Data Characteristics
Imaging equipment pharmacovigilance data originates from medical imaging diagnostic reports, equipment usage logs, adverse event reporting systems (e.g., national drug adverse reaction monitoring centers, FDA MAUDE database), and equipment manufacturer technical documentation. Data update frequencies vary. Diagnostic reports and usage logs might generate daily, while adverse event reports depend on event frequency and reporting procedures, typically summarized weekly or monthly. Document structures are diverse. Imaging diagnostic reports are often unstructured text, including examination methods, imaging findings, and diagnostic conclusions. Equipment usage logs are structured or semi-structured data, recording model numbers, serial numbers, operating parameters, and error codes. Adverse event reports include fields such as patient information, equipment information, adverse event descriptions, and handling measures. Key fields include device_id, adverse_event_description, imaging_findings, and operating_parameters.
Constraints Imposed by These Characteristics on Knowledge Base Retrieval and Recall
The heterogeneity of imaging equipment data sources and their diverse structures necessitate robust multi-modal data processing capabilities in the knowledge base to effectively integrate unstructured text and semi-structured logs. Inconsistent update frequencies require the knowledge base to support incremental updates and version management, ensuring the timeliness of recall results. Unstructured imaging findings descriptions in diagnostic reports demand advanced word segmentation and entity recognition to accurately extract key information related to adverse reactions. The large volume of specialized terminology and codes in equipment usage logs requires semantic enhancement through domain-specific dictionaries or embedding models to improve retrieval precision. Additionally, sensitive patient information in adverse event reports mandates strict adherence to data anonymization and privacy protection regulations during data processing and retrieval to prevent disclosure.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Chunk Size) | 500–800 characters | Imaging reports and adverse event descriptions are often long. Chunks that are too short can split context, affecting semantic completeness. |
Chunk Overlap Length (Overlap Size) | 100–150 characters | Ensures that critical information spanning across chunks is not lost, improving the robustness of recall. |
Recall count (Recall Count) | 8–12 items | Ensures coverage of multiple data sources and potential associations while avoiding an excessive amount of irrelevant information. |
Similarity threshold (Similarity Threshold) | 0.78–0.85 | Requires balancing recall rate and accuracy. Descriptions of imaging equipment adverse reactions can have various formulations. |
embeddingModel | text-embedding-ada-002 or domain-specific model | Prioritizes embedding models that can understand medical domain terminology to improve semantic matching accuracy. |
Rerank result count (Reranked Return Count) | top 5 items | After reranking, focuses on the most relevant few items to improve the effectiveness of the final results. |
Common Pitfalls
- Symptom: Retrieved reports have weak relevance to the query keywords, or irrelevant content appears. Reason: The chunk size is set too large, causing individual chunks to contain too much irrelevant information, diluting the key semantics.
- Symptom: Queries for equipment models or error codes fail to recall relevant log entries. Reason: During knowledge base construction, structured fields in equipment usage logs were not properly processed for keyword extraction or semantic embedding, leading to literal match failures.
- Symptom: Inputting "equipment overheating" fails to recall reports describing "equipment over-temperature." Reason: The
Similarity threshold(Similarity Threshold) is set too high, failing to capture semantic associations between synonyms or near-synonyms, resulting in insufficient recall.
Validation Steps
- Select typical adverse event queries, such as "CT machine radiation dose abnormal," and check if the recall results include multiple imaging diagnostic reports and equipment logs describing similar events. Evaluate their relevance.
- Query for specific imaging equipment models and serial numbers. Verify if the recall results include historical maintenance records or adverse event reports for that equipment, and confirm that the
device_idfield matches correctly. - Use queries containing medical professional terminology (e.g., "contrast extravasation," "artifact") to verify if the knowledge base can accurately recall relevant imaging reports or technical documents. Evaluate the completeness and contextual coherence of the recalled items.
- Simulate edge cases, such as ambiguously described adverse events, to observe the quality and diversity of recall results, ensuring the system provides valuable information even in complex scenarios.
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.