Evidence-Based Intelligent Diagnostic and Therapeutic Visualization System with Large Language Models: Multi-Turn Interaction and Multimodal Treatment Plan Generation

Abstract: Aim: Existing AI-assisted traditional Chinese medicine diagnostic tools suffer from opaque reasoning processes, passive interaction, and limited treatment plan presentation. This study proposes a knowledge-enhanced visual diagnostic system to improve the transparency and interpretability of syndrome differentiation and treatment. Methods: The system is built upon a Neo4j knowledge graph comprising 241 syndromes, 1,263 symptoms, and 2,485 relations. It incorporates a four-stage symptom matching pipeline (exact, semantic, fuzzy, and LLM verification), an information gain-driven proactive questioning strategy optimized with genetic algorithms, and a multimodal treatment presentation integrating artificial intelligence-generated illustrations, three-dimensional meridian-acupoint models, and evidence-based literature. Results: Knowledge graph constraints reduced non-standard outputs by 32%. Case studies validated the effectiveness of the interactive workflow across patient self-assessment, clinician-assisted diagnosis, and traditional Chinese medicine education. Automated paired-comparison evaluation across 30 cases further demonstrated significant improvements in diagnostic trust (Cohen's d = 1.82, p < 0.001), reduced cognitive load (improvements in four of five dimensions), and higher credibility of evidence-based references (4.21 vs. 2.95). Conclusions: The proposed system enhances the transparency of traditional Chinese medicine diagnostic reasoning and the interpretability of treatment plans through knowledge graph-driven visualization and multimodal interaction, offering a practical solution for trustworthy artificial intelligence-assisted traditional Chinese medicine applications.

• The paper presents a novel TCM diagnostic system that integrates knowledge graphs with LLM-based multi-turn interaction to drive evidence-based reasoning.
• It employs a four-stage entity normalization and information-theoretic symptom selection process to enhance diagnostic accuracy and transparency.
• The multimodal output—combining text, AI illustrations, and interactive 3D acupoint models—substantially reduces cognitive load and improves treatment adherence.

Evidence-Based Intelligent Diagnostic and Therapeutic Visualization System with LLMs for Multiturn, Multimodal TCM Decision Support

System Overview and Motivation

This work introduces an AI-assisted diagnostic and therapeutic visualization platform for Traditional Chinese Medicine (TCM) built upon a robust integration of LLMs and structured medical knowledge graphs (KGs). The system directly addresses critical bottlenecks in existing digital TCM assistants: the lack of multi-turn consultation mirroring real clinical reasoning, limited transparency into model decisions, and non-intuitive presentation of treatment advice. To this end, the authors propose a pipeline where the KG is not merely a post-hoc validator or retrieval source, but forms the backbone of dynamic, evidence-constrained diagnostic reasoning and progressive visual explanation throughout the consultation. The approach is further augmented by advanced multimodal interfaces for conveying complex treatment regimens.

Knowledge Graph as Structural Constraint and Reasoning Scaffold

A central innovation is the elevation of TCM knowledge graphs from ancillary retrieval mechanisms to the primary driver of the diagnostic interview. Unlike existing approaches which apply KGs as external context or for retrieval-augmented generation (RAG), this system utilizes the KG to iteratively restrict the candidate diagnostic hypothesis space and to structure system-initiated, information-theoretically guided questioning. Specifically, the system leverages a four-stage entity normalization pipeline for mapping free-text symptom input to standardized KG entities (exact match, semantic embedding match with BGE-M3, fuzzy string match, and LLM-based verification), ensuring high recall while minimizing mapping error.

At each conversational turn, the active candidate syndrome set is updated based on observed evidence, and the next symptoms to probe are selected based on maximal expected information gain, penalizing redundancy and encouraging discriminative questioning. This significantly enhances diagnostic efficiency and supports systematic uncertainty reduction, grounded in the underlying KG topology.

Progressive, Transparent Reasoning via Knowledge Graph Visualization

The diagnostic workflow is made transparent through dynamic subgraph visualization tightly coupled to reasoning state. At each turn, case-specific subgraphs are rendered in real time using ECharts, with categorical color mapping distinguishing symptoms from syndrome candidates and dynamic highlighting. The subgraph evolves as interactions proceed, visually exposing the evidence path, syndrome-symptom relationships, and the iterative narrowing of the hypothesis space. This design makes the diagnostic inference path both visually auditible and cognitively accessible, targeting both patient trust and clinical workflow traceability.

Advanced Multimodal Treatment Information Presentation

A persistent limitation of digital TCM assistants has been the reduction of treatment plan delivery to plain text, which fails to communicate spatial or complex procedural information. This system responds by integrating AI-generated synoptic illustrations and three-dimensional (3D) interactive acupoint models into the treatment plan presentation. During the treatment advice phase, users are provided with a complete package: structured textual instructions, dietary and lifestyle advice, AI-generated images of prescribed herbs, and spatially annotated 3D acupoint visualization (with highlighting and rotation/zoom interaction). This multimodal fusion aims to enhance comprehension and practical adherence, especially for non-expert users and populations with health literacy barriers.

LLM Hallucination Mitigation and Robust Knowledge Grounding

The system adopts a dual-pronged hallucination management strategy: (1) knowledge graph-constrained reasoning in early diagnostic stages to strictly limit candidate output within the boundaries of explicit KG structure, and (2) post-hoc reference augmentation for generated therapeutic advice. The KG acts as a tractable evidence frontier, constraining both question formulation and answer space, thus substantially reducing the possibility of unsupported or fabricated content—a key failure mode of open-ended LLMs in clinical scenarios. When generating treatment recommendations, explicit references and web page content are provided to enable downstream verifiability and user confidence.

Experimental Methodology and Numerical Findings

Evaluation is based on a stratified set of 30 authentic TCM electronic medical record cases, spanning a diverse syndrome spectrum. Three axes of performance were quantitatively assessed via LLM-as-judge protocols: diagnostic trust/transparency under KG visualization, cognitive load reduction via multimodal plan delivery, and the impact of appended references on perceived evidence quality. Scores on each metric were measured via structured Likert scales (1–5) across five dimensions per axis, using the Claude Sonnet 4.5 model as independent evaluator with deterministic output settings.

Key quantitative results:

• Diagnostic transparency and user trust significantly improved under interactive KG visualization compared to baseline (Wilcoxon signed-rank test, Benjamini–Hochberg corrected $p < 0.05$ across all trust dimensions), with large effect sizes (median Likert score gains ≥1 point in transparency and traceability).
• Cognitive load for non-expert patients interpreting treatment plans was substantially reduced with structured multimodal output (AI illustrations, 3D models), showing increased comprehensibility, execution confidence, and ease of acupoint localization (effect sizes measured by rank-biserial $r$ > 0.6), with all paired dimensions statistically significant.
• Reference augmentation for treatment advice enhanced perceptions of evidence quality, verifiability, and source credibility—again, effect sizes were moderate to large and statistically robust after multiple comparison correction.

Implications and Comparative Analysis

Relative to contemporary TCM-AI systems such as KNOWNET, BianQue, and knowledge-tuned LLM models, this system is distinct for its:

1. KG-driven, information-theoretically optimized consultation structure.
2. Continuous reasoning path visualization embedded within the diagnostic process, not merely for post-hoc browsing.
3. Deep multimodal integration in both diagnosis and therapy delivery.

The authors note that these choices are domain-targeted optimizations, not universally dominant—e.g., progressive KG-guided consultation introduces graphical interface complexity and may require further simplification for certain user cohorts. Nevertheless, the ablation and judge-based evidence in this study argue for broad gains in transparency, cognitive accessibility, and user confidence/trust relevant to both clinical integration and patient-oriented AI.

Practical and Research Outlook

Practically, this approach suggests a template for next-generation AI healthcare assistants: use KGs not as static knowledge stores but as dynamic scaffolds structuring LLM inference, inference validation, and user experience. Progressive graph-driven interaction and multimodal output can meaningfully close the gap between opaque model output and clinically actionable, explainable reasoning. Theoretically, future work may further optimize KG-constrained LLMs through tight interface between symbolic and neural pipelines or investigate adaptive visualization methods that scale to larger or more granular TCM knowledge bases. Hallucination detection could be enhanced by adversarial testing within KG-defined boundaries and by further integration of external evidence sources at every interaction stage.

Conclusion

This paper presents a rigorous, systematic design for an evidence-based, multimodal, LLM-driven TCM diagnostic and therapeutic platform. By synergistically integrating KG structural constraints, progressive visualization, and rich multimodal presentation, the system delivers quantifiable improvements in transparency, cognitive accessibility, and evidence support. These architectural principles are applicable to broader medical AI domains where trustworthy, explainable, and user-adapted decision support is essential.