Abstract
We consider frequently used scoring rules for right-censored survival regression models such as time-dependent concordance, survival-CRPS, integrated Brier score and integrated binomial log-likelihood, and prove that neither of them is a proper scoring rule. This means that the true survival distribution may be scored worse than incorrect distributions, leading to inaccurate estimation. We prove, in contrast to these scores, that the right-censored log-likelihood is a proper scoring rule, i.e. the highest expected score is achieved by the true distribution. Despite this, modern feed-forward neural-network-based survival regression models are unable to train and validate directly on right-censored log-likelihood, due to its intractability, and resort to the aforementioned alternatives, i.e. non-proper scoring rules. We therefore propose a simple novel survival regression method capable of directly optimizing log-likelihood using a monotonic restriction on the time-dependent weights, coined SurvivalMonotonic-net (SuMo-net). SuMo-net achieves state-of-the-art log-likelihood scores across several datasets with 20–100x computational speedup on inference over existing state-of-the-art neural methods and is readily applicable to datasets with several million observations.
Introduction
Survival analysis is a branch of statistics focused on predicting the time until an event occurs, such as patient recovery or mechanical failure. A defining characteristic of survival data is right-censoring: for many subjects, the event of interest has not occurred by the end of the study period, meaning we only know the event happened after a certain time.
Evaluating survival models is traditionally done using metrics like time-dependent concordance ($C^{td}$), the Brier score, or the Continuous Ranked Probability Score (CRPS). However, this research demonstrates a fundamental flaw in these metrics: they are non-proper scoring rules.
The Problem with Improper Scoring Rules
A scoring rule is "proper" if the highest expected score is achieved by the true probability distribution. If a rule is improper, a model predicting an incorrect distribution can achieve a better score than a model predicting the ground truth. This paper proves that $C^{td}$, survival-CRPS, and integrated Brier scores are all improper.
As shown in the figure above, a "fake" distribution can be constructed to yield a significantly higher concordance score than the true distribution. This implies that optimizing for these metrics can actively lead a model away from the true underlying data distribution. In contrast, the authors prove that the right-censored log-likelihood is a proper scoring rule.
SuMo-net: Direct Log-Likelihood Optimization
While log-likelihood is the ideal objective, it has been difficult to optimize in neural networks because it requires simultaneous access to the survival function $\hat{S}(t|x)$ and its derivative, the density function $\hat{f}(t|x)$.
The authors propose SuMo-net (SurvivalMonotonic-net), which utilizes a monotonic neural network architecture to make this optimization tractable.
SuMo-net enforces the non-increasing property of survival functions by ensuring weights associated with the time input $t$ are non-negative. The network outputs a scalar $h(t, x)$, and the survival function is defined as $\hat{S}(t|x) = 1 - \text{sigmoid}(h(t, x))$. Because the network is continuous and monotonic, automatic differentiation can be used to compute the density function $\hat{f}(t|x) = -\partial \hat{S}/\partial t$, enabling direct optimization of the log-likelihood.
Results and Performance
SuMo-net was evaluated against state-of-the-art methods like Cox-Time, DeepHit, and ODE-based models (SODEN, SurvNODE) across various datasets, including the massive KKBOX dataset (2.6 million observations).
Key findings include:
- Accuracy: SuMo-net consistently achieved the highest log-likelihood scores across all tested real-world and synthetic datasets.
- Efficiency: SuMo-net demonstrated a 20–100× speedup in inference time compared to competing neural methods.
- Scalability: Due to its standard feed-forward architecture, SuMo-net was the only advanced neural method capable of efficiently processing the multi-million row KKBOX dataset.
Conclusion
By identifying the impropriety of common survival metrics and providing a tractable way to optimize the proper right-censored log-likelihood, SuMo-net represents a significant step forward in survival regression. It offers a combination of theoretical soundness, state-of-the-art accuracy, and exceptional computational efficiency.
Summary updated