Predictive Model Sets Priorities for Restaurant Inspections

Using data routinely collected in the city of Chicago, data scientists have developed a predictive system to identify which of the city’s more than 15,000 restaurants should be inspected first. By focusing inspections on those most likely to have critical violations, the city can reduce exposure to potential foodborne diseases, illnesses, and unsanitary conditions before people get sick.

The city’s restaurant inspectors are each responsible for checking about 470 establishments. The goal of the data-driven project—a collaborative effort that included Allstate’s quantitative research team and the Chicago’s departments of public health and innovation and technology—was to find a way to determine which restaurants should be inspected first because of their high risk.

This approach helps inspectors find critical violations faster than when “just doing business as usual,” Gene Leynes, data scientist for the city of Chicago, said during a presentation about the project at the Domino Data Lab in June.

These critical violations, he said, are the “ones that actually make you sick.”
Critical violations are usually due to improper temperature control for food, according to Gavin Smart, Allstate’s quantitative research director. One example, he said, is the mold that develops in an ice machine.

Building the predictive model meant identifying a dozen variables that contribute to the likelihood a restaurant will fail an inspection due to a critical violation.

According to Stephen Collins, the lead analyst on the project from Allstate, the predictors of significant problems were a history of previous or serious violations, the three-day average high temperature and weather fluctuations that make it difficult to maintain a certain temperature, the location of the restaurant, nearby garbage and sanitation complaints, the length of time since the last inspection, the amount of time the restaurant has been opened, and the inspector. The variables that proved to be most important in predicting risk were the establishment’s assigned risk level, whether it had failed previous inspections, its location, and nearby sanitation complaints.

A report about the food-inspection data program, called “Food Inspection Forecasting: Optimizing Inspections with Analytics,” shows that food inspectors can be allocated more efficiently. During a two-month, double-blind pilot study of the new predictive model, 69% of 178 establishments with critical violations were found during the first half of work, compared with 55% found during normal operations. Those critical violations were found one week earlier on average than would have been found using the normal inspection system.

Chicago has an online list of all restaurants, with the restaurant’s assigned risk level (high, medium, or low), the date last inspected, and the findings from the recent inspection at https://data.cityofchicago.org/Health-Human-Services/Food-Inspections/4ijn-s7e5.