FairMoE-Surv: A Fairness-Gated Neuro-Symbolic Mixture-of-Experts for Equitable Clinical Risk Prediction with Cross-Cohort Validation

Abstract—Fairness-aware clinical risk models are often evaluated using aggregate metrics that may obscure disparities across
protected groups and provide limited evidence of generalizability.
We address both issues with FairMoE-Surv, a neuro-symbolic,
fairness-gated mixture-of-experts (MoE) architecture that combines a hybrid learned/symbolic router, clinically grouped mechanism tokens, momentum-updated population-prototype attention,
a top-2-of-6 MoE layer, demographic-parity and proxy-feature
consistency regularizers, and a calibration module. On N =
28,800 real, de-identified CIBMTR allogeneic hematopoietic-celltransplantation records spanning six self-reported race groups,
FairMoE-Surv achieves 89.2% AUROC, 88.9% accuracy, 88.4%
F1, 89.1% precision, 87.8% recall, and 88.7% concordance
index, outperforming CoxPH (83.8% AUROC), Random Survival
Forest (85.2%), DeepSurv (86.1%), DeepHit (84.8%), and MTLR
(86.9%). It also attains the lowest demographic-parity gap (9.6%
vs. 14.1–20.1%), equal-opportunity gap (9.1% vs. 15.3–23.6%),
and expected calibration error (2.8% vs. 4.1–6.7%) across all
six race groups. A five-component leave-one-out ablation shows
that every component contributes: removing the MoE layer
reduces AUROC to 84.6%, while removing the calibration module
reduces it to 86.9% but increases ECE to 3.5%. Retraining the
identical architecture on a disease-mismatched SEER breastcancer cohort (N = 4,024, three race groups) preserves FairMoESurv’s superiority but narrows its fairness margin from 27–46%
to 7–9%. These findings support cross-cohort retraining as
essential for validating fairness generalization.