TL;DR
This study leverages Google Street View (GSV) imagery and computer vision to create objective measures of neighborhood built environments in Utah. These measures are then linked to electronic health records of nearly a million patients, revealing significant associations between neighborhood characteristics and various health outcomes, such as diabetes, hypertension, and obesity. This work highlights the potential of GSV data as a scalable and cost-effective tool for understanding how our surroundings impact our health.
What Problem Does This Paper Address?
Collecting detailed information about neighborhood environments has historically been a significant hurdle in health research. Traditional methods like in-person audits are time-consuming and expensive, while surveys can be subject to biases. This limits our ability to understand how the built environment influences health outcomes on a large scale. The authors aim to bridge this gap by using readily available Google Street View images to create objective, scalable indicators of neighborhood features and assess their association with individual-level health outcomes. This moves beyond previous studies that often relied on broader, less granular data or ecological analyses.
Methods
The core of this research lies in using computer vision, specifically Convolutional Neural Networks (CNNs), to analyze Google Street View images.
- Image Collection: GSV images were collected for all primary and secondary roads in Utah, sampled at 50-meter intervals, with four images taken at each location (north, south, east, west).
- Feature Extraction: A pre-trained VGG-19 CNN model, originally trained on ImageNet, was "fine-tuned" to identify specific built environment features. These features included:
- Building type: Presence of non-single-family detached homes (indicating mixed land use).
- Roads: Single-lane roads (indicating lower urban development).
- Greenness: Street greenness (defined as at least 30% of the image showing trees/landscaping).
- Infrastructure: Presence of crosswalks and visible utility wires.
- Neighborhood Aggregation: Image-level features were aggregated to the census tract level, creating summary statistics (e.g., percentage of images with crosswalks).
- Statistical Analysis: Log Poisson regression models were used to examine the associations between tertiles of these built environment indicators and individual health outcomes, controlling for sociodemographic factors (age, race, ethnicity, etc.) and area deprivation index.
Datasets
- Google Street View (GSV) Images: Approximately 1.4 million GSV images from Utah, collected in November 2019. A subset of 18,700 images were manually labeled for training and validation of the computer vision models.
- Electronic Medical Records (EMRs): Data from Intermountain Healthcare, a large healthcare system in Utah, covering approximately one-third of the state's population (n=1,433,316 initially, with analyses focused on 938,085 urban patients with complete data). This dataset included individual-level health outcomes (diabetes, uncontrolled diabetes, hypertension, obesity, substance use disorders) and sociodemographic information for patients aged 18 and older who had a medical visit between 2017 and 2019.
Key Insights and Ideas
- GSV as a Powerful Health Predictor: The study demonstrates that objective measures derived from GSV images can predict individual health outcomes, even after accounting for individual-level factors. This suggests that our environment plays a substantial role in shaping health.
- "Green" and "Mixed-Use" Environments are Protective: Neighborhoods with more green streets and a higher prevalence of non-single-family homes (indicating mixed land use) were associated with a lower prevalence of chronic conditions like diabetes, hypertension, and obesity. This reinforces the idea that well-maintained, diverse environments promote better health.
- Visible Utility Wires as a Risk Factor: Conversely, the presence of visible utility wires was consistently associated with a higher prevalence of adverse health outcomes, including diabetes, hypertension, and obesity. This could be a proxy for neighborhood disorder or neglect, potentially impacting mental health and discouraging physical activity.
- Crosswalks: A Nuanced Relationship: While crosswalks are typically associated with walkability, this study found a surprising association with increased prevalence of hypertension and substance use disorders in areas with more crosswalks. This highlights the complexity of built environment-health relationships, suggesting other factors (like crime or specific urban contexts) might be at play.
- Scalability of GSV: The methodology offers a scalable and cost-effective way to assess neighborhood environments, overcoming limitations of traditional data collection methods.
Key Results
- Non-single-family homes were associated with:
- 17% lower prevalence of diabetes.
- 14% lower prevalence of uncontrolled diabetes.
- 27% lower prevalence of hypertension.
- 11% lower prevalence of obesity.
- Green streets were associated with:
- 10% lower prevalence of diabetes.
- 11% lower prevalence of uncontrolled diabetes.
- 16% lower prevalence of hypertension.
- 10% lower prevalence of obesity.
- Visible utility wires were associated with:
- 9-10% higher prevalence of diabetes and uncontrolled diabetes.
- 4-5% increase in obesity.
- Increased hypertension and substance use.
- Crosswalks showed a 7-9% increased prevalence of hypertension.
- Computer vision model accuracy for feature detection was high, ranging from 82.35% (non-single-family home) to 97.20% (crosswalks).
Limitations and Future Work
- Data Representativeness: While using data from a large healthcare provider, the study acknowledges potential differences in patient demographics (e.g., slightly higher proportion of females and White individuals) compared to the general Utah population. Future work could incorporate data from multiple providers and states.
- Cross-Sectional Design: The study is cross-sectional, meaning it captures associations at a single point in time. Longitudinal studies are needed to understand how changes in neighborhood environments predict changes in health outcomes over time.
- GSV Image Limitations: Computer vision models can be limited by feature size and may struggle with variations in season, zoom, or angle. Unequal time gaps in GSV updates across geographical areas can also pose challenges.
- Unmeasured Confounders: The study acknowledges that other neighborhood constructs, such as air quality, pollution, and perceived safety, were not included but could influence health.
- Future Directions: The authors suggest exploring historical GSV images to track environmental changes, developing more robust computer vision models to handle image variations, and incorporating data from multiple healthcare systems and states to enhance generalizability.
Summary updated