Bridging the gap: improve neural survival models with interpolation techniques

Survival analysis is an essential tool in healthcare for risk assessment, assisting clinicians in their evaluation and decision making processes. Therefore, the importance of using expressive and high-performing survival models is crucial. With the advent of neural networks and deep learning, a new generation of survival models has emerged, offering state-of-the-art capabilities to capture the non-linear and complex relationships inherent in multimodal patient data for survival prediction. However, these models often produce discrete outputs, resulting in survival functions that are coarse-grained and difficult to interpret. This study advances previous research by further exploring interpolation techniques as a post-processing strategy to improve the predictive accuracy of survival models. Our results show how the use of specific interpolation techniques significantly improves the concordance and calibration of survival estimates. This analysis encompasses a wide array of medical datasets, models, and interpolation techniques, demonstrating the effectiveness of the proposed approach and providing actionable insights for survival model design.