Research map / Label-efficient and robust learning / Recognition and long-tail
The Boundaries of Fair AI in Medical Image Prognosis: A Causal Perspective
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces FairTTE, a benchmark and causal framework for group fairness in time-to-event prediction from medical images, evaluating three TTE models and five fairness algorithms on AREDS, MIMIC-CXR, and ADNI.
Problem
Fairness in medical image prognosis (time-to-event prediction) has been largely overlooked; existing fairness work focuses on diagnosis tasks, and no standard metrics, datasets, or benchmarks exist for fair TTE prediction on medical images.
Approach
Uses a structural causal model to decompose the group-specific TTE likelihood into five bias sources (image feature disparity, PMI between image features and outcome/censoring, TTE distribution disparity, censoring rate disparity), links these to fairness gaps, and benchmarks fairness algorithms adapted to TTE prediction under in- and out-of-distribution shifts.
Key results
- Fairness algorithms reduce group gaps in settings with substantial bias sources (e.g., ADNI age, AREDS race, MIMIC-CXR age) but can worsen gaps when bias is small, and no method consistently outperforms DeepHit.
- Pre-training improves predictive accuracy, especially on ADNI, but p-values were above 0.05 in 18 of 24 settings for fairness improvement.
- Under distribution shift, label noise mainly degrades IBS, censoring flips degrade ranking metrics (C-td, AUC-td), and image noise degrades all metrics; fairness effects are mixed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use pre-training instead of training from scratch | TTE prediction models · ADNI and other evaluated dataset/attribute settings | predictive performance | Helped |
| Use pre-training instead of training from scratch | TTE prediction models · evaluated dataset/attribute settings | fairness | No clear effect |
| Apply fairness algorithms (SR, DI, FRL, DRO, or CSA) | DeepHit · all dataset and sensitive-attribute settings | group performance gap | Mixed |
| Apply fairness algorithms rather than DeepHit alone | DeepHit · all evaluated settings | mitigation of bias | No clear effect |
| Add noise to TTE labels (shift in Y) | TTE prediction models · distribution-shift evaluation | IBS | Hurt |
| Flip censoring indicators (shift in Δ) | TTE prediction models · distribution-shift evaluation | C^{td} | Hurt |
| Add noise to medical images (shift in X) | TTE prediction models · distribution-shift evaluation | predictive performance | Hurt |
| Introduce a distribution shift affecting one group | fair TTE prediction models · distribution-shift evaluation | fairness | Mixed |
| Increase disparities in bias sources across groups | TTE prediction models · AREDS, including race versus sex or age | group performance gap | Hurt |
Surprising findings
- Fairness algorithms often fail to significantly outperform the DeepHit baseline per the Friedman/Nemenyi test.
- Pre-training improves accuracy but has minimal impact on fairness.
Limitations
- Only the standard single-risk setting with non-informative right censoring is considered; competing risks and informative censoring are not addressed.
- Group fairness definition may degrade overall utility or remove clinically meaningful causal pathways.
- Datasets are access-controlled (AREDS, MIMIC, ADNI), and the benchmark covers only three datasets and few sensitive attributes.
Experimental setup
Datasets: AREDS MIMIC-CXR ADNI
Benchmarks: AREDS MIMIC-CXR ADNI
Metrics: time-dependent C-index time-dependent AUC integrated Brier score fairness gap between groups
Baselines: DeepHit PMF Nnet-survival DRO SR FRL DI CSA
Builds on: DeepHit PMF Nnet-survival GroupDRO Wasserstein distance
Closest papers by meaning
- FairMedFM: Fairness Benchmarking for Medical Imaging Foundation Models (NeurIPS 2024 · similarity 0.91)
- FairTune: Optimizing Parameter Efficient Fine Tuning for Fairness in Medical Image Analysis (ICLR 2024 · similarity 0.90)
- FairSeg: A Large-Scale Medical Image Segmentation Dataset for Fairness Learning Using Segment Anything Model with Fair Error-Bound Scaling (ICLR 2024 · similarity 0.89)
- Rethinking Fair Representation Learning for Performance-Sensitive Tasks (ICLR 2025 · similarity 0.87)
- Distribution-aware Fairness Learning in Medical Image Segmentation From A Control-Theoretic Perspective (ICML 2025 · similarity 0.87)
- FairCLIP: Harnessing Fairness in Vision-Language Learning (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.