Data Characteristics in Pharmacovigilance
Pharmacovigilance (PV) registration dossiers primarily include Individual Case Safety Reports (ICSRs), Periodic Safety Update Reports (PSURs), and Risk Management Plans (RMPs). These documents are typically in PDF, Word, or structured data formats (e.g., E2B XML files). Data sources are diverse, encompassing clinical trials, post-market surveillance, literature reviews, and spontaneous reports from healthcare institutions. Update frequency is high; ICSRs may have daily additions, while PSURs are usually updated biannually or annually. Document structures are complex, containing fields such as medical terminology, dosage units, timestamps, patient characteristics, drug information, event descriptions, and causality assessments. Some fields may use standard medical coding dictionaries (e.g., MedDRA, WHODrug).
Constraints Imposed by These Characteristics on Model Access and Configuration
The complexity and high update frequency of pharmacovigilance data impose specific requirements on model access. First, large volumes of unstructured text (e.g., event descriptions) demand efficient text parsing and entity recognition capabilities to accurately extract key information. Second, frequent data updates necessitate a knowledge base with rapid incremental update mechanisms to prevent knowledge staleness. Processing structured data like E2B XML requires specific parsers and mapping rules to ensure data integrity. The presence of specialized medical terminology and coding systems means models need integration or pre-training with relevant domain knowledge to improve understanding accuracy. Additionally, the existence of multilingual documents challenges the model's language processing capabilities, as some adverse event reports may originate from non-English speaking countries.
Configuration Guidelines
| Configuration Item | Suggested Value | Rationale |
|---|---|---|
Chunk size (Chunk Size) | 800–1200 characters | Balances semantic integrity of long texts with model processing efficiency, preventing information truncation or overload. |
Chunk Overlap Length (Chunk Overlap Size) | 100–150 characters | Ensures contextual continuity and reduces the risk of critical information being split at chunk boundaries. |
Recall count (Recall Count) | Top 5–8 entries | Balances recall precision with model input length limits, ensuring highly relevant document segments are retrieved. |
Similarity threshold (Similarity Threshold) | Calibrated by measurement, e.g., 0.75–0.85 | Ensures retrieved results are highly relevant to the query, avoiding the introduction of noisy information. |
Rerank result count (Reranked Return Count) | Top 3 entries | Further refines the most relevant segments after reranking, improving the quality of model responses. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Accommodates parsing time for large PSUR or RMP documents, preventing processing failures due to timeouts. |
Three Common Mistakes
- The model omits critical dosage or time information in its responses. This occurs when medical measurement units and time expressions are not adequately recognized during document parsing, or when the chunking strategy separates key numerical values from their context.
- An
Unexpected end of JSON inputerror appears in the conversation. This may be related to excessively long model output or special characters causing JSON parsing failures, especially when using certain open-source models. - The model's assessment of adverse event causality is irrelevant to the query. This typically results from a lack of contextual medical terminology or relevant clinical guidelines in the knowledge base, or from the text embedding model failing to capture deep semantic relationships between medical concepts.
How to Confirm Proper Configuration
- Upload a hybrid document containing an adverse event report, a PSUR summary, and key sections of an RMP. Verify if the model can accurately extract basic patient information, drug names, adverse event descriptions, and critical date fields.
- For an E2B XML formatted ICSR file, query the model to check if it correctly identifies and reports information corresponding to MedDRA and WHODrug codes.
- Use queries containing specific medical terminology and drug interactions. Verify if the document segments returned by the model cover the relevant specialized knowledge and manually assess their semantic relevance threshold.
The values provided are common starting points and should be measured against specific 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.