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

Our research focuses on developing an intelligent acupuncture system for diminished ovarian reserve diagnosis and treatment. This protocol can be applied to acupuncture research involving diminished ovarian reserve and other geologic reproductive disease. To begin, collect data from classical acupuncture texts, Chinese language literature from the China National Knowledge Infrastructure, or CNKI, clinical data from reproductive centers at tertiary hospitals, English language literature from the Web of Science core collection, and standardized knowledge bases.

Perform data annotation and select data according to the inclusion and exclusion criteria. Include 400 clinical cases to support the identification of diagnostic and therapeutic patterns. Using a standardized annotation platform, manually annotate approximately 30%of the data, including classical texts and semantically complex content.

With large language models, or LLMS, process 70%of the data for entity recognition, relationship extraction, and terminology normalization. Review the generated results. Using structured prompt templates designed to guide LLM, perform automated extraction.

Once the model role is defined as a domain specific knowledge extraction expert, include a few shot learning example in JavaScript object notation or JSON triplet format to improve extraction performance. Using JSON validation tools, serialize validated knowledge triplets into independent JSON files named according to their source documents to maintain traceability. Store the finalized knowledge graph in a graph database using Python-based frameworks.

Visualize the knowledge graph using built-in graph database tools to support inspection and validation. Develop the application system based on the constructed knowledge graph to enable clinical decision support and process user input data accordingly. Using the bidirectional long short-term memory conditional random field or bi-LSTM CRF model, implement natural language understanding for named entity recognition and entity linking.

Build the model architecture with 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. Train the model on manually annotated clinical and literature data with annotations performed independently by domain experts. Using Cohen's Kappa Coefficient, assess inter-annotator agreement with a threshold of at least 0.85 and resolve discrepancies through expert consensus.

During inference, extract structured entities including symptoms, diseases, and syndrome patterns from clinical input data. Map these entities to standardized nodes in the knowledge graph. Using a dual mode strategy, perform knowledge retrieval, combining keyword-based and semantic-based approaches.

Use vector representations to identify relevant unstructured knowledge. Merge keyword-based and semantic-based retrieved results. Identify and remove duplicates and structure results for downstream processing.

Adjust weights for gynecological and diminished ovarian reserve related contexts. Using natural language generation techniques, generate diagnostic and treatment outputs and convert structured knowledge into clinically interpretable responses. After applying the developed system in a clinical setting, recruit participants diagnosed with diminished ovarian reserve according to the inclusion and exclusion criteria.

Collect and record patient demographic and clinical information in a structured format. Generate diagnostic outputs and acupuncture prescriptions based on the input data. Once the system performance is evaluated using diagnostic accuracy and treatment appropriateness, compare system generated syndrome differentiation results with results provided by three senior gynecological traditional Chinese medicine physicians serving as the reference standard.

Calculate the diagnostic agreement rate as the proportion of consistent cases among all cases. Evaluate treatment appropriateness using a panel of three senior acupuncturists with at least five years of clinical experience and senior professional titles and have experts independently assess system generated prescriptions using a five-point Likert scale. Evaluate interrater agreement using Kendall's coefficient of concordance.

A total of 374 Chinese language sources, 267 English language articles, and 82 classical traditional Chinese medicine texts were incorporated into the knowledge graph. The full knowledge graph contained 16, 558 entities across seven predefined categories and 80, 084 relations spanning 21 relation types. Among 90 patients, age ranged from 22 to 54 years with a peak at 32 years and the highest concentration in the 31 to 35-year age group, self-rating anxiety scale assessment showed that 57%of patients had normal scores.

28%had subclinical anxiety levels and smaller proportions had mild, moderate, or severe anxiety. Diagnostic agreement between system generated outputs and expert clinician assessments was 86.7%Treatment appropriateness had a mean score of 4.65 on a five point Likert scale. Interrater agreement among expert evaluators measured using Kendall's coefficient of concordance was approximately 0.13, suggesting a statistically significant level of agreement, although the overall concordance remained limited.

This approach enables researchers to access the diagnostic agreement and the acupuncture prescription appropriateness. Researchers can additionally perform large language mode based reasoning analysis, multicenter validation, and the long-term eventual follow-up using this procedure. Future studies can build on this work by expanding applications to additional disease integrating multimodal data and conducting prospective studies.