P1-84 Hidden markov random field for spatial and spatial-temporal risk mapping

Lamiae Azizi, Florence Forbes, Myriam Charras-Garrido, David AbrialJournal of Epidemiology & Community Health 2011 · DOI

Introduction

Risk mapping is a fundamental tool in epidemiology, allowing researchers to identify geographic regions with significantly high or low disease prevalence. Understanding these spatial patterns is crucial for uncovering the underlying mechanisms of disease spread and for the efficient allocation of public health resources. Traditionally, risk mapping has been treated as a two-step process: first, the risk for each area is estimated, and subsequently, a separate classification or clustering step is applied to define "hotspots."

However, this disjointed approach has significant drawbacks. It often fails to account for the spatial correlation inherent in epidemiological data—where the risk in one region is naturally related to its neighbors—and can be highly sensitive to noise. This research proposes a unified framework that integrates risk estimation and classification into a single, automated step using Hidden Markov Random Fields (HMRF).

Methodology

The core of the proposed method is a Finite Mixture Model where the allocation of data points to specific risk categories is governed by a Potts model. This approach treats the "true" risk level of a region as a hidden (latent) state that influences the observed disease data.

Key Technical Components

  • Hidden Markov Random Field (HMRF): By modeling the risk as a latent state, the framework can better represent the uncertainty in observed data.
  • Potts Model: This spatial correlation process acts as a "smoothing" mechanism. It ensures that neighboring regions are more likely to share the same risk classification, reflecting the geographical reality that diseases rarely respect arbitrary administrative boundaries.
  • Mean-Field Theory Approximation: Exact inference on HMRFs is often computationally prohibitive (NP-hard). To address this, the authors utilize a mean-field approximation within an Expectation-Maximization (EM) algorithm. This allows for efficient parameter estimation and simultaneous risk classification.

Overcoming Initialization Sensitivity

A common challenge with EM algorithms is their sensitivity to initial parameters, which often leads to the algorithm getting stuck in local optima. This is particularly problematic in cases of rare or non-contagious diseases where the signal-to-noise ratio is extremely low.

The authors introduce a robust initialization strategy specifically designed to overcome this sensitivity. By providing a more reliable starting point, the model can accurately identify risk zones even when the disease prevalence is very small, making it a powerful tool for studying rare conditions.

Applications and Results

The methodology was validated using both simulated data and real-world epidemiological records:

  1. Bovine Spongiform Encephalopathy (BSE) in France: The model was applied to the "Mad Cow Disease" dataset. BSE presents a significant challenge because it is non-contagious and has a very low prevalence rate. The framework successfully identified risk patterns that traditional models might miss, demonstrating its sensitivity in low-signal environments.
  2. Spatio-Temporal Extension: The research also introduces an extension of the HMRF model to the spatio-temporal context. By accounting for temporal dependencies alongside spatial ones, the model provides more comprehensive cues regarding how risk evolves over time. This extension was illustrated through simulated data, showing improved accuracy over static spatial models.

Conclusion

This HMRF-based framework represents a significant advancement in epidemiological risk mapping. By automating the classification process and incorporating spatial-temporal dependencies, it provides a more accurate and robust representation of disease dynamics. The unified approach not only simplifies the workflow for epidemiologists but also enhances the ability to detect emerging threats in scenarios where data is sparse or the disease signal is weak. Future work may further explore the computational scaling of this model for global-scale datasets and its application to a wider variety of contagious and non-contagious diseases.

Summary updated

Abstract

Risk mapping in epidemiology enables the epidemiologist to identify regions with high or low risk of contamination and understand the underlying mechanisms of the spread of the disease. In this work we are presenting a method of risk mapping based on finite mixture models, in which the allocation to the mixture components is modelled through a correlated process, the Potts model. The inference is performed using an approximation of the Expectation Maximisation (EM) algorithm based on the mean-field theory. One advantage of this model is that the classification of the risk is done automatically and not performed in a second step as in current risk mapping. Methods We are presenting also a way of initialisation able to overcome the sensitivity of this algorithm to its initial parameters. Combining the proposed model to this way of initialisation is leading to good results even in the case of animal non contagious diseases, in which the risk level is very small. This is illustrated in both simulated data and real data: The bovine spongiform Encephalopathy disease in France. We will also introduce an extension of this model to the spatial-temporal context since taking in account the temporal dependencies besides the spatial ones usually provides more useful cues. This methodology will be illustrated on simulated data.
BibTeX
@article{azizi2011p,
  author = {Lamiae Azizi and Florence Forbes and Myriam Charras-Garrido and David Abrial},
  title = {{P1-84 Hidden markov random field for spatial and spatial-temporal risk mapping}},
  journal = {Journal of Epidemiology \& Community Health},
  year = {2011},
  doi = {10.1136/jech.2011.142976c.77},
}
Figure