Monoclonal Antibody Pharmacovigilance: Model Integration and

Data Characteristics

Monoclonal antibody (mAb) pharmacovigilance data originates from clinical trial reports, real-world studies, post-market surveillance (e.g., FDA FAERS, EMA EudraVigilance databases), and academic literature. This data exists in both structured (e.g., database records, XML files) and unstructured (e.g., free-text descriptions, case reports) formats. Update frequency varies: post-market surveillance data is typically aggregated quarterly or annually, while clinical trial data updates periodically as trials progress. Document structures are diverse, including standardized ICH E2B reports, Case Safety Reports (CSRs), and medical journal articles. Fields and units are highly specific, for example: drug names (e.g., Adalimumab), adverse event terms (using MedDRA codes), dosages (mg/kg or mg), routes of administration (e.g., intravenous, subcutaneous), treatment duration (days, weeks), and patient characteristics (e.g., age, weight, comorbidities).

Constraints on Model Integration and Configuration

The highly structured and specific nature of mAb data imposes particular requirements on model integration. The presence of standard terminology systems like MedDRA codes means models must identify and map these terms when processing adverse events, avoiding simple text matching. Data heterogeneity from multiple sources requires the integration layer to handle diverse data stream formats and perform effective standardization preprocessing. For instance, free-text descriptions of adverse events need Named Entity Recognition (NER) to extract key information and match it with MedDRA terms. Inconsistent update frequencies necessitate considering data freshness during model training and inference, potentially requiring incremental learning or periodic full updates. Dosage unit differences (mg/kg vs. mg) and varied administration route expressions require models to perform unit conversion and semantic normalization when understanding and correlating this information, ensuring accurate association analysis between drug exposure and adverse reactions.

Configuration Settings

Configuration Item	Suggested Value	Rationale
`Chunk size` (Segment Length)	`500–800 characters`	mAb pharmacovigilance reports often contain detailed medical histories and adverse event descriptions. Longer segments help preserve context and prevent truncation of critical information.
`Recall count` (Recall Count)	`Top 10–15 items`	mAb adverse reaction patterns can be complex. More relevant document snippets provide comprehensive background to aid model judgment.
`Similarity threshold` (Similarity Threshold)	`0.75–0.85`	Given the precision requirements of `MedDRA` terminology, a higher similarity threshold helps recall drug or adverse event information highly relevant to the query intent.
`Rerank result count` (Rerank Return Count)	`Top 5 items`	After reranking, the top few results typically have the highest quality and cover the core issues.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	Parsing large clinical trial reports or post-market aggregate data can be time-consuming. Extending the timeout prevents parsing failures due to large file sizes.
`embeddingModel`	`bge-large-zh` or `text-embedding-ada-002`	`bge-large-zh` offers strong semantic understanding for Chinese medical texts. `text-embedding-ada-002` performs stably in multilingual and general domains. Select based on actual data language distribution.

Common Pitfalls

Inaccurate or missing adverse event terms in model results. This occurs when MedDRA codes are not preprocessed or the model lacks fine-tuning for specific medical terminology.
connection refused errors when uploading large XML-formatted ICH E2B reports. This typically indicates incorrect network port configuration or resource limits for local vllm containers or OneAPI services.
Poor model performance when handling questions involving dosage unit conversion. This occurs when the prompt does not explicitly instruct the model to perform unit normalization or the model lacks relevant knowledge.

Validation Steps

Select a typical mAb pharmacovigilance report containing various adverse events and dosage information. Ask a question about a specific adverse event's dosage correlation. Check if the model accurately identifies the drug, adverse event, and dosage, and provides results as requested.
Upload a complex case safety report with free-text descriptions. Check FastGPT system logs to confirm no timeout or parse error during file parsing and that key entities are correctly extracted.
Use a query containing MedDRA codes. Verify that the document snippets recalled by the model include the corresponding coding information. Manually evaluate if the semantic relevance of the recalled content meets the expected threshold.

Note: The values provided are common starting points. Measure 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.