intra (Figure 4C). The combined gravitational force of the star i is calculated as follows:

where F_intra is the link force connecting the MPs within the same MD, with a magnitude of unit force ‖F‖, θ is the threshold set by the user, and similarityij is the similarity between MP i and MP j. d^ij is the unit vector of the force direction from MP i to MP j.

In addition, each star i is subjected to spring forces Fspringi from 9 ME-type poles uniformly distributed on the circular boundary of the star map:

where numij is the number of MEs of type j in MP i, and num_i is the total number of MEs in MP i. d^ij is the unit vector of the force direction from MP i to V_j.

To prevent occlusion, each star i is subject to the collision force Fcollisioni of stars that may overlap it. The final combined force Fcollisioni is calculated as follows:

This multiforce design achieved semantically coherent clustering (F_spring), representation of cross-document similarity and structural relationships (F_gravity), and clear visualization (F_collision) (Figure 4B). Stars within the same constellation are assigned a uniform color, allowing users to quickly identify them even on a densely populated map. This color is modulated by a sequential luminance gradient [63] to encode the star count: brighter constellations denote a greater number of constituent stars, enabling users to efficiently pinpoint pattern-dense regions. The luminance of star borders encodes the ME count, providing information density cues (T2). To preserve macrocontext awareness during the analysis of MPs and subsequent tasks, a topic navigation list at the bottom of the sidebar (Figure 4) visualizes the topic information of selected MDs. In the MP star map and subsequent views, users can click on MP/MES nodes or document cards to highlight the topic of the corresponding document. This mechanism enables on-demand backtracking, ensuring users remain oriented within the macrolevel topic context. This design facilitates progressive schema construction [33,34] and supports iterative exploratory learning.

Leveraging the concepts within MPs to create visual representations of hierarchical structures can significantly increase the efficiency of knowledge navigation [64,65]. Therefore, we extract MEs from each MP (T1), compose the MES, and visualize the resulting data via a radial dendrogram. Compared with alternative text visualization methods, such as word clouds or Sankey diagrams, this approach effectively captures both inclusion and relational connections among elements while accurately displaying hierarchical structures. To further enhance visualization, we propose an algorithm that groups MESs containing shared MEs into the same tree branch, thereby emphasizing co-occurrence relationships between the MESs (T3). The algorithm is outlined in Textbox 1.

For each MP, the extracted MEs form an MES, which acts as a node. The newly added node set N_add, along with the root node of the tree r_tree, is provided as input. For each node n_i in N_add, the algorithm retrieves its associated MEs via GetMEs(n_i) function. It then traverses the tree through TraverseTree to identify the set M_checked of co-occurring MEs, along with the node set N containing these MEs. They are all grouped into the set N_children, which are subsequently added as children to the appropriate parent nodes.

If multiple nodes intersect, their common ancestor is identified by CommonFatherNode, and the new nodes are added as children of this ancestor. The nodes are added directly to the root if no common ancestor is found. The process concludes by returning the updated tree structure.

Based on the algorithm described above, the association analysis view is constructed (Figure 4C). To manage the cognitive load and ensure visual consistency, the system uses a synchronized color encoding strategy and a progressive disclosure mechanism. Categorical colors are assigned to ME nodes based on their entity types, while for MES nodes, the color is derived from a weighted mixture of its constituent ME types. To maintain cross-view coherence, this encoding scheme is applied uniformly across both the association analysis and focused sectional views, supported by a persistent legend in the sidebar that ensures color-to-type mappings remain readily accessible throughout the analytical process.

The risk of information overload is further mitigated by initially displaying only the aggregated MES nodes, allowing users to perceive the high-level relational structure without being overwhelmed by individual entities. Users can then interactively expand specific MES nodes to reveal their constituent MEs as independent child nodes. This progressive disclosure strategy enables users to manage information density dynamically according to their analytical needs.

These design choices facilitate the perception of complex hierarchical and relational patterns, thereby supporting schema construction and clinical reasoning by externalizing the learner’s cognitive processes [66,67]. Furthermore, the system architecture supports the dynamic addition of new MPs, enabling real-time reconstruction of the association tree. This capability allows seamless integration of new clinical information into the existing cognitive framework of the user, promoting iterative, exploratory learning and long-term memory retrieval [68].

To enable granular analysis of a specific ME or MES, the focused sectional view (Figure 4D) uses a “planetary layer” metaphor to organize complex co-occurrence data. The central ring functions as a donut chart that encodes the proportional distribution of ME classifications. If the analytical focus is a single ME, this layer identifies its specific classification. Surrounding this core, the mantle layer uses a polar-axis-aligned area chart to visualize co-occurring MESs. In this arrangement, each radial axis represents a distinct MES, with the axis height being Nik∗ht, which represents the number of MEs belonging to class k in the ith MES. To emphasize semantic continuity across the dataset, areas representing the same ME classification are connected across adjacent axes. Compared with traditional area or stream charts, this radial layout effectively portrays the compositional richness and semantic associations of the target entity within a compact structural signature.

To bridge the gap between visualization and raw text, the system allows users to click a radial axis to query the underlying database and retrieve the corresponding MD, which is then displayed as a document card in the panel below. This design enables users to revert to the original evidence for detailed, word-by-word analysis. To further prevent highlighting overload in the text view, the system implements a selective filtering strategy. While the card highlights various extracted MEs, users can click specific semantic regions in the polar area chart to isolate and highlight only the ME category of interest. This interaction triggers the document card to automatically scroll to the relevant text segment, enabling an efficient and focused review of the original evidence without cognitive distraction.

The resulting structural signatures illustrate the contextual association patterns unique to each focused ME/MES (T4). Furthermore, the mantle’s radial patterns facilitate the interactive comparison of MESs with similar ME configurations. This comparative analysis fosters deeper cognitive engagement and more enduring conceptual impressions, supporting the learner’s ability to internalize complex clinical relationships [67].

To facilitate knowledge retention and structure the analytical process, we developed the provenance view (Figure 4E). This view enables users to capture high-fidelity snapshots from any analytical view on demand. These snapshots function as interactive nodes on a scalable, zoomable canvas, allowing users to freely organize layouts and establish logical links between visual findings. By externalizing exploration history into an explicit visual learning path, users can revisit and inspect detailed analytical states at any time, effectively transforming ephemeral interactions into organized knowledge assets.

To comprehensively evaluate the usability, effectiveness, and expertise acquisition capabilities of MExplore, we used a mixed methods evaluation strategy comprising three components: (1) case studies with domain experts to assess the practicality, (2) a comparative user study with medical students to quantify learning outcomes, and (3) semistructured interviews with experts to evaluate usability and future potential of the system.

As a web-based visual analysis system, MExplore was developed using the d3.js and Django frameworks. A Microsoft Windows platform with a 2.71 GHz Intel Core i5-7300 CPU and 8 GB of memory was used as the front-end page server. The evaluation experiments were performed via a Google Chrome web browser.

The case study can demonstrate feasibility and usability in performing real-world tasks [69]. Therefore, we conducted case studies with domain experts (E3 and E5) to evaluate MExplore within authentic analytical scenarios. Through extensive discussion, the experts identified three critical stages of medical expertise acquisition as core tasks: (1) the discovery of areas of interest, (2) the association analysis, and (3) the construction of illness scripts. The session commenced with a 20-minute tutorial on MExplore’s features, followed by 1 hour of autonomous exploration focused on the predefined tasks. To capture nuanced qualitative feedback, a think-aloud protocol was used, which encouraged participants to externalize their thought process, ask questions, or provide verbal comments throughout the session.

The case study begins with a real scenario with exposed bone in the facial area. A CT scan confirmed the presence of osteonecrosis. The experts entered the keywords bone, expose, CT, and osteonecrosis in MExplore. This query generated an MD space where, upon setting the similarity threshold to 0.5, a distinct structure comprising 3 nebulae (A1, A2, and A3) emerged (Figure 6A).

The color of nebula A1 corresponded to topic 2 (Gastrointestinal and Organ Disorders), whereas A2 and A3 aligned with topic 1 (Oral and Maxillofacial Surgery). Consistent with the patient’s clinical presentation in the maxillofacial region, the experts disregarded A1 to concentrate their analysis on A2 and A3. A detailed exploration of A3 revealed that it contained MDs describing symptoms highly congruent with the patient’s condition. Notably, these MDs demonstrated strong connectivity to nodes in topic 3 (Pharmacology and Immunology), which also appeared in the recommendation list, confirming their high relevance to the search query. Consequently, E3 selected these MDs, along with those in A3, and transitioned them to the MP star map. The maximum size of the subgraph was set to 10, which was determined during the experts’ free exploration and proved effective for segmenting common MDs.

In the MP star map (Figure 6B), E5 first examined subgraphs close to dis and sym to identify the specific conditions. Three subgraphs were selected to review the text. After these MDs were compared with the original MDs (Figure 6C), the experts concluded that the target disease was medication-related osteonecrosis of the jaw (MRONJ). E5 commented, “Compared with standard search engines, MExplore facilitates the discovery of trustworthy MDs. The structured visual representations of knowledge units enable users to identify the target disease through interactive reasoning rather than randomly searching through uncertain sources.”

Upon identifying the target disease, the experts conducted an association analysis to determine the key factors related to the illness script. As defined by Feltovich and Barrows [29], illness scripts consist of 3 components: EC, FT, and CQ. In this case, the disease’s name suggested a connection to the medication, leading the experts to select subgraphs that distribute bias toward dru (Figure 6B1). Additionally, the subgraphs adjacent to the target disease with a high total number of MEs were selected (Figure 6B2).

The selected subgraphs were organized into the association analysis view (Figure 7). The experts first analyzed the MES (Figure 7A), which is distinguished by a high proportion of dru MEs based on its color encoding. Upon clicking to expand this MES, it was found to contain the targeted therapy drugs, Anlotinib, which is closely associated with the liver cancer ME, indicating the patient had a history of malignancy and related drug treatment. The text in Figure 7D further corroborated this by containing lung cancer and targeted therapy, noting specifically that the patient was using zoledronic acid. Therefore, E3 concluded that cancer status and related medications were factors associated with EC and FT. Regarding CQ, E3 analyzed the MES characterized by high bod and sym proportions (Figure 7B and C), identifying bone exposure and exudate as primary CQ factors. E3 concluded, “Performing association analyses on real-case factors can be integrated into clinical reasoning training, enhancing learners’ clinical analysis skills.”

To construct a comprehensive illness script, the experts began by analyzing the factors presented in case 2. E5 was initially interested in the role of zoledronic acid as an EC for the development of MRONJ, so he focused on it and generated a focused sectional view (Figure 8A1). The relevant MES includes a mic ME potentially related to the FT. A review of the corresponding MD (Figure 8A2) revealed that this ME pertains to osteoclasts, detailing the drug’s pharmacological action of inhibiting osteoclast growth, which may impair bone repair. The MD further indicates that when bisphosphonates are used in cancer patients, concurrent dental procedures such as tooth extraction may trigger MRONJ (Figure 8A3).

Subsequently, E3 examined a medical history MES involving tooth extraction, such MESs typically feature a relatively high density of pro and dis MEs (Figure 8B1). By examining the MDs of associated chief complaints and other medical history MESs (Figure 8B2 and B3), multiple co-occurring pro MEs were found to mention osteomyelitis, which was also referenced in Figure 8A3, suggesting that osteomyelitis may have developed as a CQ of MRONJ, likely secondary to an underlying infection.

To obtain a more definitive CQ, E5 performed a focus analysis of MRONJ (Figure 8C1), reviewing the MDs corresponding to the MESs that have sym and equ MEs (Figure 8C2 and C3). The analysis revealed that bone exposure and pus discharge were prominent clinical symptoms, whereas necrotic bone and radiolucent areas observed on panoramic radiographs emerged as distinctive imaging features. These findings can be used as reliable criteria for clinical diagnosis.

This analysis enabled the experts to construct a comprehensive illness script for MRONJ as follows:

As E3 noted, “MExplore enables learners to construct illness scripts in a short period, facilitating rapid clinical reasoning and enhancing long-term retention of script knowledge.”

The experts all recommended MExplore for its user-friendly interface and interaction capabilities. The system’s hierarchical visualization, organized by data granularity and incremental exploration ability, significantly enhances user understanding and analysis. Although some of the experts (E5-E8) had no prior experience with visual analytics systems, they were able to easily comprehend and effectively use the system after receiving minimal training. E4 concurred that the ME-based exploration learning process in MExplore allows learners to focus on the core concepts of the target domain while creating an independent exploration path. E2 suggested incorporating features such as exploration playback and snapshot functionality to further increase the system’s usability. In summary, the experts highly praised the system’s visual design and interaction approach, asserting that it holds significant potential to improve the efficiency of knowledge acquisition for medical professionals.

As shown in Figure 10, the experts believe that most of the key analytical requirements have been well met. E9 noted that the design of the association analysis view and the focused sectional view enabled users to correlate and concentrate their analysis on key knowledge points, thereby identifying relevant and valuable information—an essential aspect of the knowledge acquisition process. E8 noted that compared with search-based learning methods or chatbots, which have become popular in recent years, this approach can exercise the learner’s thinking ability and improve knowledge retention and accumulation. E5 further highlighted, “In addition to assisting novices in acquiring expertise, MExplore can also support experts and researchers in disease studies by aiding in research, summarization, and the discovery of new features.”

Although MExplore uses a model pretrained on Chinese corpora, the framework is not strictly bound by linguistic characteristics. Transformer-based architectures have demonstrated robust concept understanding across diverse languages when fine-tuned on domain-specific data [76]. Consequently, the pipeline can support cross-lingual applications by incorporating domain-specific clinical corpora.

As experts indicated in the Analysis Requirements section, MExplore’s extraction and knowledge association capabilities benefit a broad spectrum of users, not limited to novices. It can further assist senior clinicians and researchers in identifying novel disease features and advancing clinical studies.

To ensure broad generalizability, our retrieval and filtering mechanisms primarily leverage the semantic embedding similarity of unstructured text. However, we recognize that practical medical education often requires precise constraints. Our framework allows users to incorporate a deterministic metadata filtering layer (eg, filtering by visit date, department, and document type) into the search module tailored to their specific datasets. This enables the framework to meet curriculum-specific requirements and adapt to diverse educational goals beyond unstructured content analysis.

The methodology can be extended beyond medicine to other domains requiring structured knowledge exploration, such as jurisprudence. By fine-tuning predictive models on annotated domain-specific texts, the MExplore framework can be repurposed to visualize entity relationships and facilitate knowledge acquisition in nonclinical fields.