LLM-Assisted Longitudinal Adverse Event Tracking in Clinical Trials

AI-assisted framework for standardized, longitudinal CTCAE-based adverse event (AE) extraction, normalization, grading, and patient-level timeline consolidation in oncology clinical trials.

---

Overview

Adverse event (AE) documentation is central to safety monitoring in oncology clinical trials. However, AE identification, CTCAE mapping, grading, and attribution remain largely manual, labor-intensive, and inconsistent across visits.

We present an end-to-end LLM-assisted framework that:

• Extracts AEs from clinical notes and laboratory records
• Normalizes events to CTCAE v5.0
• Assigns severity grade and attribution
• Flags immune-related and serious AEs
• Maintains patient-level longitudinal AE trajectories

The system integrates LLM reasoning with ontology-guided semantic retrieval and deterministic grading logic to support real-world safety reporting workflows.

---

Clinical Context

CTCAE-Based Safety Reporting

Adverse events in oncology trials are standardized using:

• Common Terminology Criteria for Adverse Events (CTCAE v5.0)

CTCAE defines:

• Preferred AE terms
• Severity grades (1–5)
• Standardized terminology for regulatory reporting

This framework explicitly aligns extraction and normalization to CTCAE v5.0.

---

Pipeline Overview

Framework Architecture

Core Modules

1. Multi-source Ingestion

• Physician progress notes
• Research nurse notes (AE listings removed)
• Structured laboratory panels

2. LLM-based AE Extraction

• Context-aware symptom extraction
• Temporal anchoring
• Structured JSON output

Example:

Example:

```json
[
  {
    "AE_term": "diarrhea",
    "grade": 2,
    "attribution": "Possible",
    "start_date": "2024-01-12",
    "end_date": "ongoing",
    "immune_related": "No",
    "serious": "No"
  }
]

3. CTCAE Semantic Normalization

• Embedding-based Top-3 retrieval
• Fine-tuned MedCPT model
• Ontology-aligned mapping

4. Longitudinal AE Timeline Modeling

• Onset tracking
• Resolution handling
• Grade evolution across visits
• Recurrence detection