Value-Monotonicity Matters: A Concordance Loss for Deep Survival Prediction

2026-07-18

AI Summary

This study addresses a critical misalignment in deep survival analysis: models are typically trained using likelihood-based losses but evaluated via the concordance index (C-index), leading to unreliable training dynamics and model selection, especially in small-sample, high-censoring oncology datasets. The authors first demonstrate that likelihood loss can decrease while C-index remains stagnant, revealing suboptimal optimization directions. To resolve this, they propose the monotonic Sigmoid Concordance Loss (SCL), which closely approximates \(1 - \text{C-index}\) and guarantees that loss reduction strictly corresponds to improved ranking performance. SCL is end-to-end differentiable and reduces to ranking SVM in the linear case. Across 18 multimodal datasets, SCL achieves C-index scores comparable to or better than likelihood-based methods, with training loss exhibiting rank correlations of 0.96–0.99 with C-index—dramatically outperforming the weak or even negative correlations (−0.03 to 0.53) observed with likelihood loss.

Abstract

Deep survival models are evaluated almost exclusively by the concordance index (C-index), yet they are commonly trained using likelihood objectives such as the Cox partial likelihood, discrete-time negative log-likelihood, and DeepHit likelihood. This mismatch is usually considered acceptable because the C-index can be recomputed on validation data during training. However, for end-to-end training of high-capacity encoders on small, heavily censored oncology cohorts, frequent C-index evaluation is computationally expensive, making the loss value itself an important signal for monitoring, early stopping, and model selection. We show that likelihood losses are unreliable for this purpose and propose a value-monotone concordance loss. We prove that every strictly proper survival likelihood admits directions where the loss decreases while the C-index remains unchanged, causing the loss value to decouple from ranking performance. We then study a sigmoid concordance loss (SCL), whose value approximates one minus the C-index up to a temperature term, ensuring that lower loss corresponds to higher C-index during optimization. The loss is architecture agnostic and reduces to a convex survival ranking support vector machine for linear models. Across eighteen datasets from four modalities using a unified five-fold cross-validation protocol, SCL achieves discrimination comparable to standard likelihood losses and is the best or within one standard deviation of the best C-index. Unlike likelihood losses, SCL maintains a strong correlation between loss value and C-index during training, with rank correlations of 0.96 to 0.99 compared with -0.03 to 0.53 for likelihood losses. Calibration measured by the integrated Brier score is comparable. SCL provides a value-monotone optimization objective whose value can serve as a reliable surrogate for the C-index during expensive end-to-end training.

Problem

• concordance index
• survival prediction
• value-monotonicity
• likelihood loss
• deep learning

Innovation

• value-monotonicity
• concordance loss
• deep survival analysis
• C-index
• survival prediction