Data Characteristics for This Category
Data for medical record quality control and registration preparation primarily originates from internal hospital systems: Electronic Medical Records (EMR), Hospital Information Systems (HIS), LIS/PACS, and adverse drug/device event reporting systems. This data is a mix of structured and unstructured formats. Structured data includes patient demographics, diagnostic codes (e.g., ICD-10), treatment codes, and lab/imaging results. This data typically resides in databases or CSV files. Unstructured data includes physician's handwritten progress notes, nursing notes, surgical records, imaging reports, and pathology reports, primarily in PDF, DOCX, or plain text formats. Data updates frequently, especially for inpatient records, which are continuously generated during a patient's stay. Document structures vary; progress notes are usually chronological, while lab and imaging reports follow fixed templates. Fields and units strictly adhere to medical standards, for example, blood pressure in mmHg, temperature in Celsius, and drug dosages in mg or IU.
Constraints on "Citation and Traceability" from These Characteristics
The high sensitivity and strict compliance requirements of medical record quality control data impose specific constraints on citation and traceability. Due to patient privacy and medical malpractice liability, every citation must precisely trace back to a specific field or paragraph in the original medical record, ensuring accuracy and verifiability. Unstructured text contains numerous medical terms, abbreviations, and colloquialisms, requiring high-precision text segmentation and semantic understanding to prevent RAG models from hallucinating or taking information out of context during citation. Real-time data updates necessitate regular incremental updates to the knowledge base to ensure citation timeliness. Furthermore, data format differences across systems make data cleaning and standardization crucial before citation, impacting retrieval effectiveness and traceability accuracy. Therefore, the citation mechanism must support unified management and fine-grained positioning of multi-source heterogeneous data.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale for Recommendation |
|---|---|---|
Chunk size (Chunk Size) | 300–500 characters | Medical record text paragraphs are typically short. Chunks that are too long may introduce irrelevant information, while chunks that are too short may lose context. |
Recall count (Recall Count) | 8–12 entries | Needs to cover multiple relevant medical records and diagnostic reports to ensure comprehensive information. |
Similarity threshold (Similarity Threshold) | 0.75–0.85 | Medical terminology requires high precision. A threshold that is too low may introduce inaccurate citations, while one that is too high may miss relevant information. |
Rerank result count (Reranked Return Count) | 5 entries | Ensures that the most relevant, high-quality citations are displayed first, reducing cognitive load. |
Knowledge Base Update Frequency | Once daily | Addresses daily updates in medical record data, ensuring citation timeliness. |
Citation Source Display Format | DocumentID:page number:paragraph range (File ID:Page Number:Paragraph Range) | Precisely traces back to the specific location in the original medical record file for manual verification. |
Common Pitfalls
- Citation results contain medical record text irrelevant to the query. This occurs when knowledge base chunk granularity is too large, causing the RAG model to retrieve irrelevant information along with relevant data.
- Some citation sources display "No reference in local knowledge base" or are empty. This happens when specific fields or unstructured text are not correctly indexed during data preprocessing, preventing traceability.
- The system reports
Knowledge base ID not found. This indicates an incorrect application variable configuration, failing to correctly point to the knowledge base ID, or improper knowledge base permission settings.
How to Verify Configuration
- Randomly select 10 registration application questions and verify that the citation sources accurately point to the specific location and content within the original medical record files.
- Check the knowledge base update logs to confirm that incremental update tasks execute as scheduled and do not report
parsing errororindexing failederrors. - Simulate inputting questions containing medical abbreviations and colloquialisms, then observe whether the citation results correctly parse and provide relevant evidence.
The values provided are common starting points and should be measured against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-21.