Data Characteristics for this Category
Antibody-drug conjugate (ADC) adverse event data originates primarily from clinical trial reports, real-world evidence (RWE) studies, post-market pharmacovigilance systems, and academic journals. This data exists as unstructured text, semi-structured tables, and structured databases. Unstructured text includes patient medical records, narratives from individual case safety reports (ICSRs), and medical literature abstracts. These contain extensive free-text descriptions of symptoms, signs, diagnoses, and treatment processes. Semi-structured data commonly appears in clinical trial safety summary tables, with fields such as dose, administration schedule, adverse event codes (e.g., MedDRA), severity, and onset time. Structured data typically consists of standardized ICSR databases with more regulated fields. ADCs have a unique toxicity profile, often involving hematological, hepatic, renal, and neurological toxicities. Their adverse event descriptions may include specific biomarker changes and pathological findings. Data update frequency is high during clinical trials. Post-market updates depend on reporting systems, usually occurring periodically, but severe adverse event reports may require immediate processing.
Constraints Imposed by these Characteristics on Model Integration and Configuration
The multi-modal nature of ADC pharmacovigilance data, especially the large volume of unstructured text, requires a focus on text processing capabilities during model integration. Professional terminology, abbreviations, and context-dependent information in free text mean that simple keyword matching is insufficient for accurate adverse event identification. This necessitates integrating models with advanced natural language understanding (NLU) capabilities, such as entity recognition, relation extraction, and event detection. The unique toxicity profile and adverse event descriptions of ADCs imply that models require domain knowledge or fine-tuning with domain-specific data to improve the accuracy of identifying this specific information. The periodicity of data updates, particularly the immediate requirements for severe adverse events, challenges real-time model inference and data synchronization mechanisms. Furthermore, inconsistent data formats from different sources, such as tabular structures in clinical trial reports versus narrative text in post-market reports, demand complex data preprocessing and standardization before model integration to ensure data quality and consistency.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale for Recommendation |
|---|---|---|
Chunk size (Segment Length) | 500–800 characters | Balances semantic completeness and model processing efficiency, preventing information loss or context fragmentation from overly long or short segments. |
Recall count (Recall Count) | 10–20 entries | Ensures coverage of relevant adverse event information while controlling the input volume for model inference. |
Similarity threshold (Similarity Threshold) | 0.75–0.85 | Balances recall and precision, preventing the retrieval of irrelevant documents, especially for specialized medical terminology. |
Rerank result count (Reranked Return Count) | 5–8 entries | Further refines recall results, improving the relevance of the final output presented to the user. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Provides sufficient parsing time when processing clinical reports or literature containing large amounts of unstructured text. |
embeddingModel | text-embedding-ada-002 or domain-fine-tuned model | Offers strong semantic understanding for medical text, or improves the vectorization effect for ADC adverse event descriptions through domain fine-tuning. |
Three Common Mistakes
- Model testing returns a
403 status code (no body)error. This typically indicates an incorrect API Key configuration or insufficient permissions, leading to the model service denying access. - Knowledge base retrieval results contain numerous irrelevant or low-relevance documents. This may be due to a
Similarity threshold(Similarity Threshold) set too low, or theembeddingModelfailing to effectively capture the unique semantic information of the ADC domain. - In a workflow, the conversational model fails to correctly identify medical terms containing spaces. This occurs when prompt engineering does not adequately consider tokenization or entity recognition boundaries, causing the model to split terms.
How to Confirm Proper Configuration
- Upload typical ADC adverse event reports to the knowledge base. Perform searches using precise medical terminology. Check the relevance of the recall results, ensuring highly relevant documents are ranked prominently.
- In a test environment, simulate submitting text containing complex adverse event descriptions. Observe if the model accurately identifies key entities (e.g., drugs, symptoms, dosages) and event relationships. Check the output logs for any anomalies.
- After configuration, run multi-turn dialogue tests. Ask questions about specific ADC toxicity profiles. Evaluate the accuracy and completeness of the model's responses. Check if the responses adequately cite knowledge base content.
- Review system logs. Ensure no timeout errors or frequent API call failures occur during file parsing, vectorization, and model inference, especially for large documents.
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.