Bias in Large AI Models for Medicine and Healthcare: Survey and Challenges

📖 Overview

This repository contains the resources, taxonomy, and data associated with our survey paper: "Bias in Large AI Models for Medicine and Healthcare: Survey and Challenges".

Large AI models (including LLMs, LVMs, and LMMs) are transforming healthcare, yet they risk perpetuating or amplifying medical biases. This project provides a comprehensive synthesis of 55 representative studies, organizing the literature into a clear taxonomy of bias, detection methods, and mitigation strategies.

Figure 1: An overview of bias in Large AI models for medicine and healthcare.

🎯 Key Contributions

• Taxonomy: A dual taxonomy categorizing bias by Medical Scenarios (e.g., triage, education) and Clinical Specialties (e.g., cardiology, oncology).
• Resources: A structured index of Large AI Models and Datasets used in bias research.
• Methodology: A review of current techniques for Bias Detection (e.g., counterfactual testing) and Mitigation (pre-, in-, and post-processing).
• Future Directions: Identification of open problems such as the fairness-accuracy trade-off and global health inequities.

---

🧩 Taxonomy of Medical Bias

We categorize medical bias along two principal axes to facilitate precise identification and mitigation.

1. Bias Across Medical Scenarios

• Clinical Decision Support: Disparities in diagnostic reasoning or treatment planning.
• Patient Communication: Biased triage advice or health counseling via chatbots.
• Medical Documentation: Stereotypes or hallucinations in report generation and summarization.
• Medical Education: Misrepresentation in generated case vignettes or training materials.

2. Bias Across Clinical Specialties

Our survey covers biases identified in specific domains, including:

• 🫀 Cardiology
• 🫁 Pulmonology
• 🦀 Oncology
• 🦠 Infectious Disease
• 👁️ Ophthalmology
• 🧠 Mental Health & Psychiatry

---

🛠️ Resources

🤖 General Large AI Models for Medical Bias Research

Below are selected models analyzed in the survey.

Model Name	Family	Parameter Size	Open Source?
GPT-4	GPT	≥ 175B	No
GPT-3.5	GPT	≥ 175B	No
Claude-3.5	Claude	≥ 175B	No
LlaMa-3	LlaMa	≥ 175B	Yes
Qwen-2.5	Qwen	≥ 175B	Yes
Deepseek-V3	Deepseek	≥ 175B	Yes

🤖 Medical Large AI Models for Medical Bias Research

Below are selected models analyzed in the survey.

Model Name	Family	Parameter Size	Open Source?
Med-PaLM 2	PaLM 2	≥ 175B	No
Meditron	LlaMa-2	70B-175B	Yes
PMC-LlaMa	LlaMa	10B-70B	Yes
LLaVA-Med	LLaVA	1B-10B	Yes
ClinicalBERT	ClinicalBERT	< 1B	Yes

📊 Datasets for Bias Detection

We have compiled datasets across three modalities: Text, Image, and Multimodal.

• Text: MedQA, PubMedQA, MIMIC-IV, AMQA, BiasMD.
• Image: CheXpert, MIMIC-CXR, HAM10000, ODIR, Fitzpatrick17k.
• Multimodal: LLaVA-Med, ROCO, PMC-OA.

---

⚙️ Methodology

Bias Detection Techniques

• Input Generation: Creating synthetic patients or mutating existing clinical vignettes (e.g., changing "Male" to "Female").
• Evaluation Metrics:
  • Answer Consistency: Measuring robustness across demographic changes.
  • Fairness Metrics: Demographic Parity, Equalized Odds.
  • Human Expert Assessment: Physician review for complex scenarios.

Bias Mitigation Strategies

1. Pre-processing: Data augmentation and rebalancing before training.
2. In-processing: Model fine-tuning (e.g., FairCLIP), loss function modification.
3. Post-processing: Prompt engineering (Chain-of-Thought), output rewriting, and ensembling.

---

🚀 Open Problems & Opportunities

Based on our analysis, we highlight the following urgent research directions:

1. Unified Foundations: Defining "medical fairness" distinct from general AI fairness.
2. Standardized Benchmarks: Moving beyond ad-hoc testing to rigorous, scalable benchmarks.
3. Real-World Validation: Continuous monitoring of models in deployed clinical settings.
4. Global Health Equity: Addressing the lack of representation for non-Western populations and languages.
5. Fairness-Accuracy Trade-off: Investigating how debiasing affects diagnostic performance.

---

📝 Citation

If you find this survey or repository helpful, please cite our work:

@article{xiao2025bias,
  title={Bias in Large AI Models for Medicine and Healthcare: Survey and Challenges},
  author={Xiao, Ying and Chen, Zhenpeng and Huang, Jen-tse and Chen, Wenting and Liu, Yepang and Li, Kezhi and Mousavi, Mohammadreza and Dobson, Richard and Zhang, Jie},
  year={2025}
}

This Readme file is generated by Gemini-3