Deep Discrete-Time Survival Analysis with Guaranteed Monotonicity

Abstract:

Discrete-time neural survival models trained with binary cross-entropy are attractive due
to their simplicity. However, they can produce invalid patient-specific survival curves that
increase over time when survival probabilities at different time points are learned without
structural constraints. We propose Kaplan–Meier Net (KMNet), a discrete-time neural sur
vival model that predicts interval-wise conditional survival probabilities and constructs the
survival curve through a Kaplan–Meier style product, guaranteeing non-increasing survival
predictions by design. KMNet is trained with a censoring-aware weighted binary cross
entropy objective and is further augmented with a smooth ranking term that compares
individuals using the conditional survival probability at the event interval of the anchor
observation, which differs from the global ranking losses used in existing deep survival mod
els. We evaluate KMNet on eight benchmark datasets and compare it with seven strong
neural baselines. Across datasets, KMNet achieves the best overall average rank in both
time-dependent concordance and integrated brier score, while consistently producing valid
survival curves.

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Bo_Dai1

Submission Number: 8625