Abstract
Survival analysis often faces a trade-off between the simplicity of linear models like Cox Proportional Hazards (CPH) and the predictive power of "black-box" deep learning. This paper proposes a Mixture of Cox Experts (MoCE) framework that bridges this gap. By utilizing a set of linear experts coordinated by a gating function and optimized via a Variational Lower Bound (ELBO), the model achieves state-of-the-art performance while maintaining clinical interpretability.
Introduction
Survival analysis predicts the time until an event of interest occurs, such as patient mortality or equipment failure. While the Cox Proportional Hazards (CPH) model is the industry standard, its linear assumption often fails to capture complex, non-linear relationships in high-dimensional medical data. Deep learning alternatives like DeepSurv offer improved accuracy but lack the transparency required for clinical decision-making. This work investigates whether a semi-parametric Mixture of Experts (MoE) approach can provide the best of both worlds: the performance of deep learning and the interpretability of linear models.
Methodology
The core of the proposed approach is the Mixture of Cox Experts (MoCE). Instead of a single monolithic model, MoCE employs multiple "experts"—each a simple linear Cox model—and a "gating function" that assigns patients to the most appropriate expert based on their features.
The authors introduce several key technical innovations:
- ELBO Objective: To address overfitting and mathematical instability in previous MoCE implementations, the authors derive a Variational Lower Bound (ELBO) of the log-likelihood. This provides a more robust and theoretically grounded objective for training.
- Gating Strategies: The model supports both linear gating for maximum transparency and non-linear gating (via Multi-Layer Perceptrons) for increased flexibility.
- Inference Modes: The framework allows for "Hard-Gating" (selecting the single best expert) or "Soft-Gating" (averaging expert predictions), enabling a choice between absolute interpretability and maximum predictive performance.
Experimental Setup
The model was evaluated on three benchmark clinical datasets:
- METABRIC: 1,980 breast cancer patients with gene expression and clinical data.
- SUPPORT: 9,105 seriously ill hospitalized adults.
- Rotterdam-GBSG: 2,232 patients from combined breast cancer studies.
Results and Discussion
The ELBO-MoCE model consistently outperformed or matched state-of-the-art baselines, including Random Survival Forests and DeepSurv.
Performance and Stability
On the METABRIC dataset, Soft-Gating MoCE achieved a Concordance Index (C-index) of 0.6658, surpassing standard CPH (0.6306) and DeepSurv (0.6434). Crucially, the ELBO-based training proved significantly more stable than previous approximations (like RT-MoCE), which often suffered from performance degradation during training due to overfitting.
Interpretability
Unlike deep neural networks, MoCE allows clinicians to inspect the linear coefficients of the specific expert assigned to a patient subgroup. This reveals which features (e.g., age, tumor size) are the primary drivers of risk for that particular population, providing the transparency required in medical settings.
Conclusion
This research demonstrates that "overly parameterized" deep models are not always necessary for survival analysis. A well-structured semi-parametric architecture like MoCE can deliver superior predictive power without sacrificing the transparency essential for healthcare. Future work will focus on relaxing the constant baseline hazard assumption to allow for individual-specific risk "shapes" over time.
Summary updated