Food Protection Trends - January/February 2020 - 56

GENERAL INTEREST PAPER

Predictive Models for Food Code Violations
Jim Hartman, R.S., M.S.
Retired Public Health Sanitarian, 196 S. Grant Ave., Columbus, OH 43215, USA

SUMMARY
Many health departments use hypothetical "risk classification" concepts, often including establishments' use of "risky"
food processing procedures, to prioritize inspections. In this
paper, descriptions of some strategies for scheduling inspections based on this sort of perceived risk will be followed by
a review of more fine-tuned classification models based on
inspection history and other factors. A strategy for scheduling
inspections based on detection of sporadic cases of foodborne
illness will then be outlined. The advent of social media, machine learning and portable digital technology (smart phones)
have begun to revolutionize this field, although some of the
investigative techniques used in earlier work have not yet been
applied to the newer approaches. The goals of this review are to
bring some of these strategies together and offer recommendations for advancing them.
OVERVIEW
Through much of the latter part of the twentieth century,
public health practitioners searched for ways to assess the
risk of a restaurant causing an outbreak of foodborne disease.
The goal of predictive modeling in this context is to use risk
for scheduling and focusing inspection programs. Because
of the difficulty of measuring risk directly, easier-to-measure
characteristics or performance measures have been proposed.
Because we often hear the words "analog" and "digital,"
it may be instructive to ponder exactly what they mean. An
analog clock has hands that sweep around a dial to indicate
the passage of time. However, this movement "is not time
itself: it's a representation or analogy of time" (28). In a
similar way, the presence of certain features is not risk itself,
but can represent it. A more complicated computer model,
to be described later, uses concepts from earlier models using
categories, but without employing any analogies to risk.
Other models to be described really are digital: they use the
presence or absence of illness, similar to 1s and 0s. Of course,
merely being digital does not imply being more accurate than
analog methods.
All these methods have the potential to benefit restaurant
patrons by identifying food safety problems earlier than
they would otherwise be found, thus preventing many cases
of foodborne illness. In addition, because some of these
illnesses are due to infectious agents, the general public,
including people who rarely or never eat out, could benefit.
Application of predictive modeling could help public health

Food Protection Trends January/February

agencies in charge of conducting inspections by helping them
achieve their public health goals even if they miss some of
their mandated inspections. Finally, restaurants themselves
stand to benefit in that their food preparation practices could
be improved and their legal liability diminished.
However, much of the work in these areas has appeared in
information technology journals rather than in food safety
or public health publications. Furthermore, information
technologists have not applied food safety concepts such
as HACCP. This article is an attempt to spread the word
about these methods to a broader audience and, it is hoped,
improve them.
EARLY MODELS
The U.S. Food and Drug Administration (FDA) promulgated the United States' first "Ordinance regulating eating
and drinking establishments," a mimeographed document,
in 1935 (25). The 1976 Food Service Sanitation Manual (24)
recommended semi-annual inspections. The idea of adjusting the inspection frequencies of food service operations
(restaurants) and retail food establishments (markets, etc.)
according to some theory of their risk of causing illness has
been evolving ever since.
Kaplan and EI-Ahraf (14) were among the first researchers
to tabulate data on reported outbreaks of foodborne illness
for the purpose of estimating risk according to the type of
establishment involved. Ten years later, Irwin et al. (11) set
up a case-control study to examine the violations reported
on the last routine inspection report for the restaurant that
had caused each of 28 outbreaks. Controls were matched to
them by routine inspection date. The best predictor of which
food services would later cause an outbreak was found to
be "any improper food protection practice" (11). Bryan (3)
suggested basing inspection frequencies on the presence of
foods often implicated as vehicles of foodborne illness, risky
food processing steps, and average daily patronage. Briley
and Klaus (2) used ideas from Bryan, as well as Kaplan and
EI-Ahraf 's, and added a statistic based on the average score
from the last five inspections. Wodi and Mill (27) also used
a predicted risk score based on the last two inspection scores
and critical items violated.
Columbus models
Two predictive models have been proposed at the Columbus,
Ohio, Health Department (now Columbus Public Health).

http://www.foodprotection.org/files/food-protection-trends/jan-feb-20-hartman.pdf

Food Protection Trends - January/February 2020

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