K-Survival Means

Traditional clustering algorithms like K-Means are designed to group data based on feature similarity, often ignoring the specific survival outcomes (time-to-event) that are critical in medical and reliability research. This paper asks whether a modified K-Means approach can explicitly incorporate survival data into the clustering objective to produce more meaningful, well-separated patient or system subgroups.

The author introduces K-SurvMeans, an extension of the classic K-Means algorithm. Instead of minimizing distance in feature space, K-SurvMeans uses an objective function based on the log-rank test statistic to maximize the survival differences between clusters. Because this objective is non-differentiable, the author employs Particle Swarm Optimization (PSO) to find the optimal cluster centers. To handle high-dimensional data and improve efficiency, the method is also implemented in a low-dimensional latent space using PCA.

Experiments across several benchmark survival datasets demonstrate that K-SurvMeans consistently achieves higher separation between survival distributions compared to existing deep learning-based survival clustering methods. While deep learning models often identify more clusters, they frequently struggle to maintain statistically significant differences between them. K-SurvMeans, particularly when operating in a latent space, successfully balances the discovery of finer-grained subgroups with the requirement that these groups exhibit distinct survival behaviors.

This approach provides a simpler, more interpretable alternative to complex deep learning models for survival-based stratification. By focusing on survival outcomes during the clustering process, researchers can identify patient or system subgroups that are more clinically or operationally relevant for risk assessment and personalized intervention planning.