Data Characteristics
Dermatology pharmacovigilance data originates from national drug adverse reaction monitoring centers, regulatory bodies (e.g., FDA Adverse Event Reporting System, FAERS), pharmaceutical company spontaneous reporting systems, and medical literature. Update frequencies vary; regulatory databases typically release quarterly or annually, while internal company systems update in real-time. Document structures are diverse, including unstructured case reports, semi-structured adverse event report forms, and structured drug labels and clinical trial reports. Fields cover patient demographics, medication history, adverse reaction descriptions (signs, symptoms, severity, outcome), diagnostic information, relevant lab results, suspected drug information (lot number, dosage, administration), and concomitant medications. Adverse reaction descriptions often contain medical terminology and free text. Units include dosage (mg, g, IU), frequency (times/day, week), and time (days, months, years).
Constraints on Source Citation and Traceability
Diverse data sources and varying update frequencies in dermatology pharmacovigilance require a citation mechanism that integrates information with different timeliness. The presence of unstructured and semi-structured documents makes accurate key information extraction and traceable link establishment challenging. Free text and medical terminology in adverse reaction descriptions demand high accuracy in semantic understanding and information matching, directly impacting traceability granularity. For example, patient symptom descriptions like skin erythema and papules require accurate association with specific drug adverse reaction classifications. Furthermore, standardizing and parsing units like dosage and frequency are critical for accurate citation content. These constraints necessitate focusing on text preprocessing, entity recognition, and relationship extraction during knowledge base construction, along with providing detailed metadata for citation sources to ensure traceability to original reports or literature.
Configuration Settings
| Configuration Item | Suggested Value | Rationale |
|---|---|---|
Chunk size | 500-800 characters | Balances the completeness of adverse reaction descriptions in dermatology case reports with model context window limitations. |
Recall count | Top 10-15 entries | Diagnosis and differential diagnosis of dermatological adverse reactions often involve multiple similar symptoms. Increasing recall count improves relevant information coverage. |
Similarity threshold | 0.75-0.85 | Balances recall precision and recall rate. Avoids missed recalls due to medical terminology differences while reducing the introduction of irrelevant content. |
Rerank result count | Top 5 entries | Further optimization through a reranking model after initial recall. Focuses on the most relevant key information, reducing model processing load. |
MAX_TOKEN | 2048-4096 | Accommodates the length of detailed descriptions and multi-dimensional information in dermatology adverse reaction reports, ensuring context completeness. |
temperature | 0.3-0.5 | The pharmacovigilance domain demands high accuracy and factuality. Lower temperature values help generate more stable and reliable citation content. |
Common Mistakes
- AI model-returned citation information lacks a knowledge base ID or original document link, preventing traceability to the specific source. This occurs if external API calls do not correctly parse the
detail: trueparameter, or if detailed citation return is not enabled in the knowledge base configuration. - The system fails to accurately associate symptom descriptions like "skin itching with erythema" with corresponding drug adverse reaction entries. This typically results from overly coarse knowledge base chunking or insufficient semantic understanding of medical terminology by the text vectorization model.
- Errors in handling drug dosage units, such as misidentifying "twice daily, 5mg each time" as "five times daily, 2mg each time," lead to citation data discrepancies. This stems from a lack of standardized parsing for units and values during data preprocessing, or inaccurate regular expression matching.
Verification Steps
- Randomly select 10-20 typical dermatology drug adverse reaction queries. Check if cited knowledge base segments in each answer point to original reports or literature, and verify the consistency of the cited content with original sources.
- For queries containing specific drug dosages and frequencies, verify the accuracy of relevant numerical values and units in the AI-returned citation content against original data.
- Test dermatology adverse reaction symptom descriptions of varying complexity. Observe if the AI accurately identifies and recalls relevant medical terminology, diagnostic criteria, or differential diagnosis information, and assess its coverage within the knowledge base.
- Regularly review model-generated citation traceability links to ensure they are valid and accurately navigate to specific locations or relevant sections of the original data source.
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.