Survival analysis in healthcare can be plagued by demographic imbalances due to fairness-oblivious model assumptions or training data biases, especially in clinically consequential domains, such as allogeneic hematopoietic cell transplantation (HCT). To address this problem, we introduce Fair Survival Analysis Transformer (FairSurvTrans), a transformer model for survival forecasting that draws ideas from the design of a large language model (LLM). The model applies positional encoding and multi-head self-attention mechanisms to structured clinical data while incorporating a dynamic fairness-conscious loss function that penalizes performance inequality across the racial subgroups. When trained using synthetically balanced HCT data, FairSurvTrans demonstrated superior predictive performance with an overall C-index value of 0. 6738 0. 0031, while achieving the lowest fairness penalty (0. 0002 0. 0001) among benchmark models such as CoxPH, DeepSurv, and XGBoost-AFT. The model also made well-calibrated risk predictions for 1-year survival while maintaining an accuracy consistent across the six predefined racial groups. These results demonstrate the utility of transformer models inspired by LLM architectures in equitably facilitating interpretable, clinically actionable survival modelling, enabling more responsible deployment of AI in healthcare.
Kalmani et al. (Tue,) studied this question.