Data Characteristics
CAR-T cell therapy pharmacovigilance data originates from clinical trial reports, real-world evidence (RWE) databases, hospital electronic health record (EHR) systems, patient reports, and regulatory adverse event submissions. This data is typically semi-structured or unstructured, including free-text descriptions, medical terminology codes (e.g., MedDRA terms), laboratory test results, and imaging reports. Update frequency is higher during clinical trials, potentially weekly or monthly. Post-market surveillance is more dispersed; adverse event reports are immediate, but aggregated analysis might occur quarterly or annually. Document structures are complex, encompassing multi-page PDF reports, Word documents, or JSON/XML data exchange files. Fields cover patient demographics, diagnosis, treatment plans, adverse event descriptions, severity, onset time, outcomes, and relevant laboratory indicators (e.g., cytokine levels, complete blood count). Units vary, such as time units (days, weeks), dosage units (mg/kg), and concentration units (pg/mL).
Constraints on Model Integration and Configuration
The complexity of CAR-T cell therapy pharmacovigilance data imposes specific requirements on model integration and configuration. The large volume of semi-structured and unstructured text data demands robust text processing capabilities and longer context windows. The coexistence of periodic and real-time data updates requires models to support scheduled incremental updates and real-time event responses. Diverse document structures necessitate flexible file parsers capable of handling various formats. The presence of medical terminology codes means the model needs medical knowledge graph understanding or semantic matching via vector embeddings. Diverse fields and units, especially numerical data from laboratory indicators, require the model to accurately identify values and units during extraction and perform normalization. These constraints collectively determine the need for models supporting long texts and strong multimodal parsing capabilities within the FastGPT platform, along with appropriate knowledge base segmentation strategies, retrieval algorithms, and reranking mechanisms.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 16000 | CAR-T pharmacovigilance reports often contain detailed clinical descriptions and multi-round follow-up records, requiring a large context window to capture complete information. |
Chunk size (Segment Length) | 800–1200 characters (characters) | Ensures each segment contains a complete adverse event description or key laboratory result, while avoiding redundancy from excessive length. |
Recall count (Retrieval Count) | Top 10 entries (top 10) | The complexity of clinical event descriptions requires retrieving more potentially relevant knowledge snippets to improve accuracy. |
Similarity threshold (Similarity Threshold) | 0.75 | Balances recall and precision, ensuring a match with highly relevant medical terms and clinical manifestations related to adverse event descriptions. |
Rerank result count (Reranked Return Count) | Top 5 entries (top 5) | After initial retrieval, reranking further refines the results, focusing on the most relevant evidence and reducing the model's processing burden. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds (seconds) | Parsing large PDF or scanned documents can be time-consuming; increasing the timeout prevents parsing failures. |
Common Mistakes
- Model returns empty fields for drug adverse events because it lacks pre-training or fine-tuning on medical terminology and abbreviations, failing to accurately identify key information.
- After uploading a large clinical trial report file, the system remains unresponsive for an extended period or displays a "file parsing timeout" error. This occurs because the
PARSE_FILE_TIMEOUT_SECONDSparameter is set too low, failing to accommodate the parsing time for complex documents. - When retrieving CAR-T related adverse reactions, the model recalls numerous side effects unrelated to CAR-T therapy. This happens because the knowledge base segmentation strategy is too coarse, failing to effectively distinguish adverse events under different treatment modalities.
Validation Steps
- Upload a clinical case report containing typical CAR-T adverse reactions (e.g., cytokine release syndrome, neurotoxicity). Observe if the model can accurately extract and summarize the main adverse events.
- Test with structured adverse event reports containing MedDRA codes. Verify if the model can correctly identify and associate them with corresponding medical terms.
- Validate that when new CAR-T therapy-related adverse event reports are added to the knowledge base, the model can promptly reflect this new information in queries, ensuring the incremental update mechanism is effective.
The values provided are common starting points and should be measured against the reader's 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.