Time-to-event regression using partially monotonic neural networks

David Rindt, Robert Hu, David Steinsaltz, Dino Sejdinovic

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 that, in contrast to these scores, 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 the 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: The Problem with Improper Scoring Rules

Survival analysis, or time-to-event regression, is central to fields ranging from medicine to engineering. A critical challenge in this domain is right-censoring, where the event of interest is not observed for all subjects within the study period. To evaluate models under these conditions, researchers frequently employ metrics such as time-dependent concordance ($C^{td}$), survival-CRPS, and the integrated Brier score.

However, this paper demonstrates a fundamental theoretical flaw: these popular metrics are not proper scoring rules for right-censored data. A scoring rule is proper if and only if the true data distribution achieves the optimal expected score. As illustrated in Figure 1, an improper score like $C^{td}$ can actually assign a better (higher) value to an incorrect distribution than to the true one.

Figure 1: Comparison of True and False distributions under the improper time-dependent concordance score. The false distribution achieves a higher score (0.749) than the true distribution (0.541).

In contrast, the authors prove that the right-censored log-likelihood is a proper scoring rule. Despite its theoretical superiority, many modern neural network models avoid direct optimization of log-likelihood due to computational intractability or the need for complex numerical integration.

Methods: SuMo-net

To address the limitations of improper metrics and the complexity of existing likelihood-based models, the authors introduce SurvivalMonotonic-net (SuMo-net). SuMo-net is a partially monotonic neural network designed to model the survival function $\hat{S}(t|x)$ directly.

Figure 4: The SuMo-net architecture. By constraining weights associated with time ($t$) to be non-negative ($w_{ij} \in \mathbb{R}^+$), the network ensures a monotonic survival function.

The architecture (Figure 4) consists of two main components:

  1. Covariate Processing: A standard feed-forward network processes the input features $x$.
  2. Monotonic Time Integration: The processed features and time $t$ are fed into a second network. By enforcing non-negative weights for all paths originating from $t$, the model guarantees that $\hat{S}(t|x)$ is monotonically non-increasing with respect to time.

This design allows for the direct calculation of the probability density function $\hat{f}(t|x) = -\frac{\partial}{\partial t} \hat{S}(t|x)$ via automatic differentiation. Consequently, SuMo-net can optimize the right-censored log-likelihood directly without resorting to proxy metrics or expensive ODE solvers.

Results and Scalability

The researchers evaluated SuMo-net against state-of-the-art (SOTA) methods on several datasets, including SUPPORT, METABRIC, and the massive KKBOX churn dataset (over 2.6 million observations).

Key findings include:

  • Superior Accuracy: SuMo-net consistently achieved the best log-likelihood scores across all benchmarks.
  • Computational Efficiency: SuMo-net provides a 20–100x speedup in inference compared to methods like Cox-Time or SODEN. This efficiency stems from its ability to compute survival probabilities in a single forward pass.
  • Scalability: The model handles millions of observations with ease, whereas many competing neural survival models struggle with memory or time constraints at that scale.

Conclusion

The proof that common survival metrics are improper highlights a significant risk in model selection for time-to-event data. By introducing SuMo-net, the authors provide a theoretically sound and computationally efficient alternative. SuMo-net’s use of monotonic constraints enables direct optimization of the proper log-likelihood score, leading to more accurate and scalable survival modeling.

Summary updated

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 that, in contrast to these scores, 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 the 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--100$\times$ computational speedup on inference over existing state-of-the-art neural methods, and is readily applicable to datasets with several million observations.
BibTeX
@inproceedings{rindt2022time,
  author = {David Rindt and Robert Hu and David Steinsaltz and Dino Sejdinovic},
  title = {{Time-to-event regression using partially monotonic neural networks}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2022},
  url = {https://openreview.net/forum?id=KBWUUhVfT4},
}
Figure