Geographical clustering of cancer incidence by means of Bayesian networks and conditional Gaussian networks

José M. Peña, I. Izarzugaza, E. Aldasoro, Pedro LarrañagaAISTATS 2001

This paper presents a sophisticated approach to public health mapping, using probabilistic graphical models to uncover geographical patterns in cancer data. By applying Bayesian and Conditional Gaussian Networks to cancer incidence rates in the Basque Country, the authors demonstrate how unsupervised learning can identify distinct regional "health profiles" that correlate with industrialization and population density.

The Challenge of Geographical Health Mapping

Public health officials often need to understand how diseases like cancer are distributed geographically to allocate resources or investigate environmental risks. However, raw data is often "noisy" and lacks clear labels—we don't inherently know which towns belong together in a high-risk or low-risk group.

This is a classic data clustering problem with a twist: the "cluster" is a hidden variable we must infer from observed data (cancer rates). The challenge lies in finding a model that can handle the complex dependencies between different types of cancer while accurately grouping towns that share similar risk profiles.

Methods: Probabilistic Graphical Models

The authors compared two types of probabilistic graphical models to tackle this unsupervised learning task:

  1. Bayesian Networks (BNs): Used for discrete data. The cancer incidence rates were "binned" into four categories.
  2. Conditional Gaussian Networks (CGNs): Used for continuous data, assuming that incidence rates follow a normal (Gaussian) distribution within each cluster.

To "learn" these models from data where the cluster labels are missing, they employed two advanced algorithms: Bayesian Structural EM (BS-EM), an iterative process that searches for the best network structure while simultaneously estimating the missing cluster assignments, and BS-BC+EM, an improved version for discrete data that uses the "Bound and Collapse" method to speed up and stabilize the learning process.

Essentially, the algorithms look for a structure where, once you know which cluster a town belongs to, the incidence rates of different cancers become mostly independent—a sign of a clean, logical grouping.

Datasets and Study Area

The study utilized data from the cancer registry of the Autonomous Community of the Basque Country (ACBC) in Spain, covering the years 1986 to 1994. The dataset included 231 towns for the male database and 221 for the female database, filtered for towns with sufficient population to provide reliable estimates.

The study area: The Autonomous Community of the Basque Country in Spain, highlighting major cities like Bilbao, San Sebastián, and Vitoria.

Features consisted of age-standardized cancer incidence rates for the six most frequent cancer types for each sex (e.g., lung, prostate, and bladder for men; breast, colon, and uterus for women). Rates were calculated per 100,000 inhabitants and adjusted to a "standard" population to ensure fair comparisons between towns with different age demographics.

Key Results and Geographical Consistency

Across almost all models, a clear pattern emerged. The "Cluster 1" group consistently mapped to the most densely populated and industrial towns. These areas showed higher incidence rates for nearly all six major cancer types compared to the rural "Cluster 0."

Geographical clustering results for discrete and continuous models with varying cluster sizes (k=2 and k=3).

Major industrial cities like San Sebastián and Bilbao consistently fell into the high-incidence cluster. Vitoria, the third major city, showed a more unique profile, sometimes shifting clusters depending on the model. When the model was forced to find three clusters ($k=3$) instead of two, it further subdivided the industrial group, identifying specific "hotspots" with even higher incidence rates.

Interestingly, the learned networks showed very few links between different cancer types. This suggests that the "cluster variable" captures the bulk of the shared risk; once you know a town is in the "industrial cluster," knowing its lung cancer rate doesn't tell you much more about its prostate cancer rate.

Discrete vs. Continuous Modeling

The authors found that discrete Bayesian Networks (BNs) were more realistic than continuous Conditional Gaussian Networks (CGNs). The CGNs' strict assumption that cancer rates follow a perfect bell curve (normality) was often too rigid for real-world health data, leading to less intuitive results. The discrete BNs provided a more robust characterization, successfully identifying that the most populated areas exhibited higher cancer rates in 3 out of 6 cancer types for women and nearly all types for men.

Limitations and Future Directions

The authors identify several hurdles for future research:

  • Data Sparsity: Many towns in the region have small populations. Even with age standardization, a few cases in a tiny village can create a statistical "spike" that doesn't reflect a long-term trend.
  • The Normality Trap: The failure of the CGNs suggests that future work should explore non-Gaussian continuous models or more sophisticated discretization techniques.
  • Causality vs. Correlation: While the link between industrial areas and higher cancer rates is suggestive, these models identify clusters, not causes. Future research would need to integrate environmental data (like air quality or soil contaminants) directly into the Bayesian Network to move from description to explanation.

Summary updated

BibTeX
@inproceedings{pena2001geographical,
  author = {José M. Peña and I. Izarzugaza and José Antonio Lozano and E. Aldasoro and Pedro Larrañaga},
  title = {{Geographical clustering of cancer incidence by means of Bayesian networks and conditional Gaussian networks}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2001},
}
Figure