TCSurv: Time-based Clustering for Reliable Survival Analysis

TMLR Paper6841 Authors

Abstract: Survival analysis is critical in healthcare for predicting time-to-event outcomes such as disease progression or patient survival. While deep learning excels at capturing meaningful representations from complex clinical data and has improved performance in deep survival models, it inherently struggles with reliability and robustness, challenges that are especially significant when deploying these models in real-world clinical practice. Out-of-distribution (OOD) detection, designed to identify or flag samples that deviate from the training distribution, has become a key method for evaluating AI reliability across fields. This capability is especially important in clinical applications, where noisy or heterogeneous patient data can lead to incorrect assessments; yet, OOD detection remains underexplored and challenging in deep survival analysis due to the need to handle both censored and observed samples, which are unique to this domain. In this study, we address this critical gap by introducing TCSurv, a novel time-base clustering approach for survival analysis that handles both observed and censored samples for robust OOD detection. TCSurv initializes cluster centers using in-distribution data, creating time-specific clusters that anchor model predictions for both observed and censored samples. Experiments in real-world clinical data, including Alzheimer’s dementia progression, and benchmark medical imaging datasets demonstrate that TCSurv effectively distinguishes OOD samples without compromising survival performance compared to existing deep survival analysis frameworks.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Martin_Mundt1

Submission Number: 6841