TLDR: The research introduces Fairness-Aware Survival Modeling (FASM), a novel approach designed to mitigate algorithmic bias in healthcare survival predictions. FASM addresses both intra-group and critical cross-group risk ranking disparities over time, which are often overlooked by traditional models. Applied to SEER breast cancer data, FASM significantly improves fairness in risk predictions for Black and White patients, particularly in the mid-term follow-up, while maintaining predictive performance comparable to fairness-unaware models. This advancement aims to promote more equitable clinical decision-making.
As machine learning models become more common in healthcare, there’s a growing concern that these models might unintentionally worsen existing health inequalities. This is particularly true in survival analysis, where models predict how long patients might live or when an event might occur. Traditional models can sometimes perpetuate biases found in clinical data, leading to unfair predictions for different demographic groups.
A new research paper introduces a solution called Fairness-Aware Survival Modeling (FASM). This innovative approach aims to tackle algorithmic bias in survival predictions, focusing on both how individuals are ranked within their own group and, crucially, how they are ranked compared to individuals in other groups over time. This cross-group ranking fairness is often overlooked but is vital for equitable healthcare decisions.
The Challenge of Bias in Survival Prediction
In fields like oncology, where survival models are used to estimate patient risks over time, existing health disparities can be amplified. For instance, Black women often face delays in screening and treatment for breast cancer, leading to shorter survival times compared to White women. When models are trained on such data, they can learn and reinforce these biases. Additionally, the dynamic nature of survival outcomes, including factors like patient follow-up (censoring), can vary unevenly across populations, further complicating fair model development.
Conventional fairness methods usually focus on ensuring equal performance within specific subgroups. However, they often miss “cross-group ranking disparities.” This means a model might accurately rank patients within, say, Black and White groups separately, but still place high-risk Black patients below lower-risk White patients. Such misranking can have serious implications for treatment prioritization and resource allocation.
Introducing Fairness-Aware Survival Modeling (FASM)
FASM is designed to address these complex challenges. It works by first generating a collection of “nearly-optimal” survival models – models that perform very well but might differ in how they use specific patient characteristics. From this collection, FASM then selects the model that best minimizes bias, using a specially developed Model Selection Index (MSI) that considers multiple fairness metrics.
The framework evaluates fairness using several key metrics:
- Intra-group ranking bias: This looks at disparities in how well the model ranks patients within their own demographic group (e.g., comparing C-index and integrated AUC across groups).
- Cross-group ranking bias: This is a critical innovation, assessing how individuals from one group are ranked relative to those from another. It uses metrics like cross concordance index (xCI) and time-specific cross AUC (xAUC) to ensure that risk rankings are consistent and fair across different subgroups.
By incorporating these metrics, FASM ensures that models not only predict accurately but also prioritize equity in clinical decision-making.
Applying FASM to Breast Cancer Prognosis
The researchers applied FASM to a large dataset of breast cancer patients from the Surveillance, Epidemiology, and End Results (SEER) Program, focusing on Black and White women. Breast cancer was chosen as a representative case due to its well-documented health disparities.
The study compared FASM with two baseline Cox proportional hazards models: one that included race as a variable (“CoxPH”) and one that excluded it (“Under-blindness”).
Key Findings
The results were compelling. FASM demonstrated significantly improved overall fairness. It showed minimal intra-group ranking bias and, importantly, the smallest cross-group bias. This means FASM was much better at ensuring that high-risk patients from one group were not systematically ranked below lower-risk patients from another group.
Crucially, FASM achieved these fairness improvements while maintaining predictive performance comparable to the fairness-unaware models. While FASM had a marginally lower integrated AUC (0.827) compared to CoxPH (0.833), it struck the most balanced trade-off between accuracy and equity.
Time-stratified evaluations showed that FASM maintained stable and low cross-group ranking disparities over a 10-year follow-up period, with the greatest improvements observed during the mid-term. In contrast, the conventional CoxPH model showed significant disparities, especially in the first year, and the “Under-blindness” model (which excluded race) showed worsening disparities over time.
The FASM model also yielded more balanced risk predictions between Black and White patients, reducing the systematic assignment of higher risk scores to Black patients seen in the CoxPH model. While some disparities re-emerged in early and late follow-up, they remained smaller than those observed with other models, highlighting the dynamic nature of these disparities and the need for ongoing systemic interventions alongside algorithmic solutions.
Also Read:
- FnRGNN: A Multi-Level Approach to Fairer Predictions in Graph Neural Networks
- Evaluating Fairness in Graph Neural Networks on Knowledge Graphs: A New Benchmark
Implications for Equitable Healthcare
This research underscores the importance of addressing both intra-group and cross-group ranking biases in clinical risk prediction. FASM offers a practical and generalizable approach for developing survival models that prioritize both accuracy and equity. By enabling the development of clinical decision tools that promote fair access to timely, life-saving care, FASM represents a significant step forward in trustworthy machine learning in healthcare.
For more detailed information, you can read the full research paper here.
