Investigating Trustworthiness of Nonparametric Deep Survival Models for Alzheimer's Disease Progression Analysis

Alzheimer's Dementia (AD) affects millions, and survival analysis models can help predict disease progression. But as deep learning models become more accurate, a new study from Thrasher et al. (arXiv:2605.04063) warns they may also perpetuate bias. The team tested several nonparametric deep survival models on AD datasets and found they often produced unfair predictions for marginalized groups—particularly along lines of sex, race, and education level. Previous work focused on model accuracy but largely ignored fairness, making this one of the first rigorous fairness audits in AD progression analysis.

The researchers introduced two novel fairness metrics: Time-Dependent Concordance Impurity (TDCI) and Kaplan-Meier Fairness (KMF). TDCI measures how a model's ranking of patient risk varies across sensitive groups over time, while KMF compares predicted survival curves against actual outcomes per group. Using these metrics, they demonstrated that standard deep survival models like DeepSurv and Cox-CC exhibit significant bias—up to 15% disparity in some subgroups. The study also analyzed feature importance, showing that age and cognitive test scores dominate predictions, but models still rely on sensitive attributes. This work provides both diagnostic tools and a call to action for developing fairer AI in clinical decision support.