Investigating the Association Between Text-Based Indications of Foodborne Illness from Yelp Reviews and New York City Health Inspection Outcomes (2023)

TL;DR

The paper addresses the gap in real-time foodborne illness surveillance by evaluating whether text-based signals from Yelp reviews, detected by the HSAN classifier, align with official NYC restaurant inspections in 2023. It integrates Yelp, NYC DOHMH inspection data, and Census tract geography, aggregating signals at the tract level and assessing linear and rank-based associations using Pearson $r$ and Spearman $\rho$, alongside nonparametric group comparisons. The analysis finds only weak overall associations ($r = 0.03$, $\rho = 0.05$ citywide) with notable borough-level variability, suggesting that online illness mentions reflect complementary rather than redundant information to formal inspections. The study highlights the potential for hybrid surveillance at finer spatial/temporal resolutions and outlines an agenda for address-level analyses, temporal alignment, and covariate control to enhance early-warning capabilities for public health agencies.

Abstract

Foodborne illnesses are gastrointestinal conditions caused by consuming contaminated food. Restaurants are critical venues to investigate outbreaks because they share sourcing, preparation, and distribution of foods. Public reporting of illness via formal channels is limited, whereas social media platforms host abundant user-generated content that can provide timely public health signals. This paper analyzes signals from Yelp reviews produced by a Hierarchical Sigmoid Attention Network (HSAN) classifier and compares them with official restaurant inspection outcomes issued by the New York City Department of Health and Mental Hygiene (NYC DOHMH) in 2023. We evaluate correlations at the Census tract level, compare distributions of HSAN scores by prevalence of C-graded restaurants, and map spatial patterns across NYC. We find minimal correlation between HSAN signals and inspection scores at the tract level and no significant differences by number of C-graded restaurants. We discuss implications and outline next steps toward address-level analyses.

Figures (4)

• Figure 1: Bivariate choropleth of mean HSAN score (x-axis classes) and mean restaurant inspection score (y-axis classes) by NYC Census tract, 2023.
• Figure 2: Univariate tract maps and review density.
• Figure 3: Scatter of tract mean HSAN score vs. mean inspection score by borough, 2023. Each point represents one Census tract.
• Figure 4: Distribution of tract mean HSAN scores by number of C-graded restaurants (1--2, 3--4, 5+).