💻 computer science

Ablation Study of a Fairness Auditing Agentic System for Bias Mitigation in Early-Onset Colorectal Cancer Detection

This study demonstrates that an agentic AI system enhanced with Retrieval-Augmented Generation (RAG) significantly improves the accuracy of fairness auditing for early-onset colorectal cancer detection models compared to standalone large language models or agents without retrieval capabilities.

Original authors: Amalia Ionescu, Jose Guadalupe Hernandez, Jui-Hsuan Chang, Emily F. Wong, Paul Wang, Jason H. Moore, Tiffani J. Bright

Published 2026-03-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Amalia Ionescu, Jose Guadalupe Hernandez, Jui-Hsuan Chang, Emily F. Wong, Paul Wang, Jason H. Moore, Tiffani J. Bright

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are building a very smart robot doctor to help diagnose a specific type of cancer called Early-Onset Colorectal Cancer (EO-CRC). This is cancer that strikes people under 50, and sadly, it doesn't affect everyone equally. Some groups (like Black and Hispanic patients, or those with less money) often get diagnosed later and have a harder time surviving.

The big worry is: What if our new robot doctor accidentally learns these same unfair biases? If the robot is biased, it might miss the cancer in the people who need help the most, making the health gap even wider.

This paper is about building a "Fairness Auditor"—a special team of AI robots designed to check if the main robot doctor is being fair before it ever sees a patient.

Here is how the researchers tested their system, explained simply:

1. The Two-Worker Team

Instead of just one big brain, the researchers built a team of two specialized AI agents (think of them as two different employees):

Employee #1: The "Detective" (Domain Expert Agent)
- Job: This agent reads thousands of medical articles to find out who is at risk. It asks: "Are there specific groups of people who get this cancer later or die sooner?"
- Goal: To create a list of "sensitive attributes" (like race, gender, or income) that the main robot needs to watch out for.
Employee #2: The "Judge" (Fairness Consultant Agent)
- Job: This agent takes the Detective's list and decides how to measure fairness. It asks: "What math rules should we use to make sure the robot isn't ignoring the Detective's findings?"
- Goal: To pick the right fairness metrics (like "False Negative Rate") to ensure the robot doesn't miss cancer in vulnerable groups.

2. The Big Question: Do They Need a Library?

The researchers wanted to know if these AI workers needed help from an external library (a database of facts) or if they could just rely on what they already knew from their training.

They tested three different "brains" (AI models of different sizes: Small, Medium, and Giant) in three different scenarios:

The "Memory Only" Test: The AI tries to answer using only what it memorized during training (no outside help).
The "No Library" Agent Test: The AI acts like a worker but isn't allowed to look up facts in the library.
The "Library" Agent Test (RAG): The AI is a worker with a library. It can look up the latest medical papers and fairness rules while it works.

Note: "RAG" (Retrieval-Augmented Generation) is just a fancy way of saying "Look it up before you answer."

3. The Results: The Library Wins (Mostly)

Here is what they found, using a simple analogy:

For the Detective (Employee #1):
- The Result: Giving the Detective access to the library was a huge game-changer.
- The Analogy: Imagine asking a student to write a report on a specific disease. If they only use their memory, they might get the general idea but miss the specific details about who is most affected. If you give them a stack of the latest research papers (the library), their report becomes much more accurate and matches what a human expert would say.
- Why? The AI's memory isn't perfect about specific, niche medical disparities. It needs the library to stay grounded in reality.
For the Judge (Employee #2):
- The Result: The library helped, but it wasn't as critical.
- The Analogy: The Judge is like a math teacher. They already know the rules of fairness (like "treat everyone equally") from their training. They don't need to look up a dictionary to know what "fairness" means. However, the library still helped the biggest AI models perform slightly better.
- The Catch: The biggest AI model (120B parameters) was so smart that it sometimes didn't even bother opening the library! It thought it knew enough on its own.

4. The Size Matters (But Not Always)

The researchers tested small, medium, and giant AI models.

The Giant Model (120B): Was the smartest overall, but it's heavy and expensive to run (like a supercomputer).
The Small Model (8B): Was surprisingly good if it had the library.
The Lesson: You don't always need the most expensive, giant brain. A smaller, cheaper brain can do an excellent job if you give it access to a good library (RAG). This is great news for hospitals that don't have millions of dollars for supercomputers.

5. Why This Matters

Right now, hospitals are installing AI to help diagnose patients. But if we don't check these AIs for bias, they might accidentally hurt the people who need help the most.

This paper proves that we can build an automated "Fairness Auditor" that:

Reads the latest medical research.
Checks if the AI is being fair to different groups.
Does this automatically, without needing a team of human experts to stare at code for months.

In a nutshell: To make AI fair in healthcare, don't just rely on the AI's memory. Give it a library to check its facts, and even a small, affordable AI can become a powerful guardian of equity.

1. Problem Statement

The rapid integration of AI into clinical decision support systems (e.g., 70% of U.S. hospitals) has outpaced the capacity for rigorous safety monitoring. This creates a critical "expertise bottleneck" where:

Algorithmic Bias: Unmonitored models often perpetuate structural biases (e.g., under-referring minority patients or misdiagnosing specific demographics), leading to "silent failures."
Context Dependency: Determining appropriate fairness metrics and sensitive attributes requires nuanced clinical judgment that generic automation tools lack.
Scalability & Latency: Manual auditing by cross-disciplinary expert teams is slow, resource-intensive, and cannot keep up with the pace of model deployment.

The authors aim to determine if an Agentic AI system (autonomous agents driven by Large Language Models) can autonomously audit biomedical ML models for fairness, specifically in the context of Early-Onset Colorectal Cancer (EO-CRC), a condition marked by significant racial, ethnic, and socioeconomic disparities.

2. Methodology

The study proposes a two-agent architecture equipped with Retrieval-Augmented Generation (RAG) to synthesize evidence from clinical literature and recommend fairness strategies.

System Architecture

Domain Expert Agent (Agent 1):
- Task: Synthesizes clinical literature to identify disparity drivers (e.g., race, SES, gender) and generates a narrative rationale.
- Knowledge Source: A corpus of 39 journal articles (22 on EO-CRC disparities, 17 "noise" articles).
Fairness Consultant Agent (Agent 2):
- Task: Uses Agent 1's output to recommend specific fairness metrics (e.g., Equal Opportunity, False Negative Rate Parity) and sensitive attributes for model evaluation.
- Knowledge Source: A curated library of formal fairness metric definitions.

Experimental Design (Ablation Study)

The authors evaluated the system across three Large Language Model (LLM) scales: Llama 3.1 (8B), GPT-OSS (20B), and GPT-OSS (120B). They compared three configurations for each model:

Pretrained LLM-only: The model relies solely on internal parametric knowledge.
Agent without RAG (Agent NR): The agent framework exists but does not retrieve external data.
Agent with RAG (Agent R): The full system with dynamic retrieval from external knowledge bases.

Evaluation Metrics

Semantic Similarity: Outputs were compared against expert-derived ground truth statements using cosine similarity over sentence embeddings.
Tool Usage: Frequency of retrieval tool invocation was tracked.
Statistical Analysis: Kruskal–Wallis tests and Wilcoxon rank-sum tests with Holm correction were used to determine significance.

3. Key Contributions

Agentic Framework for Clinical Auditing: Demonstrates a scalable, autonomous workflow for identifying bias and recommending fairness metrics in high-stakes clinical domains.
Ablation of RAG and Agent Orchestration: Provides empirical evidence on the specific contributions of retrieval (RAG) versus model scale and agent structure in reducing hallucinations and improving alignment with expert knowledge.
Resource-Constrained Viability: Evaluates the feasibility of running these systems on smaller, locally deployable models (8B parameters) versus large-scale models (120B).

4. Key Results

Performance by Agent Role

Agent 1 (Disparity Identification):
- RAG was critical. The "Agent with RAG" configuration consistently achieved the highest semantic similarity scores across all model scales (8B, 20B, 120B).
- Statistical Significance: RAG significantly outperformed both LLM-only and Agent-without-RAG conditions ( $p < .001$ ).
- Tool Usage: The retrieval tool was invoked in 100% of runs, indicating that external evidence is essential for identifying specific clinical disparities.
Agent 2 (Fairness Metric Selection):
- Variable Impact of RAG. The benefit of RAG depended heavily on model scale.
  - 8B Model: RAG significantly improved performance.
  - 20B Model: No significant difference was found between configurations, suggesting the model's internal knowledge was sufficient for this specific task.
  - 120B Model: RAG and Agent frameworks improved performance over LLM-only, but the difference between "Agent with RAG" and "Agent without RAG" was not statistically significant ( $p = 0.24$ ).
- Tool Usage: The 120B model invoked the retrieval tool in only 85% of runs, suggesting larger models can sometimes rely on internal pretraining for fairness concepts.

Model Scale Comparison

Llama 3.1 (8B): With RAG, it achieved strong performance (Mean Similarity ~0.80 for Agent 1), making it a viable option for resource-constrained environments.
GPT-OSS (120B): Achieved the highest absolute semantic similarity scores (Mean ~0.83 for Agent 1), but required significantly more computational resources.

5. Significance and Conclusion

Scaling Fairness Audits: The study proves that agentic systems with RAG can effectively scale the auditing of clinical AI, reducing the reliance on scarce human experts for initial disparity identification.
Task-Dependent Retrieval: The necessity of RAG is task-dependent. Knowledge-intensive tasks (identifying specific demographic disparities) strictly require external retrieval to avoid hallucinations. Conceptual tasks (selecting fairness metrics) may rely more on the model's internal pretraining, especially at larger scales.
Practical Deployment:
- For resource-constrained settings, the 8B model paired with RAG offers a practical, high-performance solution that can run on local hardware (e.g., laptops).
- For high-resource settings, the 120B model with RAG provides the highest alignment with expert standards.
Future Directions: The authors call for broader evaluations across more clinical domains and direct expert assessment of the clinical validity (not just semantic similarity) of the generated recommendations.

In summary, this paper establishes that Retrieval-Augmented Generation is a non-negotiable component for accurate clinical disparity detection in AI auditing systems, while the necessity of large model scales varies depending on the complexity of the reasoning task.