🤖 machine learning

K-Survival Means

This paper introduces K-SurvMeans, a novel K-Means extension for survival data clustering that optimizes cluster centers to maximize pairwise survival differences using Particle Swarm Optimization and a learned low-dimensional latent space, demonstrating superior performance in separating survival distributions compared to existing deep learning methods.

Original authors: Abdallah Alabdallah

Published 2026-07-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Abdallah Alabdallah

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to sort a massive crowd of people into different teams. Usually, you'd group them by what they look like or what they wear—maybe all the red-shirted people go in one group, and the blue-shirted people in another. But what if the real story isn't about their clothes, but about how long they stay at the party before leaving? In the world of medicine and statistics, this is called survival analysis. It's the art of predicting not just what happens to someone, but when it happens, like how long a patient might live after a diagnosis or how long a machine might run before breaking. The tricky part is that sometimes people leave the party early (the event happens), and sometimes they just walk out before the party ends (the event hasn't happened yet, known as "censored" data). Scientists have long used a classic tool called K-Means to sort people into groups based on their features, but this tool is a bit blind; it sorts by appearance without caring if the groups have different "staying power" at the party. This paper asks a simple but powerful question: Can we teach this sorting tool to care about the time people stay, so the groups we find are truly different in how long they survive?

Enter K-SurvMeans, a new and clever upgrade to the old K-Means algorithm, proposed by Abdallah Alabdallah. Think of the original K-Means as a teacher who sorts students into groups based on how similar their backpacks are. K-SurvMeans, however, is a teacher who sorts students based on how long they are likely to stay in the classroom before the bell rings. Instead of just looking at the features (the backpacks), this new method looks at the survival outcome (the time until the bell) and uses that information to decide where everyone sits. The goal is to make sure that the groups formed are as different as possible from each other in terms of their survival times. If Group A leaves the room very quickly and Group B stays for hours, that's a perfect split. If both groups leave at the same time, the split is useless, even if they look different.

To find these perfect groups, the authors had to solve a tricky math puzzle. The usual way to sort things (like K-Means) uses a smooth, sliding path to find the best answer, but the math for comparing survival times is "bumpy" and doesn't slide smoothly. So, the authors used a different strategy called Particle Swarm Optimization. Imagine a swarm of birds flying around a forest looking for the best berry patch. Each bird (or "particle") represents a possible way to sort the data. They fly around, sharing information about where they found good berries (good groupings) and adjusting their flight paths to find the absolute best spot. In this case, the "berries" are groupings where the survival differences between clusters are huge. The algorithm tests thousands of these "bird swarms" to find the arrangement that maximizes the separation between groups.

The paper also tackles a problem called the "curse of dimensionality," which is like trying to find a needle in a haystack that keeps getting bigger and bigger. When there are too many features to look at, the math gets too messy for the birds to fly efficiently. To fix this, the authors created a version called K-SurvMeans (Latent). They first shrink the data down into a simpler, lower-dimensional space (like compressing a giant, detailed map into a small, easy-to-read sketch) and then let the bird swarm sort the groups there. This makes the search faster and helps find clearer, more distinct groups.

When the authors tested this new method on several real-world datasets (including medical data like FLCHAIN, SUPPORT, METABRIC, and NWTCO), the results were quite promising. They compared K-SurvMeans against the old K-Means, a version of K-Means with the "sketch" compression, and some complex, deep-learning-based methods that try to do the same thing. The findings suggest that K-SurvMeans, especially the "Latent" version, is very good at its job. In many cases, it managed to create groups where 100% of the pairs of clusters showed a statistically significant difference in survival times. For example, on the FLCHAIN dataset, K-SurvMeans (Latent) found 5 distinct groups, and every single pair of those groups was clearly different in how long the patients survived.

In contrast, the deep-learning methods (like SCA and VaDeSC) often found more groups, but those groups weren't always as different from each other. It's like the deep learning models found 15 different teams, but many of them had players who left the party at almost the exact same time, making the teams less useful for understanding risk. The authors also noted that while the standard K-Means (without survival info) sometimes found groups that looked different in survival time, it usually found fewer groups and didn't capture as much variety in the population as the new method did.

The paper concludes that K-SurvMeans is a strong, simple, and effective way to sort survival data. It suggests that by directly optimizing for survival differences, we can get clearer, more meaningful groups than by just looking at features or using overly complex deep learning models. However, the authors are careful to note that this method isn't perfect yet; it can get slow if the data is huge or if you try to find too many groups at once. They also point out that unlike the deep learning models, K-SurvMeans doesn't currently predict individual survival curves for each person, just the groups. But for the specific task of finding distinct, well-separated groups of patients or systems, this new "survival-aware" sorting tool suggests a very effective path forward.

Technical Summary: K-Survival Means (K-SurvMeans)

Problem Statement
Survival analysis is critical for modeling time-to-event data in fields such as healthcare, predictive maintenance, and risk assessment. A fundamental challenge in this domain is patient stratification: identifying subgroups with distinct risk profiles to support personalized interventions. While clustering techniques are natural candidates for discovering such subgroups, conventional algorithms like K-Means operate solely on input feature spaces. They seek groups of observations that are similar in feature space but do not explicitly consider survival outcomes (time-to-event and censoring status) during the optimization process. Consequently, clusters generated by standard methods may exhibit similar survival distributions, limiting their utility for risk stratification. Although deep learning approaches (e.g., SCA, VaDeSC) have been developed to jointly learn latent representations and survival objectives, they often introduce substantial model complexity, require large datasets, and necessitate extensive hyperparameter tuning. Traditional clustering algorithms remain underexplored in the context of survival data despite their simplicity and interpretability.

Methodology
The paper proposes K-SurvMeans, a novel extension of the K-Means algorithm designed specifically for survival data clustering. The core innovation is the explicit incorporation of survival outcomes into the clustering objective function to maximize pairwise survival differences between clusters.

Objective Function: Unlike standard K-Means which minimizes intra-cluster variance, K-SurvMeans defines a loss function based on the log-rank test. For a set of $k$ cluster centers, the algorithm assigns observations to the nearest center and calculates the log-rank statistic ( $\chi^2_{ij}$ ) and p-value ( $p_{ij}$ ) for every pair of clusters. The objective is to maximize the sum of log-rank statistics for pairs that are statistically significant (below a threshold $\alpha$ ). The loss function is formulated as:
$L(C; X, k) = - \left( \sum_{0 \le i < j \le k-1} \mathbb{1}(p_{ij} < \alpha) \right) \left( \sum_{0 \le i < j \le k-1} \chi^2_{ij} \right)$
The goal is to find cluster centers $C^*$ that minimize this loss.
Optimization: Since the objective function involving the log-rank test and indicator functions is non-differentiable, the authors employ Particle Swarm Optimization (PSO). In this framework, each particle represents a candidate solution encoded as the concatenation of all $k$ cluster center coordinates. Particles iteratively update their positions based on personal and global best fitness scores, where fitness is determined by the objective function value.
Dimensionality Reduction: To mitigate the "curse of dimensionality" and improve optimization efficiency, the method is extended to operate in a learned low-dimensional latent space. A dimensionality reduction function $f_\theta$ (e.g., PCA) maps the original high-dimensional input to a lower-dimensional space $z$ . The clustering optimization is then performed on these latent representations. In the current formulation, the encoder is learned independently of the clustering loop (fixed during optimization), though the authors note that joint learning is a potential future direction.

Key Contributions

Algorithmic Extension: The introduction of K-SurvMeans, which adapts the centroid-based K-Means framework to survival analysis by replacing the Euclidean distance objective with a survival-driven log-rank objective.
Optimization Strategy: The application of Particle Swarm Optimization to solve the non-differentiable clustering problem, allowing for direct maximization of survival separation without requiring gradient-based backpropagation.
Latent Space Integration: A framework that combines dimensionality reduction with survival clustering to handle high-dimensional data and reduce the search space complexity.
Benchmarking: A comprehensive evaluation against both traditional baselines (K-Means with/without PCA) and state-of-the-art deep learning methods (SCA and VaDeSC) across multiple public survival datasets.

Results
Experiments were conducted on four publicly available benchmark datasets (FLCHAIN, SUPPORT, METABRIC, NWTCO). The evaluation metric was the percentage of pairwise significant differences (based on the log-rank test) across all cluster pairs.

Performance: K-SurvMeans, particularly the variant operating in the latent space (K-SurvMeans(Latent)), consistently achieved the highest percentage of significant pairwise differences compared to all other methods. It successfully identified clusters with distinct survival distributions.
Cluster Granularity: While deep learning methods (SCA, VaDeSC) often identified a larger number of clusters, these clusters frequently exhibited weaker survival separation (lower percentages of significant differences). In contrast, K-SurvMeans(Latent) balanced the discovery of a larger number of clusters with high survival separation, suggesting it captures finer-grained, clinically meaningful subgroups.
Baseline Comparison: Standard K-Means and K-Means(Latent) achieved relatively high separation percentages but tended to produce fewer clusters, capturing less population heterogeneity than the proposed method.
Visualization: Kaplan-Meier survival curves demonstrated that clusters found by K-SurvMeans(Latent) showed clear separation with limited overlap, outperforming the deep learning baselines in terms of distinct survival profiles.

Significance and Claims
The paper claims that K-SurvMeans offers a compelling alternative to complex deep learning models for survival clustering. By explicitly optimizing for survival separation, the method produces clusters that are more useful for risk stratification and decision-making. The authors emphasize that the method retains the simplicity and interpretability of centroid-based clustering while effectively incorporating survival information.

The authors modestly acknowledge limitations, specifically the method's scalability to very high dimensions and large numbers of clusters, which they address via dimensionality reduction. They also note that, unlike SCA and VaDeSC, K-SurvMeans currently focuses on clustering and does not predict individualized survival curves. Future work is planned to integrate embedding learning directly into the optimization process and to extend the method to predict individualized survival curves. The full implementation is made available on GitHub to facilitate reproducibility.

Technical Summary: K-Survival Means (K-SurvMeans)

More like this