Food Protection Trends - January/February 2020 - 59

would show the actual risk associated with specific foods
and processing errors and determine how much deviation
from control is necessary to cause illness. For example,
we know the percentage of outbreaks caused by improper
cooling, but we do not know the percentage of improperly
cooled foods that cause illness. What is the probability that
bean soup cooled in a huge pot for eight hours will cause
illness? To answer this question, we would have to investigate
meals eaten by people who did and people who did not get
sick. Many (or most) inspections in the jurisdiction would
be targeted to investigate restaurants or other sources of
foods identified by the web site. A data-processing system
capable of tabulating and comparing food processes used
by restaurants involved with cases and those involved with
controls would be necessary.
For example, suppose the case had cramps and diarrhea
and a stool test revealed the presence of Bacillus cereus. As
Table 2 shows (4), attention would focus on foods eaten
7-12 hours earlier. Table 3 (26) suggests focusing attention
on starchy foods that may have been held at incubating
temperatures. Under the theory that improper hot holding
of starchy food caused the illness, either such a food would
not appear in the food history of any controls matched to this
case, or the food would have been held at a safe temperature.
NRICGP reviewers rejected each proposal. The 1996
review panel summary, for example, suggested that "the
proposed work may exceed the capabilities of the [Principal
Investigator] ... the proposed testing using PCR methods
may not be readily available and may be extremely costly ...
[and a] study design limited to a small case-control approach
to investigate foodborne exposures may provide a clearer
focus for future studies." This proposal is cited here because
it used a unique method for identifying cases, and the case
follow-up methods appear to be applicable to later work.
RECENT MODELS
For many decades, it has been possible to call in complaints to the health department by telephone; now it is
also easy to file complaints via government websites. These
methods, however, are notoriously subject to underreporting.
In the past few years, papers have appeared about the use of
social media to monitor disease - so-called infodemiology
and infoveillance (18). Oldyrod et al. (18) did a computer-assisted search of 5 databases and reviewed 5,239 papers
discussing use of these methods for disease surveillance and
found the papers focused mostly on influenza-like illness;
only 10 discussed use of consumer-generated data for foodborne illness surveillance.
Harrison et al. (7) used custom software to analyze
294,000 New York City restaurant reviews on the business
review website Yelp (www.yelp.com) and considered 468 of
them to have been reportable as foodborne illness outbreaks,
although only 3% of those had actually been reported to the
New York City Department of Health and Mental Hygiene.

Of 129 reviewers describing 2 or more sick persons or
scombroid poisoning or a neurological illness, 102 refused to
be interviewed by investigators.
Kang et al. (13) also used Yelp reviews and claimed to be
the first to compare them with official restaurant inspection
results. They "scraped" Seattle restaurant reviews from 2006
to 2013 and found that over half the restaurants had no
corresponding inspection records. They claim 82% accuracy
in predicting the inspection results where these results
existed. Their work generated tables similar to Table 4.
Interestingly, they found that restaurants that reviewers called
"pretentious" were hygienic, whereas when reviews called
restaurants "cheap" or singled out specific food ingredients,
e.g., noodles, "one can extrapolate that the overall experience
probably was not glorious."
These methods do not seem to depart radically from earlier
methods, and they have a critical drawback: they are time- and
resource-consuming. However, they also have an important
advantage: they supplement traditional reporting systems.
Chicago's food safety inspection forecasting model
In 2002, Chicago Mayor Rahm Emanuel issued an
executive order making "Open Government Data" a priority
and requiring all city agencies to make "all appropriate data
sets" available for public use (5). The city won a grant from
Bloomberg Philanthropies in 2013 to develop innovations
they could share with other cities. A working group, including
the data science team at Allstate Insurance, chose the idea of
a predictive model for food code violations partly because
the city had 15,000 food service operations but only 36 food
sanitarians (17), which comes out to 2.4 hours per inspection,
including travel time, if each location received an average of
2 inspections per year. The model they developed was tested
over 2 months in 2014 and went into use in 2015 (17).
The right side of Table 1 shows the 9 predictor variables
Chicago uses (22). Chicago tested the model by using
it to generate inspection schedules and compared those
inspection results to those resulting from inspection
assignments generated in the ordinary way. Managers did
not inform sanitarians which lists were generated by the
model. The result was that 55% of the inspections scheduled
in the usual way found critical violations, whereas 69% of
the inspections scheduled by the model resulted in critical
violations. Also, the model identified the problem restaurants
(of which there were 35) 7.5 days earlier than the normal
scheduling method would have.
Chicago published their model as "open source code" so
that it could be used and possibly improved on in other jurisdictions. But as of early 2016, only one other jurisdiction had
used it - Montgomery County, Maryland, near Washington,
D.C. They also found 27% more violations in the first month
than would have been found with the normal scheduling and
found them 7.5 days earlier (22, 23).

January/February Food Protection Trends

http://www.yelp.com

Food Protection Trends - January/February 2020

Table of Contents for the Digital Edition of Food Protection Trends - January/February 2020