Spatial Clustering for Carolina Breast Cancer Study

Hongqian Niu, Melissa Troester, Didong Li2024 · DOI

TL;DR

This paper introduces Gaussian Process Spatial Clustering (GPSC), a novel algorithm designed to identify meaningful clusters in geospatial data by considering both location and associated characteristics. GPSC outperforms traditional clustering methods on complex datasets, offering a more nuanced understanding of how socioenvironmental factors influence health outcomes, as demonstrated in its application to North Carolina census tract data.

What Problem Does This Paper Address?

Understanding how socioenvironmental factors influence health outcomes, particularly for diseases like cancer, is a critical but challenging task. Geospatial data, which combines location with various characteristics (like demographics, socioeconomic status, and environmental pollution), presents unique difficulties for traditional clustering algorithms. These methods often struggle to simultaneously account for the spatial relationships between locations and the patterns within their associated features. The paper aims to bridge this gap by developing a method that can effectively cluster geographically dispersed data while considering the complex interplay of spatial and non-spatial information, specifically in the context of health disparities.

Methods

The core contribution is the Gaussian Process Spatial Clustering (GPSC) algorithm. GPSC builds upon the flexibility of Gaussian Processes (GPs), which are powerful tools for modeling functions. Unlike standard clustering methods that might treat spatial and covariate information separately or rely on custom distance metrics, GPSC models the underlying functional relationships between covariates and a response variable within each cluster.

Here's a breakdown:

  • Gaussian Processes (GPs): GPs are used as a flexible regression tool to model the relationship between features (covariates) and a response variable. They provide a probabilistic framework for prediction and uncertainty quantification.
  • Functional Clustering: Instead of clustering data points directly based on distance, GPSC clusters locations based on the similarity of the functions that describe the relationship between their covariates and the response. This means that even if two locations are geographically distant, they can be grouped together if they exhibit similar patterns in how their characteristics relate to health outcomes.
  • Iterative Algorithm: GPSC works iteratively. It starts with an initial assignment of data points to clusters, fits a GP model to the data within each cluster, and then reassigns points to the cluster whose GP model best predicts their response. This process repeats until convergence.
  • Theoretical Guarantees: The paper provides theoretical support for GPSC, including proofs of convergence and analysis of its performance under noisy conditions.

Datasets

The primary dataset used for the real-world application is from the Carolina Breast Cancer Study (CBCS). This dataset includes:

  • North Carolina census tracts: Each tract is represented by its spatial coordinates (longitude, latitude).
  • Socioeconomic data: Aggregated from the American Community Survey, reflecting advantage and disadvantage.
  • Environmental pollution data: From the U.S. Environmental Protection Agency (EPA) National Air Toxics Assessment (NATA).

The paper also describes three simulation studies using synthetic data to evaluate GPSC's performance under various conditions (linear and nonlinear functions, noisy responses, and overspecified cluster numbers). These simulations serve as benchmarks for comparing GPSC against other clustering algorithms.

Key Insights and Ideas

  • Beyond Simple Proximity: The most significant insight is that meaningful clusters in health-related geospatial data are not solely defined by geographic proximity. Instead, they are often characterized by similar functional relationships between socioenvironmental factors and health outcomes. GPSC captures this by focusing on the predictive power of these relationships within clusters.
  • Handling Complex Relationships: Traditional methods struggle when the relationship between spatial location and characteristics is complex or when geographically distant areas share similar functional patterns. GPSC's GP-based approach is inherently better equipped to model these intricate relationships.
  • Interpretable Clusters for Health Disparities: By identifying clusters with distinct functional profiles, GPSC can help pinpoint marginalized communities or areas that might be overlooked by simpler clustering methods. This has direct implications for targeted public health interventions.
  • Robustness to Noise and Overspecification: The theoretical analysis and simulations demonstrate that GPSC is robust to measurement noise in the response variable and can handle situations where the number of specified clusters is higher than the true number, a common challenge in real-world applications.

Key Results

  • Improved Clustering Performance: In simulation studies, GPSC consistently outperformed traditional clustering algorithms (K-means, spectral clustering, hierarchical clustering, DBSCAN) and their spatial/supervised variants, particularly in scenarios with nonlinear relationships and noisy data.
  • Finer Detail in Real-World Data: When applied to the North Carolina census tract data, GPSC identified clusters with finer detail than K-means, particularly in less dense areas of the state. This allowed for a more nuanced understanding of socioeconomic and environmental advantage/disadvantage.
  • Interpretability: The GPSC algorithm, when applied to the NC data, produced 3 interpretable clusters representing low, medium, and high socioeconomic and environmental advantage, which were more detailed than a baseline aggregation of 8 latent classes.
  • Theoretical Guarantees: The paper provides theoretical backing for GPSC's convergence and its ability to correctly cluster data under certain conditions, even with noise.

Limitations and Future Work

  • Tuning Parameters: Like most clustering algorithms, GPSC has tuning parameters, most notably the number of clusters (L). The paper acknowledges that choosing L often requires domain expertise and is beyond the scope of the current work.
  • Kernel Choice: While the theoretical guarantees are kernel-independent, the practical performance of GPSC can depend on the choice of the GP kernel. The authors suggest that more nuanced anisotropic or nonstationary kernels might be beneficial for highly heterogeneous datasets, and designing such kernels remains an open problem.
  • Scalability: For very large datasets, the standard GP regression complexity of O(n³) can be a bottleneck. The authors mention that scalable GP methods exist and could be applied to reduce runtime to O(n log n).
  • Multi-output Extension: The current paper focuses on single-output cases (one response variable), but the authors note that extension to multi-output cases is straightforward.
  • Domain Expertise: The authors emphasize that GPSC results should be used as a supplementary tool for domain experts to identify marginalized communities and areas, rather than as a sole decision-making method. Future work could involve integrating domain expertise more directly into the algorithm or developing more sophisticated sensitivity analyses to validate findings across different geographic definitions.

Summary updated

Abstract

In the Carolina Breast Cancer Study (CBCS), clustering census tracts based on spatial location, demographic variables, and socioeconomic status is crucial for understanding how these factors influence health outcomes and cancer risk. This task, known as spatial clustering, involves identifying clusters of similar locations by considering both geographic and characteristic patterns. While standard clustering methods such as K-means, spectral clustering, and hierarchical clustering are well-studied, spatial clustering is less explored due to the inherent differences between spatial domains and their corresponding covariates. In this paper, we introduce a spatial clustering algorithm called Gaussian Process Spatial Clustering (GPSC). GPSC leverages the flexibility of Gaussian Processes to cluster unobserved functions between different domains, extending traditional clustering techniques to effectively handle geospatial data. We provide theoretical guarantees for GPSC's performance and demonstrate its capability to recover true clusters through several empirical studies. Specifically, we identify clusters of census tracts in North Carolina based on socioeconomic and environmental indicators associated with health and cancer risk.
BibTeX
@misc{niu2024spatial,
  author = {Hongqian Niu and Melissa Troester and Didong Li},
  title = {{Spatial Clustering for Carolina Breast Cancer Study}},
  year = {2024},
  doi = {10.1142/9789819807024\_0025},
}
Figure