A Study on an Intelligent Diagnosis and Treatment Assistant System for Acupuncture in Diminished Ovarian Reserve Based on a Knowledge Graph

Summary

A gynecology-specific acupuncture knowledge graph and decision-support system for diminished ovarian reserve (DOR) was developed using literature, classical texts, and clinical data. Clinical evaluation demonstrated 86.7% diagnostic agreement and high prescription scores, with statistically significant but limited inter-expert agreement, supporting its role as an auxiliary clinical decision-support tool.

Abstract

Diminished ovarian reserve (DOR) is increasing in incidence and affecting younger women. Acupuncture has been used clinically for DOR; however, practice remains subjective, experience-dependent, and lacks standardized guideline implementation. Building on our previous general intelligent acupuncture system, this study targets gynecological diseases and uses DOR as the clinical entry point to develop a specialized knowledge graph and decision-support system. This study included 82 classical acupuncture texts, Chinese and English literature, and clinical data. Knowledge was extracted using a hybrid approach combining large language models and manual annotation to construct a full acupuncture knowledge graph and a DOR-specific gynecology subgraph. The system employs BiLSTM-CRF-based entity recognition and dual-mode retrieval to automate syndrome differentiation and acupuncture prescription generation. The full graph contains 16,558 entities and 80,084 relations, while the gynecology subgraph contains 8,677 entities and 27,092 relations. In a clinical evaluation of 90 patients with DOR, the system achieved a diagnostic agreement rate of 86.7% compared with expert assessments. The mean prescription appropriateness score was 4.65. Inter-rater agreement among experts, measured using Kendall’s coefficient of concordance, was W ≈ 0.13 (P < 0.05), indicating statistically significant but limited agreement. These findings suggest that the system can provide structured support for acupuncture-based syndrome differentiation and prescription generation in DOR, with potential as an auxiliary clinical decision-support tool.

Introduction

Diminished ovarian reserve (DOR) is characterized by a decline in the quantity and/or quality of oocytes, leading to reduced fertility, menstrual irregularities, and hormonal imbalance¹. The incidence of DOR has increased in recent years and is progressively affecting younger populations. Epidemiological data from the United States indicate an increase from 19% in 2004 to 26% in 2011², while a report from the Chinese Medical Association documented a prevalence of 13.58% among infertile patients undergoing assisted reproductive technology^2,3.

Current management strategies in Western medicine include controlled ovarian hyperstimulation, adjuvant pharmacological treatments, and hormone replacement therapy. However, these approaches are associated with several limitations, including low oocyte retrieval rates, limited improvement in live birth outcomes, high treatment discontinuation rates, contraindications, and potential adverse effects^4,5.

Acupuncture has been investigated as a complementary therapeutic approach for improving ovarian function in patients with DOR. Evidence from clinical studies and systematic reviews suggests potential benefits^6,7,8. Nevertheless, acupuncture diagnosis and treatment remain highly dependent on practitioner experience, vary across different theoretical schools, and lack standardized integration with modern clinical data. These limitations restrict the broader implementation of acupuncture guidelines and increase the learning barrier for clinicians without formal training in traditional Chinese medicine.

Recent advances in artificial intelligence (AI) and knowledge graph (KG) technologies have enabled new approaches for structuring and utilizing complex medical knowledge in acupuncture and TCM^9,10. Prior studies include visual analyses of acupuncture research trends¹¹, acupoint recommendation models based on graph convolutional networks¹², knowledge graph frameworks incorporating prompt learning for acupuncture and tuina¹³, and systematic reviews of TCM knowledge graph construction and applications¹⁴. Furthermore, the integration of knowledge graphs with large language models (LLMs) has emerged as a promising direction, as demonstrated by systems such as TCM-KLLaMA for prescription generation and case-based question-answering systems combining LLMs with knowledge graphs^15,16. These studies highlight the potential of AI-driven approaches to support structured reasoning and decision-making in TCM. In addition, complex network methodologies have been applied to analyze acupuncture theoretical systems, providing methodological support for modeling structured medical knowledge¹⁷.

Our group has previously developed a general acupuncture knowledge graph for infertility, including the Transformer-based dual-channel ACUBERT model¹⁸ and a large-scale acupuncture knowledge base with multi-hop reasoning capability¹⁹. However, these systems were designed as general-purpose frameworks and lack disease-specific optimization for gynecological conditions. In addition, they have not been clinically validated in a DOR-focused application scenario.

In this context, the present study focuses on gynecological reproductive disorders, using DOR as the clinical entry point. Building upon the existing general acupuncture knowledge graph, a domain-specific knowledge graph tailored to DOR was constructed, and an intelligent clinical decision-support system was developed. The system was evaluated using real-world clinical cases to assess diagnostic consistency and prescription appropriateness, aiming to provide structured and reproducible support for acupuncture-based clinical decision-making.

Protocol

This study was conducted in accordance with institutional ethical standards and was approved by the Ethics Committee of the General Hospital of Eastern Theater Command (Approval No. DZQH-KYLLFS-25-38). Written informed consent was obtained from all participants prior to inclusion. All patient data were anonymized and de-identified to ensure privacy protection. The research tools for this protocol have been listed in the Table of Materials.

1. Knowledge graph construction

The knowledge graph was constructed using a multi-source data integration strategy. Data collection and annotation were performed by trained personnel with formal education in Traditional Chinese Medicine (TCM) and acupuncture. Data sources included classical acupuncture texts, Chinese-language literature from CNKI, English-language literature from the Web of Science Core Collection, clinical data from reproductive centers at tertiary hospitals, and standardized knowledge bases (The detailed list of classical texts is available in Supplementary File 1). These sources were selected to ensure both historical depth and contemporary clinical relevance. A total of 400 clinical cases were included to support the identification of diagnostic and therapeutic patterns. The overall workflow of knowledge graph construction is illustrated in Figure 1.

Figure 1: Knowledge graph construction workflow. Overview of the data collection, extraction, and integration process used to construct the knowledge graph. Please click here to view a larger version of this figure.

Inclusion criteria required that all data be directly related to acupuncture diagnosis and treatment, with complete and clearly structured content suitable for entity and relationship extraction. Classical texts were required to be critically collated and free of obvious errors, while modern literature was limited to peer-reviewed publications or documents issued by authoritative institutions. Clinical data were required to be de-identified and ethically approved. Exclusion criteria included duplicate records, incomplete text fragments, content unrelated to acupuncture diagnosis and treatment, and unsupported or erroneous statements.

A hybrid knowledge extraction approach was implemented, combining manual annotation and automated extraction. Approximately 30% of the data, including classical texts and semantically complex content, was manually annotated by domain experts using a standardized annotation platform. The remaining 70% of the data were processed using large language models (LLMs) for automated extraction, including entity recognition, relationship extraction, and terminology normalization. All automatically generated results were reviewed and corrected by experts to ensure accuracy and consistency.

Automated extraction was performed using structured prompt templates designed to guide LLM behavior (Template examples are presented in Supplementary File 2). These templates defined the model role as a domain-specific knowledge extraction expert and incorporated few-shot learning examples in JSON triplet format to improve extraction performance. Preprocessed literature abstracts were submitted through application programming interfaces (APIs), and parallel processing strategies were implemented to improve computational efficiency. Extraction outputs were parsed and validated using JSON validation tools to ensure structural integrity. Validated knowledge triplets were serialized into independent JSON files, named according to their source documents to maintain traceability.

The finalized knowledge graph was stored in a graph database and implemented using Python-based frameworks. Visualization was performed using built-in graph database tools to support inspection and validation.

Entities in the knowledge graph were categorized into seven types: diseases, symptoms, syndrome patterns, acupoints, meridians, treatment methods, and effector substances and biological mechanisms. Relationships between entities were constructed to reflect clinical diagnostic and therapeutic logic, supporting multi-hop reasoning and structured inference. The ontology and relationship structure are illustrated in Figure 2. The detailed knowledge graph schema and examples of relationship constraints are provided in Supplementary File 3 and Supplementary File 4.

Figure 2: Knowledge graph ontology and relationship model. Schematic representation of entity types and relationships within the knowledge graph. Core entities include disease, symptom, syndrome, acupoint, and treatment method, and they illustrate diagnostic and therapeutic relationships.
Please click here to view a larger version of this figure.

2. Application system development

The application system was developed based on the constructed knowledge graph to enable clinical decision support. The system processes user input data and performs structured analysis through natural language processing and knowledge graph reasoning.

Natural language understanding was implemented using the BiLSTM-CRF model for named entity recognition and entity linking; the model architecture is shown in Figure 3. The model architecture consisted of an embedding layer trained on domain-specific corpora, a bidirectional long short-term memory layer to capture contextual information, and a conditional random field layer to ensure optimal sequence labeling. The model was trained on manually annotated clinical and literature data, with annotation performed independently by domain experts. Inter-annotator agreement was assessed using Cohen’s kappa coefficient, with a threshold of ≥0.85. Discrepancies were resolved through expert consensus.

Figure 3: BiLSTM-CRF model architecture. Schematic diagram of the BiLSTM-CRF model used for named entity recognition, including embedding, BiLSTM, and CRF layers.
Please click here to view a larger version of this figure.

During inference, the model extracted structured entities, including symptoms, diseases, and syndrome patterns, from clinical input data. These entities were mapped to standardized nodes in the knowledge graph using terminology normalization, cosine similarity-based vector matching, and contextual disambiguation.

Knowledge retrieval was performed using a dual-mode strategy that combined keyword and semantic-based approaches. Keyword retrieval utilized full-text indexing for precise matching of structured entities, while semantic retrieval employed vector representations to identify relevant unstructured knowledge. Retrieved results were merged, deduplicated, and structured for downstream processing. A multi-dimensional ranking strategy was applied to prioritize results based on exact matching, semantic similarity, and keyword relevance, with weighting adjusted for gynecological and DOR-related contexts.

The system generated diagnostic and treatment outputs using natural language generation techniques, converting structured knowledge into clinically interpretable responses.

3. Clinical application and evaluation

The developed system was applied in a clinical setting to evaluate its practical performance. Study participants were patients diagnosed with diminished ovarian reserve at a tertiary hospital in Nanjing who consented to receive acupuncture treatment.

Participants were eligible if they were aged 20 to 40 years and met diagnostic criteria for diminished ovarian reserve according to the 2016 POSEIDON criteria and the 2020 Chinese Expert Consensus. Diagnostic indicators included reduced anti-Müllerian hormone levels and antral follicle count < 5. Patients were excluded if they had conditions affecting fertility, serious comorbidities, or had received acupuncture treatment for fertility within the previous three months. The patient inclusion and exclusion flowchart is shown in Figure 4.

Figure 4: Patient inclusion and exclusion flowchart. Flowchart showing patient selection. A total of 127 patients met inclusion criteria, and 90 patients were included after applying exclusion criteria and accounting for incomplete data. Please click here to view a larger version of this figure.

Patient data, including demographic and clinical information, were collected and recorded in a structured format. Participants completed standardized questionnaires to capture symptom and clinical information (the detailed questionnaire is provided in Supplementary File 5). The system generated diagnostic outputs and acupuncture prescriptions based on the input data.

System performance was evaluated using diagnostic accuracy and treatment appropriateness. Diagnostic accuracy was assessed by comparing system-generated syndrome differentiation results with those provided by three senior gynecological TCM physicians, who served as the reference standard. The diagnostic agreement rate was calculated as the proportion of consistent cases among all cases.

Treatment appropriateness was evaluated by a panel of three senior acupuncturists with at least five years of clinical experience and senior professional titles. Experts independently assessed system-generated prescriptions using a 5-point Likert scale. Inter-rater agreement was evaluated using Kendall’s coefficient of concordance (W).

Results

The constructed knowledge graph integrates heterogeneous data from classical traditional Chinese medicine texts, modern clinical literature, and real-world case records, following established principles of knowledge graph construction and evaluation. A total of 374 Chinese-language sources, 267 English-language articles, and 82 classical TCM texts were incorporated, ensuring both historical depth and contemporary clinical relevance (Figure 5).

Discussion

The present study describes a knowledge graph-based intelligent decision-support system tailored for acupuncture treatment of diminished ovarian reserve, representing a transition from general-purpose traditional Chinese medicine knowledge modeling to a disease-specific clinical application.

Compared with conventional acupuncture practice, which relies heavily on practitioner experience, the proposed system introduces a structured and reproducible diagnostic framework. By integrating multi-sou...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
BiLSTM-CRF	Open Source (https://github.com/jmhIcoding/bilstm-crf)	N/A	Deep learning model architecture employed for named entity recognition in clinical text extraction.
CNKI Database	China National Knowledge Infrastructure	N/A	Literature repository utilized for the retrieval of relevant Chinese-language academic publications.
JSON Viewer & Formatter	Megaish	1.6.0	Software tool used for JSON data structure viewing and validation.
KGs Annotation Platform	Nanjing KGI Data Technology Co., Ltd.	N/A	Web-based collaborative annotation platform utilized for the manual labeling of medical entity and relationship corpora to support knowledge graph construction.
Large Language Model API	Moonshot AI	N/A	Kimi LLM API employed for semantic understanding and automated knowledge extraction from biomedical literature; interface compatible with OpenAI API format.
MySQL	Oracle Corporation	8.0.42	Relational database management system used for the storage of structured clinical data.
Neo4j Browser	Neo4j,Inc.	5.26.10	Graph database interface used for the interactive visualization and exploratory querying of the gynecological knowledge graph.
Python	Python Software Foundation	3.11.4	Primary programming language utilized for automated literature crawling, LLM API invocation, multiprocessing coordination, data parsing, and knowledge graph construction.
Web of Science Core Collection	Clarivate Analytics	N/A	Literature repository utilized for the retrieval of relevant English-language academic publications.