Biomedical Large Language Models and Prompt Engineering for Causality Assessment of Individual Case Safety Reports in Pharmacovigilance

This study evaluated the performance of biomedical Large Language Models combined with specific prompting strategies and causality algorithms on 150 Individual Case Safety Reports, finding that while the Medicine LLaMA-3 8B model with Chain-of-Thought prompting achieved the highest agreement with human experts (64%), these models remain suboptimal for reliable individual-level causality assessment in pharmacovigilance.

The Gist

Imagine you are a detective trying to solve a mystery: Did a specific medicine cause a patient's bad reaction, or was it just a coincidence?

In the world of medicine, this is called Causality Assessment. It's a crucial job. If a drug is dangerous, we need to know so we can warn the public. If it's safe, we don't want to panic unnecessarily.

Traditionally, this job is done by human experts (like senior doctors and pharmacists) who read through messy, incomplete patient reports and ask themselves a checklist of questions: Did the reaction happen right after taking the pill? Did it stop when they stopped the pill? Is there another explanation?

This paper is about a new experiment: Can we teach a super-smart computer (an AI) to be this detective?

The Experiment: The AI Detectives

The researchers didn't just use any computer. They used Biomedical Large Language Models (LLMs). Think of these as "detectives who have read every medical textbook in the library," rather than general AI that has read the entire internet (which includes a lot of nonsense).

They tested five different detective teams, each made of:

1. The AI Model: Three different "brains" (TinyLlama, Medicine LLaMA-3, and MedLLaMA).
2. The Prompt Strategy: How they asked the AI to think. Did they say, "Just guess"? Or did they say, "Think step-by-step like a human" (Chain-of-Thought)? Or "Break the problem into tiny pieces" (Decomposition)?
3. The Rulebook: Two different official checklists used by real doctors (The Naranjo Scale and the WHO-UMC method).

They gave these AI teams 150 real-life mystery cases (some from regular drugs, some from vaccines) and asked them to solve the cases. Then, they compared the AI's answers to the answers given by two real human experts.

The Results: The Good, The Bad, and The "Almost"

1. The "Medical School" Advantage
The AI models trained specifically on medical books performed much better than general AI models. It's like comparing a detective who has read the Encyclopedia of Crime to one who has only read Wikipedia. The medical AI got about 64% agreement with the human experts. That's a big jump from the 34% agreement seen with general AI in previous studies.

2. The "Step-by-Step" vs. The "Story"
The AI did surprisingly well when using the Naranjo Rulebook. This rulebook is like a math test: "Question 1: Yes/No. Question 2: Yes/No. Add up the points." The AI was good at following these rigid steps.
However, the AI struggled with the WHO Rulebook. This one is more like writing an essay or telling a story about the patient. The AI got confused, often giving answers that didn't match the human experts at all. It seems the AI is better at math than at writing a coherent medical story.

3. The "Hallucination" Problem
Even the best AI team made mistakes.

• The "Echo" Effect: Sometimes, the AI would just repeat the question back to you instead of answering it (like a parrot).
• The "Confident Wrong" Effect: If a patient report was missing information, the human expert would say, "I don't know." The AI, however, would confidently guess an answer. It's like a student taking a test who doesn't know the answer but writes a long, confident paragraph anyway.
• The "Missing Link": The AI was bad at checking if a side effect was officially listed in the drug's manual (because the AI didn't have access to that specific manual during the test) and was bad at spotting if another disease caused the problem.

The Big Takeaway

Can we replace human doctors with AI for this job yet? No.

The AI is a very fast, very knowledgeable intern, but it's not a Senior Detective yet.

• It's faster: It can read reports instantly.
• It's knowledgeable: It knows medical terms well.
• But it lacks judgment: It can't always tell the difference between a "maybe" and a "definitely," and it sometimes makes up reasons to support its answer.

The Future

The authors suggest that in the future, we shouldn't just ask the AI to "solve the case." Instead, we should use a hybrid approach:

• Let the AI do the heavy lifting (reading the reports, checking the facts).
• Let the human expert make the final call, using the AI's work as a helpful assistant.

Think of it like GPS vs. a Human Driver. The GPS (AI) is great at knowing the roads and traffic rules, but the Human Driver (Expert) is still needed to make the final decision when the road gets muddy or the map is wrong.

In short: Biomedical AI is getting much better at being a medical detective, but it still needs a human partner to ensure the verdict is fair, accurate, and safe.

Technical Summary

1. Problem Statement

Pharmacovigilance relies heavily on the causality assessment of Individual Case Safety Reports (ICSRs) to determine the likelihood that a drug or vaccine caused an adverse event (AE) or adverse event following immunization (AEFI).

• Current Bottleneck: Manual assessment by human experts is time-consuming, resource-intensive, and increasingly unsustainable due to surges in reporting volumes.
• Limitations of Existing AI: Previous studies using general-purpose Large Language Models (LLMs) showed suboptimal performance in individual-level causality assessment, often attributed to a lack of domain-specific training data and inadequate prompt strategies.
• Research Gap: There is a lack of evidence regarding whether biomedical-specific LLMs (trained on medical literature), when combined with advanced prompt engineering (e.g., Chain-of-Thought) and established causality algorithms, can reliably replicate human expert reasoning and classification.

2. Methodology

The study employed a comparative experimental design to evaluate the agreement between biomedical LLMs and human experts.

• Dataset:
  • Total: 150 ICSRs.
  • Sources: 140 cases from the FDA Adverse Event Reporting System (FAERS) and 10 clinically confirmed myocarditis/pericarditis cases from the Vaccine Adverse Event Reporting System (VAERS).
  • Diversity: Included newly approved drugs, gene/cell therapies, common prescriptions, orphan drugs, and controlled substances.
• Models Evaluated: Three open-source biomedical LLMs were selected based on Hugging Face Open LLM Leaderboard performance (as of Jan 2025):
  1. TinyLlama 1.1B (Afrideva)
  2. Medicine LLaMA-3 8B (QuantFactory)
  3. MedLLaMA v20 (JL42)
• Prompt Strategies:
  • Chain-of-Thought (CoT): Encourages step-by-step reasoning.
  • Decomposition: Breaks complex tasks into sub-tasks.
  • Note: Prompts adhered to CLEAR principles (Concise, Logical, Explicit, Adaptive, Reflective).
• Causality Algorithms:
  1. Naranjo Algorithm: A 10-question weighted scale (Doubtful, Possible, Probable, Definite).
  2. WHO-UMC Algorithm: A structured 4-step process (Eligibility, Checklist, Algorithm, Classification).
• Gold Standard: Two human experts (a pharmacovigilance pharmacist and a senior medical doctor) manually assessed all 150 cases using both algorithms.
• Evaluation Metrics:
  • Agreement: Measured via Gwet's Agreement Coefficient 1 (AC1) and percentage agreement.
  • Performance: Accuracy, Sensitivity, Specificity, Precision, Recall, and F1-score.
  • Qualitative Analysis: Error classification (e.g., instruction drift, prompt echoing, hallucinations, lack of justification).

3. Key Contributions

• Domain-Specific vs. General Models: Demonstrated that biomedical LLMs significantly outperform general-purpose LLMs in causality tasks, nearly doubling inter-rater agreement with human experts (reaching ~64% vs. ~34% in prior studies).
• Algorithm Sensitivity: Revealed that the choice of causality algorithm is a critical determinant of performance. Structured, itemized algorithms (Naranjo) yielded much higher alignment than narrative-dependent frameworks (WHO-UMC).
• Error Taxonomy: Provided a detailed classification of failure modes specific to biomedical LLMs in regulatory contexts, including "prompt echoing" (linguistic mimicry without reasoning) and "instruction drift."
• Reproducibility: The study adhered to FAIR principles, releasing all code, prompts, and hyperparameter configurations via a public GitHub repository.

4. Key Results

• Best Performing Configuration: The Medicine LLaMA-3 8B + Naranjo Algorithm + Chain-of-Thought combination achieved the highest agreement with human experts for final causality classification (64%; AC1 = 0.644).
• Algorithm Disparity:
  • Naranjo: Models showed moderate-to-strong agreement on structured questions (e.g., dechallenge/rechallenge outcomes) but struggled with semantic tasks like "alternative causes" and "temporal plausibility."
  • WHO-UMC: Performance dropped significantly. The Medicine LLaMA-3 8B + WHO-UMC + CoT combination achieved only 23.3% agreement (AC1 = 0.297).
• Prompt Engineering Impact: While CoT and Decomposition showed similar overall performance, the structural nature of the Naranjo algorithm (step-by-step) aligned better with CoT prompting than the more narrative WHO-UMC framework.
• Critical Failure Modes:
  • Low Sensitivity/Recall: Models frequently missed positive causal links, classifying many "Possible" cases as "Doubtful."
  • Reasoning Gaps: Even when scores were correct, models often failed to provide the necessary clinical justification or evidence, a critical flaw for regulatory "human-in-the-loop" workflows.
  • Prompt Echoing: Models sometimes mimicked the prompt structure rather than generating independent reasoning.
  • Epistemic Uncertainty: Unlike human experts who assign "unknown" to ambiguous cases, LLMs tended to generate high-certainty categorical responses despite missing data.

5. Significance and Implications

• Regulatory Readiness: While biomedical LLMs show promise, they are not yet ready for autonomous deployment in pharmacovigilance. The lack of consistent, explainable reasoning undermines their utility in highly regulated environments (e.g., EU) where transparency and traceability are mandatory.
• Human-in-the-Loop Necessity: The study reinforces that LLMs should currently serve as decision-support tools rather than decision-makers. They can assist in triaging or drafting, but human experts must validate the reasoning and final classification.
• Future Directions:
  • Agentic AI: Future iterations may require "Agentic" architectures that can access external tools (e.g., live SmPC databases) to answer questions like "Is this AE listed?" which current static models cannot do.
  • Hybrid Workflows: Optimizing the interface between structured algorithms (Naranjo) and LLM reasoning to minimize instruction drift.
  • Training Data: The need for training data that includes specific regulatory documents (SmPCs) and explicit reasoning chains for causality.

Conclusion: Biomedical LLMs represent a significant step forward in automating pharmacovigilance, offering nearly double the alignment of general models. However, their inability to consistently provide evidence-based reasoning and their sensitivity to algorithmic structure currently limit their application to a supportive role under strict human supervision.