A Novel Approach to Characterize State-level Food Environment and Predict Obesity Rate Using Social Media Data: Correlational Study

Chuqin Li, Alexis Jordan, Jun Song, Yaorong Ge, Albert ParkJournal of Medical Internet Research 2022 · DOI

TL;DR

This paper introduces a novel method to characterize state-level food environments and predict obesity rates by leveraging publicly available social media data (Yelp and MyFitnessPal). The approach successfully correlates food environment features derived from social media with actual obesity rates, offering a dynamic and cost-effective alternative to traditional survey methods.

What Problem Does This Paper Address?

The obesity epidemic is a critical public health challenge, with environmental factors playing a significant role. Understanding the "food environment" – the availability, accessibility, affordability, and acceptability of food – is crucial for addressing obesity. Traditional methods for measuring food environments, such as surveys and geographic information systems (GIS), are often costly, labor-intensive, and suffer from delays in data collection and reporting. This paper addresses the need for a more dynamic, scalable, and cost-effective way to characterize food environments and their relationship with obesity rates. The authors propose using social media data, which is organic, continuously updated, and readily available, to fill this gap.

Methods

The researchers developed a multi-step approach to quantify the food environment and predict obesity rates:

  1. Food Category Availability: They used Yelp's API to collect data on over 350,000 businesses across the US. By analyzing business categories, they determined the availability of different food types within each state.
  2. Caloric Density Estimation: To quantify the nutritional aspect of the food environment, they combined Yelp data with MyFitnessPal (MFP) data. They identified the most popular dishes for each food category from Yelp and then used MFP's extensive database to find the caloric information (calories per gram) for these dishes. This allowed them to calculate the average caloric density for each food category.
  3. Weighted Caloric Density Score: The caloric density of each category was then weighted by its availability in a given state to create a "weighted score for caloric density" for each state.
  4. Affordability and Acceptability Metrics: They also incorporated affordability (using Yelp's price categories) and acceptability (using average ratings and number of reviews from Yelp) as features.
  5. Predictive Modeling: Finally, they used these derived features (category availability, affordability/acceptability, and weighted caloric density) to build predictive models for state-level obesity rates. They employed machine learning models like random forest, support vector machine, and XGBoost, using leave-one-out cross-validation for evaluation.

Datasets

The study utilized three primary data sources:

  • Yelp: A crowd-sourced review platform used to gather information on over 353,000 businesses, including their categories, ratings, price levels, and reviews. This data was collected via the Yelp API and web scraping.
  • MyFitnessPal (MFP): A popular calorie-tracking app used to obtain nutritional information (specifically calories per gram) for popular dishes. The researchers collected nutrition records for over 37,000 dishes.
  • Behavioral Risk Factor Surveillance System (BRFSS): This US state-based health surveillance system provided the ground truth for state-level obesity rates. Data for 49 states and the District of Columbia from 2019 were used.

Key Insights and Ideas

  • Social Media as a Food Environment Proxy: The core idea is that social media platforms like Yelp, with their vast amount of user-generated content about businesses and food, can serve as a powerful and dynamic proxy for understanding the local food environment.
  • Caloric Density Matters: The study highlights that not just the availability of food, but its caloric density is significantly correlated with obesity rates. This suggests that the nutritional quality of readily available food is a key factor.
  • Disparity in Food Environments: A clear difference in food category availability was observed between states with low (Colorado) and high (Mississippi) obesity rates, with "fast food" being significantly more prevalent in the high-obesity state.
  • Affordability and Acceptability Trends: Fast-food restaurants, while generally cheaper, also tended to have lower ratings and fewer reviews compared to other restaurant types, suggesting a complex interplay of factors beyond just price.
  • Predictive Power of Social Media Data: The research demonstrates that social media data, when combined with nutritional information, can effectively predict state-level obesity rates, outperforming previous studies that relied solely on Twitter data.

Key Results

  • The study found a statistically significant correlation between the characterized food environment (using social media data) and state obesity rates.
  • A model incorporating all three feature sets (category availability, affordability/acceptability, and weighted caloric density) achieved the best performance, with a Pearson correlation of 0.796 between the predicted and actual state obesity rates.
  • The availability of "Fast Food" was found to be positively correlated with state obesity rates (Pearson correlation of 0.676).
  • The weighted score for caloric density also showed a strong positive correlation (r=0.671; P<.001) with state-level obesity rates.

Limitations and Future Work

  • Data Collection Limitations:
    • Yelp API queries are limited to 1000 results per query, potentially leading to missing data in densely populated areas.
  • MFP data may contain inaccuracies due to user-submitted information and partially matching dishes.
  • Actual Consumption Not Captured: The study infers food environment characteristics but does not directly measure individual food consumption.
  • State-Level Granularity: While the study predicts state-level obesity, a more granular analysis at the zip code level could provide deeper insights.
  • Future Work Suggestions:
    • Replicating the analysis at the zip code level to better inform local food environments.
  • Integrating socioeconomic status and local food environment data for a more comprehensive understanding of obesity correlations.
  • Developing personalized mobile health apps that use GPS data and heat maps to provide real-time feedback on food environments and physical activity.
  • Expanding the model to include other social media platforms and data sources.

Summary updated

Abstract

Our study proposed a method for characterizing state-level food environments only using continuously updated social media data. State-level food environments were accurately described using social media data, and the model also showed a disparity in the available food between states with different obesity rates. The proposed method should elastically apply to local food environments of different sizes and predict obesity rates effectively.
BibTeX
@article{li2022novel,
  author = {Chuqin Li and Alexis Jordan and Jun Song and Yaorong Ge and Albert Park},
  title = {{A Novel Approach to Characterize State-level Food Environment and Predict Obesity Rate Using Social Media Data: Correlational Study}},
  journal = {Journal of Medical Internet Research},
  year = {2022},
  doi = {10.2196/39340},
}
Figure