Detecting Spatial Patterns of Disease in Large Collections of Electronic Medical Records Using Neighbor-Based Bootstrapping (NB2)

Maria T. Patterson, Robert L. Grossman

Abstract

We introduce a method called neighbor-based bootstrapping (NB2) that can be used to quantify the geospatial variation of a variable. We applied this method to an analysis of the incidence rates of disease from electronic medical record data (ICD-9 codes) for approximately 100 million individuals in the US over a period of 8 years. We considered the incidence rate of disease in each county and its geospatially contiguous neighbors and rank ordered diseases in terms of their degree of geospatial variation as quantified by the NB2 method. We show that this method yields results in good agreement with established methods for detecting spatial autocorrelation (Moran's I method and kriging). Moreover, the NB2 method can be tuned to identify both large area and small area geospatial variations. This method also applies more generally in any parameter space that can be partitioned to consist of regions and their neighbors.

Introduction

As electronic medical records (EMR) and geocoded health data explode in volume, researchers need efficient ways to identify spatial autocorrelation—the tendency for nearby areas to have similar disease rates. Traditional methods like Moran’s I or kriging are statistically robust but can be computationally intensive or rigid when applied to massive datasets. Specifically, it is often difficult to distinguish between different scales of spatial influence: is a disease concentrated in a few tight clusters (like a localized environmental toxin), or does it follow a broad regional gradient (like socioeconomic trends)? This paper introduces Neighbor-Based Bootstrapping (NB2), a flexible Monte Carlo method designed to handle hundreds of millions of records while providing flexibility in the types of patterns it detects.

Methodology

The NB2 Algorithm

The core of the NB2 method is a resampling (bootstrapping) approach. For any given county, the algorithm estimates the disease rate by comparing its neighbors to a random selection of counties. The authors developed two distinct implementations to capture different spatial scales:

  1. The t-test implementation: This uses a paired Student’s t-test to compare prediction errors. It tends to favor tight, high-peaked clusters, making it ideal for identifying small-area variations.
  2. The log odds implementation: This counts how often the neighbor-based estimate is closer to the truth than the random one. This version is more sensitive to broad, smooth regional gradients, or large-area variations.

Dataset and Normalization

The study utilized a massive, de-identified dataset from the Truven Health MarketScan Commercial Claims and Encounters Database, covering approximately 100 million unique individuals in the US from 2003 to 2010. The analysis mapped 1.3 billion diagnostic records to 3,109 counties in the continental US. The researchers focused on 548 "3-digit" ICD-9 disease codes that had sufficient data across at least two-thirds of US counties, with all rates adjusted for age and gender using 2010 Census standards.

Results and Validation

Comparison with Traditional Metrics

NB2 results showed strong agreement with the gold-standard Moran’s I statistic, validating its effectiveness. However, NB2 provided more nuance regarding the scale of the patterns. For instance, the t-test version identified top diseases with an average spatial range of 495 km, whereas Moran’s I favored larger patterns with an average range of 625 km. Hypertensive heart disease (402) and Essential hypertension (401) consistently ranked as the most spatially autocorrelated, representing broad regional trends.

Comparison of NB2 t-test (left) and log odds (right) statistics against Moran's I.

Computational Efficiency

A key advantage of NB2 is its scalability. While 1,000 bootstrap repetitions provide high precision, the authors found that as few as 10 repetitions yielded nearly identical rank-orderings of diseases. Running the full analysis for 548 diseases with 100 bootstraps took about 3 minutes per disease on a single core, proving its viability for "big data" contexts.

Performance metrics for varying bootstrap counts (M).

Discussion

The authors categorized the discovered spatial patterns into three primary drivers:

  • Physical Environment: e.g., Lyme disease (ICD-9 088), which clusters tightly where specific ticks live.
  • Structural/Regulatory: e.g., Legally induced abortion (ICD-9 635), which shows sharp "cliffs" at state borders due to differing laws.
  • Socioeconomic: e.g., Hypertensive heart disease (ICD-9 402), which forms a broad "crescent" across the Southern US.

Conclusion

While NB2 is a powerful descriptive tool, it has limitations. The MarketScan data excludes the uninsured and the elderly, potentially underrepresenting certain disease patterns. Furthermore, NB2 identifies that a pattern exists but does not explain the underlying causality. Future work could involve applying NB2 to non-geospatial parameter spaces, such as social networks or genetic similarity, to identify non-physical clusters of disease.

Summary updated

Abstract

We introduce a method called neighbor-based bootstrapping (NB2) that can be used to quantify the geospatial variation of a variable. We applied this method to an analysis of the incidence rates of disease from electronic medical record data (ICD-9 codes) for approximately 100 million individuals in the US over a period of 8 years. We considered the incidence rate of disease in each county and its geospatially contiguous neighbors and rank ordered diseases in terms of their degree of geospatial variation as quantified by the NB2 method. We show that this method yields results in good agreement with established methods for detecting spatial autocorrelation (Moran's I method and kriging). Moreover, the NB2 method can be tuned to identify both large area and small area geospatial variations. This method also applies more generally in any parameter space that can be partitioned to consist of regions and their neighbors.
BibTeX
@article{patterson2017detecting,
  author = {Maria T. Patterson and Robert L. Grossman},
  title = {{Detecting Spatial Patterns of Disease in Large Collections of Electronic Medical Records Using Neighbor-Based Bootstrapping (NB2)}},
  journal = {Big Data},
  year = {2017},
  doi = {10.1089/big.2017.0028},
}
Figure