Towards Robust Time-to-Event Prediction: Integrating the Variational Information Bottleneck with Neural Survival Model

Armand Bandiang Massoua, Abdoulaye Baniré Diallo, Mohamed Bouguessa

Abstract: Survival analysis aims to predict the time until a specific event of interest occurs. Although neural network-based survival models perform well in extracting rich feature embeddings and outperform traditional models, they are susceptible to the intricacies of noise present in real-world data. This noise can cause these models to miss crucial information for event-time prediction while introducing irrelevant information into the feature embeddings. Furthermore, models may struggle to distinguish between relevant and irrelevant information in data-limited regimes, such as healthcare. This can lead to overfitting, resulting from spurious correlations between irrelevant information and survival outcomes. To address these problems, we introduce the Variational Information Bottleneck (VIB) regularization approach. VIB is designed to meticulously filter out both irrelevant and redundant information, resulting in more robust feature embeddings for event-time prediction. We conducted detailed experiments on several real-world survival datasets. Our approach outperforms state-of-the-art methods in event-time prediction in various evaluation metrics. Furthermore, evaluations on semi-synthetic noisy dataset demonstrate the superior noise resistance of our approach, showcasing improved generalization and robustness.