Spatial Scan for Disease Mapping on a Mobile Population

Liang Lan, Vuk Malbasa, Slobodan VuceticAAAI 2014 · DOI

Abstract

Traditional disease mapping often relies solely on residential data, overlooking risks encountered during daily activities like commuting or working. This paper proposes a novel spatial scan statistic that incorporates human mobility patterns. By modeling disease risk as a function of time spent in specific regions and employing several computational optimizations, the method effectively identifies high-risk clusters in large, mobile populations.

Introduction

In epidemiology, identifying geographic areas with unusually high infection or risk rates—known as "disease mapping"—is crucial for public health interventions. For decades, the spatial scan statistic, particularly Kulldorff’s method, has been the gold standard. However, these traditional methods typically rely on residential addresses, ignoring the significant amount of time individuals spend at workplaces, schools, and transit hubs.

This "residence-only" approach can fail to detect environmental hazards or viral outbreaks tied to non-residential locations. For instance, an outbreak at an airport might go unnoticed if the affected individuals live in disparate parts of a city. This paper introduces a mobile-aware spatial scan statistic designed to leverage abundant smartphone and transit data to identify these hidden risks.

Methodology

The core innovation lies in extending the traditional spatial scan to account for the fraction of time an individual spends in various locations.

The Logistic Model and Likelihood Ratio

The authors model the probability of an individual becoming ill using a logistic function weighted by the time spent inside a candidate high-risk region versus outside of it. A log-likelihood ratio test is then used to compare the null hypothesis (uniform risk) against the alternative hypothesis (elevated risk within a specific sub-region).

Scalability Optimizations

Processing high-dimensional movement data for a large population is computationally intensive. To achieve linear scalability, the authors introduced four key optimizations: sliding windows, sparsity exploitation, discretization (binning), and mathematical pruning.

Impact of discretization on solution accuracy and runtime.

As shown in the figure above, discretization significantly reduces computation time while maintaining a low solution difference, enabling the algorithm to handle massive datasets efficiently.

Experimental Setup

The method was validated using the EpiSims dataset, a high-fidelity synthetic simulation of Portland, Oregon. This dataset simulates the daily activities of over 1.6 million people across 240,000 locations. The authors compared their mobile-aware scan ($D_M$) against a traditional residential scan ($D_R$) to benchmark performance.

Results and Discussion

The mobile-aware scan demonstrated superior performance in identifying clusters tied to non-residential activity.

Scenario Analysis: Industrial Areas and Airports

In a simulated outbreak at a busy industrial site, the mobile-aware scan detected the high-risk region with high statistical significance ($p < 0.01$), whereas the residential scan failed ($p = 0.39$). The most extreme test involved an airport—a location with high foot traffic but almost no residents. The residential scan was completely blind to this risk, while the mobile scan successfully flagged the region.

Detection of high-risk clusters using the mobile-aware spatial scan in Portland.

Computational Efficiency

By combining discretization and sparsity optimizations, the algorithm achieved linear scalability relative to population size and location count. On a $150 \times 150$ grid, the optimized system processed the Portland data in approximately 1.7 hours, making it a viable tool for real-world public health monitoring.

Conclusion

This research highlights that residence is often a "noisy proxy" for disease exposure. By incorporating mobility data, public health officials can detect outbreaks in workplaces and transit hubs that traditional methods would miss. Future work will focus on improving robustness against noisy GPS data and integrating more complex disease progression models.

Summary updated

BibTeX
@inproceedings{lan2014spatial,
  author = {Liang Lan and Vuk Malbasa and Slobodan Vucetic},
  title = {{Spatial Scan for Disease Mapping on a Mobile Population}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2014},
  doi = {10.1609/aaai.v28i1.8765},
}
Figure