Foodborne Illness — Reporting Disparities via Yelp

The Problem

• Reviews vs Reports Gap: Many diners post about getting sick in online reviews but never notify health authorities, creating an unseen pool of incidents.
• Potential disparities: Certain demographics or locales may rely on review sites instead of official channels, leaving public agencies unaware of issues in those populations.
• Missed signals: Traditional surveillance fails to capture these crowdsourced complaints, risking that outbreaks in some communities go unnoticed.

The Solution

• Scraped, machine-classified and analyzed millions of Yelp reviews for keywords and narratives indicating foodborne illness experiences at restaurants.
• Compared the volume and pattern of Yelp-mentioned illness incidents with official food poisoning complaint data to quantify reporting gaps.
• Highlighted geographic and demographic patterns where many people complain online but far fewer file official reports, enabling health departments to identify where outreach or education is needed.

Architecture Overview

• Data Collection: Implemented web scrapers to gather Yelp review data across numerous restaurants and regions, building a large dataset of customer feedback.
• NLP Extraction: Used text processing to detect mentions of food poisoning symptoms or complaints within reviews, filtering out unrelated content.
• Machine Learning: Using manually classified data, we trained a Support Vector Machine to identify food poisiong related reviews and if reported for themselves or others.
• Data Fusion: Matched the timeline and location of Yelp-indicated incidents with corresponding official public health reports to directly compare unofficial vs. official reporting rates.
• Statistical Analysis: Analyzed seasonality and socio-demographic differences. Modeled factors (e.g., neighborhood income levels, restaurant types) associated with high Yelp complaint activity but low official reporting, to understand underlying causes of reporting disparities.
• Visualization: Created charts and heatmaps to illustrate areas of underreporting, helping stakeholders easily grasp the extent and location of the gaps.

Results and Impacts

• Revealed hotspots where Yelp reviews indicated numerous illness incidents yet official reports were scarce, signaling critical underreporting in those areas.
• Restaurant-heavy counties had the most online illness reports. This was the strongest pattern (r = 0.75).
• Richer, more educated counties reported more illness online. Higher bachelor’s degree levels (r = 0.44), high-school graduation (r = 0.40).
• Poorer counties reported less online. Higher uninsurance (r = -0.28), unemployment (r = -0.23), poverty (r = -0.18), and SNAP use (r = -0.16).
• Yelp-based food illness reporting looks useful, but it is biased toward restaurant-dense, higher-SES counties, so it should complement—not replace—traditional surveillance.

Skills and Tools Used

Cross-Project Capabilities

• Non-Traditional Data Utilization: Showcased the ability to convert consumer-generated content (reviews) into actionable health insights, a methodology applied in multiple projects.
• Scalable Analysis: Gained experience handling millions of data points and extracting meaningful patterns, skills applicable to any large-scale data challenge.
• Equity Perspective: Experience analyzing data for disparities and gaps informs efforts in other projects to ensure fair and inclusive data-driven interventions.

Published Papers/Tools

• Peer-Reviewed Study: Published findings in a public health journal, detailing how Yelp review analysis can expose hidden food safety issues. Paper
• Conference Presentation: Presented this work via a poster at a national public health informatics conference (2020), sharing the approach with a broader audience.