Data Characteristics in this Category
Intelligent triage in pharmacovigilance primarily uses data from drug inserts, clinical trial reports, real-world study data, adverse event reporting systems (e.g., national adverse drug reaction monitoring center databases), and medical literature. Data update frequencies vary. Drug inserts typically update with approval or revision cycles. Adverse event reporting systems receive continuous data. Document structures are diverse. Drug inserts are semi-structured text, containing fixed fields like indications, dosage, contraindications, and adverse reactions. Adverse event reports are often unstructured text, including patient complaints, medication history, and adverse event descriptions. Fields involve generic drug names, brand names, batch numbers, manufacturers, adverse event types, severity, causality assessment, and reporting sources. Units include dosage (mg, g), frequency (times/day), and duration (days, weeks).
Constraints Imposed by These Characteristics on "Model Integration and Configuration"
Data characteristics in pharmacovigilance intelligent triage impose specific constraints on model integration and configuration. The semi-structured nature of drug inserts requires the model to effectively parse tables and lists and extract key adverse reaction information. Unstructured adverse event reports demand higher Natural Language Understanding (NLU) capabilities, requiring processing of colloquialisms, abbreviations, and medical terminology. Inconsistent data update frequencies mean knowledge base synchronization strategies must be flexible. High-frequency data sources (like adverse event reports) should support incremental updates, while low-frequency ones (like drug inserts) can use periodic full updates. Diverse fields and units require the model to perform entity recognition and relation extraction, accurately identifying entities like drugs, symptoms, and dosages, and understanding their relationships. Additionally, for sensitive medical data, data security and privacy protection are critical considerations during model integration, ensuring compliance in data transmission and storage.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Chunk Size) | 500–800 characters | Ensures each text block contains sufficient context for adverse event descriptions and drug information correlation, preventing semantic loss from long text splitting. It also controls single chunk length to reduce model processing load. |
Recall count (Retrieval Count) | 10–15 entries | Balances recall rate and model inference efficiency. Intelligent triage scenarios may require correlating multiple drugs or symptoms. Increasing the retrieval count can improve coverage of relevant information and reduce the risk of missed diagnoses. |
Similarity threshold (Similarity Threshold) | 0.75–0.85 | Ensures retrieved results are highly relevant to the user's query, filtering out low-quality or irrelevant knowledge snippets. This threshold requires fine-tuning based on actual data distribution and model performance; too low may introduce noise, too high may lead to missed retrievals. |
Rerank result count (Reranked Return Count) | 5 entries | After initially retrieving many potentially relevant results, a reranking model refines them, providing only the most relevant few pieces of information to the final inference model. This improves the accuracy and efficiency of the final answer. |
maxContext | 4000 characters | Considers the input window limitations of current mainstream large language models. This parameter setting should accommodate the user's query, retrieved knowledge snippets, and necessary system instructions, preventing truncation or performance degradation due to excessive context length. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Drug inserts and adverse event reports may contain complex structures or large amounts of content. A longer parsing timeout prevents parsing failures due to excessively long file processing times. |
Three Common Mistakes
- Knowledge base document parsing failure or partial content loss. This occurs when document format diversity is not fully considered. For example, image text in PDF drug inserts is not OCR processed, or table structures are not correctly extracted, preventing critical information from being imported into the knowledge base.
- Model inference answers are brief or fail to fully follow instructions. This usually happens when the
maxContextparameter is set too small. When the user's query and retrieved knowledge snippets are combined, the total length exceeds the model's processing limit, forcing the model to answer based on partial information. - Intelligent triage results have poor relevance, failing to accurately match drugs with adverse reactions. This may be due to an inappropriate chunking strategy during knowledge base vectorization, where the chunk size is too large or too small. This can break the complete semantic meaning of adverse event information or confuse it with other irrelevant information, affecting similarity calculations.
How to Verify Correct Configuration
- Perform batch document uploads and parsing for various drug inserts and adverse event reports. Check if the knowledge base completely retains key field information, such as adverse event descriptions and drug dosages, and ensure no parsing error logs are output.
- Simulate typical triage scenarios (e.g., "What should I do if I experience [symptom] after taking [drug]?"). Observe if the model's output accurately references relevant drug information and adverse reaction treatment advice from the knowledge base. Compare with expected results to confirm the relevance threshold is set appropriately.
- Randomly select several documents already imported into the knowledge base. Use key phrases for retrieval. Check if the
Recall count(Retrieval Count) andRerank result count(Reranked Return Count) meet expectations. Verify that theSimilarity threshold(Similarity Threshold) of the retrieved results falls within a reasonable range, ensuring highly relevant knowledge snippets are effectively recalled.
The values provided are common starting points and should be measured against your 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.