Data Characteristics
Pharmacovigilance data in hospital operations originates from internal hospital information systems. These include electronic medical record (EMR) systems, medication management systems, and adverse event reporting systems. Data exists in a mixed format of structured and unstructured information. Structured data includes patient demographics, medication records, diagnosis results, and adverse reaction classification codes (e.g., ICD-10, WHO-ART codes). Unstructured data includes physician ward rounds notes, nurse observation records, patient chief complaints, and adverse event descriptions.
Data updates frequently. Medication information and adverse event reports are typically entered in real-time or near real-time. Document structures are diverse, ranging from standardized report templates to free-text descriptions. Fields and units include drug names, dosages, frequencies, administration routes, occurrence times (down to hours or minutes), adverse reaction symptom descriptions, severity assessments (e.g., Naranjo scale scores), and relevant laboratory test results (with units like mg/dL, IU/L).
Constraints on Knowledge Base Retrieval and Recall
The complexity of hospital operations pharmacovigilance data imposes multiple constraints on knowledge base retrieval and recall. First, real-time updates require the knowledge base to have an efficient incremental update mechanism to ensure the timeliness of recall results. Second, the mix of structured and unstructured data necessitates support for multimodal retrieval. For example, the system must match medication codes and natural language descriptions of adverse reaction symptoms simultaneously.
Diverse document structures, especially the presence of free text, challenge text segmentation strategies and entity recognition capabilities. The system must accurately identify key entities such as drugs, symptoms, and times. The specialized nature of fields and units requires the retrieval model to possess domain knowledge. It must understand medical terminology and dosage unit meanings to avoid recall biases caused by synonyms or abbreviations. For instance, a search for "allergy" must associate with specific symptoms like "rash" or "urticaria." High-precision recall is essential. Both false positives and false negatives can lead to serious medical risks.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Segment Length) | 500–800 characters | Ensures each segment contains sufficient context while avoiding information redundancy, which aids entity recognition and semantic understanding. |
Chunk Overlap Rate (Segment Overlap Rate) | 10%–20% | Maintains context continuity and reduces semantic loss due to segment truncation, especially for lengthy physician notes. |
Recall count (Recall Count) | 10–20 items | Balances recall breadth with model processing load, covering potentially relevant adverse event reports. |
Similarity threshold (Similarity Threshold) | 0.75–0.85 | Balances recall precision and recall rate, reducing false positives while avoiding missing critical adverse reaction information. |
Rerank result count (Reranked Return Count) | 3–5 items | Focuses on the most relevant high-quality results during the reranking phase, reducing the burden of subsequent manual review. |
Index Update Cycle | 1 hour | Responds to the real-time requirements of pharmacovigilance data, ensuring the timeliness of knowledge base content. |
Common Mistakes
- Retrieval results contain many irrelevant or low-relevance reports. This usually occurs when the
Similarity threshold(Similarity Threshold) is set too low, leading to an overly broad recall scope. - A user query for "palpitations caused by a specific drug" fails to recall relevant adverse event descriptions. This might happen if the
Chunk size(Segment Length) is too short, causing the palpitation symptom and drug information to be truncated into different segments. - After importing a large number of Markdown-formatted drug instructions into the knowledge base, some content is not correctly indexed. This happens when the system's default parser has insufficient support for specific Markdown syntax, leading to
PARSE_FILE_TIMEOUT_SECONDStimeouts or parsing errors.
Validation Steps
- Select typical queries covering various drugs, adverse reaction types, and severity levels. Check the
similarityscore distribution of the recall results to assess their relevance and ranking. - Regularly simulate the import of new adverse event reports. Check the knowledge base index update speed and the discoverability of new data in retrieval.
- For adverse reactions to specific drugs, design queries that include synonyms, abbreviations, and vague descriptions. Validate the knowledge base's understanding and recall capabilities for medical terminology.
- Verify the completeness and accuracy of key fields (e.g.,
occurrence time,dosage) in the recall results. Ensure information is not lost during segmentation or recall.
Note: The values provided are common starting points. Measure them 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.