~V
1237_C12.fm Page 321 Wednesday, November 12, 2003 1:14 PM
iO An Essay on the
Unrealized Potential of
Predictive Microbiology
Tom McMeekin
CONTENTS
12.1 Introduction
12.2 A Short History and the Philosophy of Predictive Microbiology
12.3 The Basics of Predictive Modeling
12.4 Addressing Concerns in Predictive Modeling
12.5 Identifying Opportunities from Predictive Modeling
12.6 Modeling Attachment to and Detachment from Surfaces
12.7 Modeling Fungal Growth
12.8 Application of Predictive Microbiology
12.9 Concluding Remarks
Acknowledgments
References
12.1 INTRODUCTION
This book presents contemporary views of the state of the art of predictive modeling
of microbial growth in foods, which, it is worth noting, has been the subject of much
research for more than 30 years. Contained herein the reader will sense that the "front
end" of the modeling process (data collection, model developing, model fitting) has
a scientifically sound basis and that the "middle bit" (tertiary models, applications
software, expert systems, etc.) should make the technology readily available.
However, one continues to have the sense that the predictive models and the
databases upon which they are based have not nearly reached their potential. This
"wrap up" chapter is intended to reinforce the view that predictive microbiology
research has addressed, and continues to address, perceived weaknesses of the
concept and offers the view that its potential as a food safety management tool
remains to be realized. The views expressed are personal, but probably shared by
others who have engaged in this type of research for many years. The style adopted
is that of an "essay" (i.e., a written composition less elaborate than a treatise) in the
2004 by Robin C. McKellar and Xuewen Lu
1237_C12.fm Page 322 Wednesday, November 12, 2003 1:14 PM
hope that a less formal presentation may encourage greater consideration by potential
users of the concept.
12.2 A SHORT HISTORY AND THE PHILOSOPHY
OF PREDICTIVE MICROBIOLOGY
Esty and Meyer 7 devised a mathematical model to describe the thermal death kinetics
of Clostridium botulinum type A spores that was used immediately, and since that
time as the basis for heat processing of nonacid canned foods. Its use continues,
despite subsequent understanding that a log-linear model may be an oversimplified
description of the rate at which death occurs. Log-linear kinetics ignore the com-
plications of "shoulders" and "tails," because the performance criterion applied, a
12-log-cycle reduction in spore numbers, introduces a very large safety margin.
Practitioners of heat processing use the time/temperature combinations derived with
great confidence in the microbiological outcome. Such confidence and immediate
and continuing application of the process suggest a single purpose rather than the
earliest introduction of the concept of predictive microbiology. This would have
implied considering the consequences of selecting less severe processing conditions.
In searching for the origin of predicting microbial behavior in foods, the trail
appears to begin with Scott 25 who researched the effect of temperature on microbial
growth on meat and wrote:
A knowledge of the rate of growth of certain micro-organisms at different temperatures
is essential to studies of the spoilage of chilled beef. Having these data, it should be
possible to predict the relative influence on spoilage exerted by the various organisms
at each storage temperature. Further, it would be feasible to predict the possible extent
of the changes in populations which various organisms may undergo during the initial
cooling of sides of beef in the meatworks when the meat surfaces are frequently at
temperatures very favourable to microbial proliferation.
Scott 24 also studied the effect of water availability on microbial growth on meat
and, it is interesting to note that while explicit models were not developed, the
knowledge accumulated was sufficient to allow shipments of nonfrozen meat from
Australia to markets in the U.K. and Europe, based on the combined effects of
temperature, water availability, and modified atmosphere. Later in this chapter we
will return to the proposition that significant benefit can be obtained from accumu-
lated knowledge or patterns of microbial population responses without transforming
that knowledge into an explicit mathematical description.
So, why bother taking the additional steps necessary to convert a response pattern
into a mathematical model? The answer is both practical and philosophical. In the
practical sense, the additional effort required to construct and validate a model will,
if properly carried out, lead to formulation of a general rule describing the effect of
the environmental responses studied on the growth, death, or survival of the target
population. A response pattern, on the other hand, is more likely to describe the
outcome of limited experimental trials, the applicability of which will require testing
2004 by Robin C. McKellar and Xuewen Lu
1237_C12.fm Page 323 Wednesday, November 12, 2003 1:14 PM
under even slightly changed conditions. Challenge tests are a well-established means
of describing a response pattern, or ensuring that a performance criterion for a
particular product/process combination is met.
Like the botulinum cook, the processing criteria or formulations scrutinized by
challenge tests will "overcook" the situation to ensure a very high probability that
the process will not "fail dangerous." The value of properly constructed and validated
models to replace, or significantly reduce, the need for expensive and time-consum-
ing challenge tests by providing a generally applicable solution is well documented.
As all microbial population responses are variable and characterized by a distribu-
tion, 18 knowledge of that distribution takes the predicted outcome to the next level
by allowing calculation of confidence limits for the predictions. The increased
precision available in turn allows greater confidence in specifying minimal process-
ing conditions, thus preserving the nutritional and quality characteristics of a product
without leading to microbial food safety or shelf-life problems.
The philosophical reasons to develop a predictive model lie in the nature of
science itself as expressed by Lord Kelvin: "When you can measure what you are
speaking about and express it in numbers you know something about it; but when
you cannot measure it, when you cannot express it in numbers your knowledge is
of a meagre and unsatisfactory kind."
Thus, quantitative science is inherently more useful than the qualitative description
of a phenomenon, in that the latter is embodied in the former. Nevertheless, we should
not lose sight of the fact that qualitative information is better than no knowledge, and
that semiquantitative response patterns may be appropriate to address a particular
question, to serve as a starting point or, as a "check and balance," when moving
towards quantitative descriptions of microbial population responses. When using
"black box" modeling approaches, such as polynomial models or artificial neural
networks, a priori knowledge is a valuable adjunct providing reality checks to ensure
that the developing model actually describes the observed biological response. 9
The history and philosophy of science likewise suggest that moving from an
empirical or phenomenological description to a mechanistic or deterministic descrip-
tion of a process represents an advance in the "good science" hierarchy. The former
descriptions are pragmatic in nature and give rise to stochastic models in which the
data are described by useful mathematical relationships. Mechanistic models, on the
other hand, have a theoretical basis, allowing interpretation of the observed response
on the basis of underlying theory; e.g., in considering microbial growth Monod 17
noted that, "There is little doubt that as further advances are made towards a more
integrated picture of cell physiology, the determination of growth constants will have
a much greater place in the arsenal of microbiology."
Of the secondary models commonly used in predictive microbiology, none can
be viewed as truly mechanistic, although some are regarded as more mechanistic
than others. Thus, Arrhenius-type temperature dependence models for microbial
growth are often thought to have a greater mechanistic basis than do Belehradek-
type (square-root or Ratkowsky-type) models. This contention would not have found
support from Belehradek 2 who, because of the origin of Arrhenius models in chem-
ical kinetics, wrote:
2004 by Robin C. McKellar and Xuewen Lu
1237_C12.fm Page 324 Wednesday, November 12, 2003 1:14 PM
~V
The problem of temperature coefficients in biology was initiated by chemists and has
suffered from the beginning from this circumstance. Attempts to apply chemical tem-
perature-velocity formulae (the Q\0 rule and the Van't Hoff-Arrhenius law) to bio-
logical processes failed because some of the temperature constants used in chemistry
(<2 10 , |i) can De sa id not to hold good in biological reactions.
While it may be argued that mechanistic models are more amenable to refinement
as knowledge of the system increases, development of fully mechanistic models for
microbial growth has been limited by inability, to date, to provide quantitative values
for all model parameters. Thermodynamic mechanistic models for the denaturation
of proteins based on the enthalpy and entropy convergence temperatures and the
heat capacity change 8 appear to apply to a wide range of proteins. 19 The growth of
bacterial populations also seems amenable to this approach as foreshadowed by Ross
and Olley in Chapter 10 of McMeekin et al. 13 and Ross 22 using Escherichia coli as
an example. The ever increasing speed of computer simulation has recently enabled
Drs. Olley, Ratkowsky, and Ross to apply this thermodynamic model to a very wide
range of bacteria, including true psychrophiles.
Heat capacity change, which is the driving force for protein denaturation, may
also explain the temperature response for bacterial growth, although it has not been
measured for whole cells. This lends support to June Olley's long held view that
the master reaction controlling temperature dependence is, in fact, the unfolding of
proteins (or other macromolecules), exposing hydrophobic groups to interactions
with surrounding water. Recent protein unfolding studies suggest that exposing polar
surfaces may also be involved. 19
When mechanistic models are invoked it is prudent to check the magnitude and
sign (+ or -) of parameter estimates. Inappropriate parameter values, e.g., enormous
activation energies, indicate that the model does not describe biological reality even
though a mechanistic basis may have been inferred. Moving to an even greater level
of "malpractice" in predictive model development one encounters situations in which
very limited data sets have been fitted to models with a surfeit of parameters, resulting
in an apparently perfect fit of the data to the model. This scenario represents an
exercise in curve fitting unique to the data set used to develop the model and, in the
end, because of false expectations based on erroneous assumptions of a mechanistic,
or even a solid, quantitative foundation, is significantly less useful than describing
a general response pattern.
12.3 THE BASICS OF PREDICTIVE MODELING
The rules for model selection were laid down by Ratkowsky in Chapter 2 of a
previous book on predictive modeling, 13 viz parsimony, parameter estimation prop-
erties, range of variables, stochastic assumption, and interpretability of parameters.
These continue to apply to the primary and secondary models described in Chapter
2 and Chapter 3 of this book and their use is supported by the model fitting techniques
described in Chapter 4.
An important prerequisite to model fitting is the need to plan ahead by selecting
an experimental design appropriate to the purpose of the study and to understand
2004 by Robin C. McKellar and Xuewen Lu
1237_C12.fm Page 325 Wednesday, November 12, 2003 1:14 PM
~V
the limitations of the method of data collection selected. While "traditional" viable
count and turbidimetric methods continue to dominate in modeling studies, several
alternatives are given in Chapter 1 of this book. The search, as is the case in
developing detection methods, is driven by:
1 . The desire to obtain results in a compressed time frame
2. Ease-of-use characteristics, including automation to reduce cost and
reduce the physical burden involved in collecting sufficient data at appro-
priate, usually close, time intervals and over an appropriate, often
extended, time period
Regardless of the method chosen, the standard remains the viable count method
against which other methods require rigorous validation. Despite the labor-inten-
sive, time-consuming nature of viable count methods, they can, with good labora-
tory practice, perform with sensitivity, accuracy, precision, reproducibility, and
repeatability.
Chapter 5 of this book provides a timely reminder that the test of a predictive
model is not how well it performs in well-controlled laboratory conditions, but how
well it predicts the behavior of microbial populations in real foods and in environ-
ments experienced under practical conditions of food production, processing, and
storage. Indeed, some researchers have advanced the opinion that initial development
of models in laboratory media represents wasted effort if subsequently the model
fails to provide an adequate description of a target organism's behavior in food. 5 In
particular, Brocklehurst (see Chapter 5, this book) and colleagues have drawn atten-
tion to important effects of food structure, including emulsions and surfaces that
may significantly affect microbial behavior.
Similar caveats on the performance and limits of models were advanced by
Ross 23 under the headings Model Applicability and Model Accuracy. The necessity
that models are applied only to relevant situations requires enunciation of the con-
ditions under which the model performs well and the boundaries beyond which
predictions should not be made. If a model is deemed appropriate, its accuracy must
also be considered, and this determination must take account of the fact that all
microbial responses are variable. Commonly used measures to evaluate model per-
formance are the bias and accuracy factors. 21 Note use of the term "evaluate" by
this author, whereas most authors use "validate" to describe the process of ensuring
that a model performs well in real foods subjected to anticipated conditions.
12.4 ADDRESSING CONCERNS IN
PREDICTIVE MODELING
While variability is recognized as a characteristic feature of response times, such as
the generation time or lag phase duration of a microbial population, that variability
is characterized by distributions such as the gamma distribution in which the variance
is proportional to the square of the response time. 18 This knowledge enables vari-
ability to be incorporated in models and confidence limits to be determined.
2004 by Robin C. McKellar and Xuewen Lu
1237_C12.fm Page 326 Wednesday, November 12, 2003 1:14 PM
~V
A more difficult suite of problems arises from the ability of microorganisms to
adapt readily to the selective pressures of the food environment. Variability and
adaptability were considered by Bridson and Gould 3 in a dichotomy described
respectively as classical vs. quantal microbiology, where it was argued that "large
populations of microorganisms obey the rules of taxonomy but the individual cells
exhibit uncertainties (caused by mutations and fluctuating local environments) which
are buried within the macropopulations" and "the functional stability of classical
microbiology masks minority subpopulations which, nevertheless, contribute to the
complex dynamics of microbial populations."
Conditions under which the adaptability of individual cells will influence the
performance of a predictive model include those where small numbers of cells are
found in a harsh environment causing decline in the viable population. This situation,
often found in food processing environments, will lead to a few viable cells repairing,
resolving the lag phase, and becoming the parent stock for the next phase of active
growth. 14
Adaptability and conditions where the rules of quantal microbiology apply give
rise to uncertainty and the remaining challenges for predictive modeling. These
now are centered on assessment of the initial conditions in a food, which will
determine the level of contamination (expressed as concentration or prevalence in
a sample), the initial physiological state of the population (or the survivors from
an original population), and the complexity of the food system, including the
microenvironment in which an individual organism is deposited. Uncertainty may
also arise if interactions occur between different components of the microbiota and
in fluctuating environments. The impact of the latter will depend on the magnitude
of the fluctuations. Approaches to deal with dynamic environments and history
effect on microbial growth and survival are discussed in Chapter 7 and Chapter 9,
respectively.
Fluctuating environments, particularly with respect to temperature, probably
represent a normal situation in food processing and during storage and distribution,
and the consensus is that lag times will be affected but that there is little (if any)
effect on generation times once the lag phase has been resolved. As an example, in
the author's laboratory, cycling Streptococcus thermophilus between 30 and 40°C
produced immediate changes to the anticipated growth rate. The microbiological
outcome of such fluctuations can, therefore, be predicted easily by a growth rate
temperature dependence model.
When temperature fluctuations are larger, a transient lag phase may be intro-
duced before exponential growth is resumed. A good example of this behavior was
provided by Baranyi et al. 1 working with Brochothrix therm osphacta. When the
incubation temperature was dropped from 25 to 5°C, the growth rate changed as
anticipated, but a shift from 25 to 3°C induced a lag phase in this psychrotolerant
spoilage bacterium.
Modeling the effects of severe fluctuations in temperature or other environmental
factors will be more difficult in that both lag and growth phases need to be considered
and, in some instances, the model will also have to account for death of a proportion
of the population. These situations provide examples where patterns of microbial
population behavior, without an explicit mathematical description, may indicate a
2004 by Robin C. McKellar and Xuewen Lu
1237_C12.fm Page 327 Wednesday, November 12, 2003 1:14 PM
practical control situation. Let us consider a temperature-based example that has
been studied at the pilot-plant level, and a laboratory-based water-activity example.
In the former, a pilot-plant-scale cheese-milk pasteurizer was used to study the
development and control of thermoduric streptococci biofilms during milk pasteur-
ization. Under normal operating conditions thermoduric streptococci grew on pas-
teurizer plates where the bulk milk temperature was between 35 and 50°C, and
were detected in the product stream after 8 to 10 h. Introducing a temperature step
change in the growth region of 55°C for 10 min with a 60-min interval between
step changes resulted in a 20-h production run without detectable growth of ther-
moduric streptococci. 11
In the latter example, Mellefont et al., 15 using optical density (OD) methods showed
that abrupt osmotic downshifts significantly increased the lag times of Gram-negative
organisms, but those of Gram-positive organisms were largely unaffected. Further
studies using viable count methods at close time intervals indicated that the apparent
increase in lag time in fact comprises a death phase, a true lag period and growth back
to the initial population levels (Mellefont, L.A., personal communication).
Returning to temperature shifts, Mellefont and Ross 16 found that downshifts
induced significant increases in the lag phase duration of Escherichia coli and
Klebsiella oxytoca that were dependent on the magnitude of the shift. These authors
suggested that lags were introduced when the culture was shifted to temperatures
beyond the normal physiological temperature range (NPTR) for growth, where cells
are required to do additional work to adjust to the new environment and the rate at
which that work is done. 20 At temperatures within the NPTR, the energy requirement
for cellular functions to proceed is unchanged, the cells are "cruising," and thus
shifts within this region are characterized by a simple change of rate without inter-
ruption to the growth cycle. This is a plausible hypothesis but much more work is
required to firm up the concept and determine the physiological significance of the
NPTR. The notion of a normal physiological range for other factors such as water
activity and acidity should also be addressed.
Thus, predicting lag phase duration in foods is considered problematic because
of the twin uncertainties of the initial physiological state of organisms and the
numbers initially contaminating a product. Despite the fact that lag phase models
can be developed in the laboratory with reasonable success (as the uncertainties
above are minimized), this has remained one of the more intractable problems in
predictive modeling.
A potential solution, however, was provided by Ross 23 using the trusted modelers
device of moving from a kinetic to a probabilistic modeling approach to confront
increasing uncertainty in describing initial conditions. The approach was to concede
that while lag times are highly variable, the variability can be reduced by using the
concept of relative lag times or "generation time equivalents," i.e., the ratio of lag
time to generation time. By this device it was shown that although lag times may
take almost any value, there is a common pattern of distribution of relative lag times
for a wide range of species across a wide range of conditions. That common
distribution of relative lag times has a sharp peak in the range of four to six generation
time equivalents accounting for >80% of lag phase duration determinations. 23 This
stochastic approach has significance for the application of predictive models in that
2004 by Robin C. McKellar and Xuewen Lu
1237_C12.fm Page 328 Wednesday, November 12, 2003 1:14 PM
when an adjustment of -5 generation time equivalents is added to generation time
predictions, good agreement of observed microbial proliferation on carcasses during
chilling was observed.
12.5 IDENTIFYING OPPORTUNITIES FROM
PREDICTIVE MODELING
In a practical sense it is clear that interrupting the exponential growth phase of a
target organism is an effective strategy to buy time before critical limits of spoilage
or pathogenic bacteria are reached. In the pasteurizer example, interrupting growth
by intermittent temperature changes has the potential to more than double the run
time of the heat exchanger with respect to thermoduric streptococci. An alternative
biofilm control strategy is to minimize microbial adhesion to the pasteurizer plates,
e.g., by a Teflon surface coating. However, this approach is inherently limited in its
efficacy, with a 50% reduction in the initial load translating to an extension in run
time equivalent to one generation time, for S. thermophilus -20 min, and a 90%
reduction to -3 generations (~1 h) at its optimum temperature.
Buying time may also be useful in the context of a sequence of processing
operations if the effect of changes occurring during each operation on microbial
population dynamics is understood. Meat processing provides an interesting case
study as regulatory authorities contemplate mandating a microbial "kill" step for
certain pathogens during the conversion of live animals to meat. This philosophy
will require the application of an intervention to achieve a reduction in pathogen
numbers and, with current technology, the options are acid or hot water treatments.
But could a better understanding of the chilling process reveal a potentially valuable
intervention step without the significant cost burden of acid or hot water cabinets
or steam pasteurization equipment? Such an opportunity is suggested by the results
of Chang et al. 4 who studied reduction of bacteria on pork carcasses associated with
chilling, concluding that "the effects of chilling techniques on microbial populations
could provide pork processors with an additional intervention for pork slaughter or
information to modify and/or improve the chilling process."
While lowering temperature and, in many chilling operations, simultaneously
reducing surface water activity may not result in as great a reduction in microbial
numbers as a heat treatment, understanding and "tweaking" these operations may
prove valuable:
1. To identify and characterize a new critical control point (CCP)
2. As part of the "farm to fork" philosophy where every operation has an
impact on the final level of risk to the consumer
Food safety professionals write often about the Hurdle concept in which multiple
barriers confront microbial contaminants to delay resolution of the lag phase and
the onset of exponential growth. However, mostly we think of applying hurdles
via the intrinsic properties of the food (water activity, pH, organic acid concentra-
tion, etc.) and the extrinsic conditions of storage (temperature, atmosphere, etc.).
2004 by Robin C. McKellar and Xuewen Lu
~V
1237_C12.fm Page 329 Wednesday, November 12, 2003 1:14 PM
It may also be useful to think of the cumulative effect of hurdles applied in a
sequence of processing operations leading to a final product. In the meat chilling
situation outlined above, hurdles leading to inactivation and stasis of microbial
populations will be supplemented by the effect of downstream processes. A case
in point is freezing cartons of meat destined for the hamburger trade, where ice
crystal formation might be expected to cause further damage to cells injured during
chilling.
12.6 MODELING ATTACHMENT TO AND
DETACHMENT FROM SURFACES
Don Schaffner, in Chapter 10 of this book, draws attention to a recent research trend
in food microbiology concerned with modeling contamination and decontamination
processes. This subject was reviewed recently by den Aantrekker et al. 6 Encompassed
within the general area, researchers will need to consider contamination of foods
and food contact surfaces (including hands), transfer of organisms to foods from
surfaces and vice versa, removal of organisms from surfaces, the potential for and
significance of recontamination of foods with small numbers of pathogens, etc.
The field of research, arising from the propensity of organisms to become
intimately associated with surfaces as a survival mechanism, has spawned a subdis-
cipline of microbiology concerned with a continuum from the early events of adhe-
sion to the formation of mature biofllms and their detachment from surfaces. Two
aspects of these types of studies will be considered briefly here. The first is modeling
attachment and detachment of biofllms, a general treatment of which was reviewed 6
under the heading Recontamination through Equipment. From this readers will be
able to identify the relevant original literature. A specific example of this type of
study, supporting the step change control strategy for S. thermophilus biofllms 11 was
reported by Lee. 12 The model developed suggested that increasing the generation
times of S. thermophilus represents the most effective way of controlling biofilm
formation and subsequent detachment into cheese-milk.
The second aspect considers the important role of fluid transport in contamina-
tion events, which, in this author's opinion, is an obvious but often overlooked
phenomenon in recent times. However, a considerable literature was generated on
the role of fluid transfer in poultry processing in the 1970s (a quarter of a century
ago), mainly by Dr. S Notermans and colleagues in The Netherlands and Dr. CJ
Thomas and colleagues at the University of Tasmania. A model formulated to
describe contamination during immersion chilling of poultry carcasses relies on the
observation that the number of bacteria transferred from processing water to the skin
of a broiler carcass is directly proportional to the number of organisms in the water.
Effectively, when the carcass (or any other surface) is removed from a liquid it
carries with it a sample of the liquid in which the density of the organisms is the
same as that of the bulk liquid. The simple relationship has found utility, e.g., in the
design of countercurrent immersion chillers in which the carcasses exit the chiller
at the point where the microbial load in the water is lowest.
2004 by Robin C. McKellar and Xuewen Lu
1237_C12.fm Page 330 Wednesday, November 12, 2003 1:14 PM
~V
12.7 MODELING FUNGAL GROWTH
The great majority of predictive modeling studies to date have described the effect
of environmental factors on bacterial growth. Viruses and parasites effectively have
no growth ecology in foods, but patterns of decline have been described, e.g., for
the effect of freezing on Trichinella in pork. As we are reminded in Chapter 11,
molds are very important food spoilage organisms and the toxins they produce may
lead, usually in the longer term, to public health problems. Despite their importance
in food microbiology, the study of predictive mycology is very limited when com-
pared with studies on bacteria. Nevertheless, sufficient material is available for
Philippe Dantigny to provide a short review of fungal modeling studies in foods,
including some interesting points for discussion such as the inability of the square-
root model to describe the effects of temperature on the kinetics of mold growth.
Insights might also arise from returning to the plant pathology literature that,
almost 50 years ago when I was a boy in Northern Ireland, was the basis of public
broadcasts predicting the likelihood of potato blight (Phytophthora infestans) or
apple scab (Venturia inequalis) problems as a result of prevailing climatic conditions
(temperature and relative humidity). In essence, these forecasts might be considered
an early example of risk assessment, predicting the incidence of plant disease rather
than human disease.
12.8 APPLICATION OF PREDICTIVE MICROBIOLOGY
To use predictive models practically in the food industry requires devices that
monitor the environmental conditions in that part of the paddock to plate continuum
of interest and a means to translate that environmental history into an estimate of
the growth, survival, or death of a target organism.
The devices available are chemical or physical monitors where the interpretative
function is built into the device, e.g., as a color change resulting from a chemical
reaction or, in physical mode, the extent of migration of a dye along a wick. 13 Implicit
in the efficacy of monitors with a built-in interpretative function is that the rate at
which the chemical or physical change occurs mimics that of microbial behavior
under the same conditions. Unfortunately, such monitors are based almost exclu-
sively on Arrhenius reaction kinetics, which deviate from biological response rates
as limits for growth are approached. Often these regions are of greatest interest in
predicting the shelf life and safety of foods.
An alternative to built-in interpretation is to construct a tertiary model, 26 usually
a spreadsheet-based program that converts a monitored temperature history into an
estimate of microbial growth potential. Many such programs are available and
prominent ones are described in Chapter 6 of this book. That chapter also describes
a most significant initiative in predictive modeling: the development of COMBASE,
initially by combining the U.S. -developed Pathogen Modeling Program and the
U.K. -developed Food MicroModel, but to which other significant databases could
be added. Further, the authors of Chapter 6 draw attention to the main pillars of
predictive microbiology software packages: databases and mathematical models.
Here they point out that while most chapters in this book deal with aspects of
2004 by Robin C. McKellar and Xuewen Lu
1237_C12.fm Page 331 Wednesday, November 12, 2003 1:14 PM
~V
mathematical modeling, more emphasis needs to be placed on the value of databases
and the scientific study termed Bioinformatics. The power of the database, perhaps
through the agency of a mathematical model or algorithm, is described via the use
of Expert Systems that provide decision support. These include the possibility of
in-line real-time systems.
Such developments highlight the interface of predictive microbiology with infor-
mation technology systems, allowing the application of models in food safety man-
agement strategies including Hazard Analysis Critical Control Point (HACCP),
Quantitative Risk Assessment, and Food Safety Objectives. Much has been written
about specific opportunities for predictive models to underpin these strategies that
need not be repeated here, e.g., see Chapter 8 of this book. Further, there is a
particular role for predictive models in comparing the microbiological outcomes of
various operations in food manufacture. Predictive models based on detailed knowl-
edge of the microbial ecology of any product/pathogen combination enable us to:
1 . Quantify the effect of preservation technologies on microbial populations
2. Optimize existing or suggest new processing procedures
3. Indicate risk management options
4. Identify regulations that are unwarranted
5. Support the need for outcome-based regulations
6. Enable equivalence determinations
The last two applications are likely to become prominent in the global debate on
regulatory frameworks to assure the safety of food in international trade and attendant
market access issues that sit alongside the protection of public health.
However, reticence to realize the potential value of predictive modeling appli-
cations continues as exemplified by comments from the Food Safety and Inspection
Service (FSIS) of the USDA. In July 2002, FSIS issued a notice entitled "Use of
microbial pathogen computer modeling in HACCP plans" (www.usda.fsis.gov) that
presented a particularly negative account of the potential to use microbial pathogen
computer modeling (MPCM) programs in the development and use of HACCP plans.
While acknowledging that MPCM may be useful in "supporting hazard analyses,
developing critical limits and evaluating the relative severity caused by process
deviations," FSIS states categorically, "It is not possible or appropriate to rely solely
upon a predictive modeling program to determine the safety of food and processing
systems." Furthermore, "determining pathogen growth and survival, and controlling
it in food products often requires complete and thorough analysis by an independent
microbiology laboratory, challenge studies and surveys of the literature."
In effect, FSIS listed the elements of a properly conducted and independently
evaluated predictive modeling study:
• Thorough analysis of the literature to reveal general patterns of microbial
population behavior
• Challenge studies to validate that the proposed model is applicable in
practice
2004 by Robin C. McKellar and Xuewen Lu
1237_C12.fm Page 332 Wednesday, November 12, 2003 1:14 PM
~V
• Independent evaluation to verify that an MCPM program consistently
allows a process to meet agreed critical levels at identified CCPs
There is, of course, a requirement for those proposing the use of predictive
models in HACCP plans to ensure that the model predictions are used conservatively.
This is a fact well recognized by predictive modeling researchers who promote
caution in the use of models, e.g., McMeekin and Ross. 14 These authors recognized
clearly that
there are two criteria that must be satisfied if any form of application software is to be
used effectively. The first is a properly developed and validated model and the second
the ability of an operator to interpret correctly the microbiological significance of the
results. Poorly performing models coupled with poor interpretation continue to be the
greatest threat to widespread use of predictive microbiology as a technology with the
potential to assure the microbiological quality and safety of foods.
A somewhat different perspective on the use of predictive models to evaluate
food safety was provided by an International Food Technology (IFT) committee in
their report "Evaluation and Definition of Potentially Hazardous Foods." This was
prepared for the FDA (Dec. 31, 2001; IFT/FDA Contract No 223-98-2333, Task
Order No 4).
In this report IFT recognizes that certain foods have combinations of pH, a w ,
preservatives, etc., that restrict microbial growth and may, therefore, not require
refrigeration to protect public health. Under certain circumstances time alone can
be used to control product safety, e.g., "if S. aureus is a concern the Pathogen
Modeling Program V.5.1 could be used to estimate the time of storage where the
pathogen could grow."
The committee further suggests that general growth models such as Pathogen
Modeling Program should be used conservatively or in combination with challenge
testing. However, if an in-house model has been developed and validated for a
particular food, it could be used itself or with challenge testing.
12.9 CONCLUDING REMARKS
This is only the third major book devoted to predictive microbiology in approxi-
mately 30 years of scientific endeavor in the field. The first, 13 referred to in the
Preface of this book, appeared in 1993; the second, 23 which focused on predictive
models for the meat industry, was not widely distributed.
The chapters in this book cover the entire gamut of predictive modeling research
and it is appropriate that the "middle" chapter, Chapter 6, describes database devel-
opment, the fulcrum on which predictive modeling swings.
Here we find a description of how new knowledge is accumulated and stored in
databases that, when properly constructed, are dynamic information repositories to
which new information can be added at any time and from which that found to be
"dodgy" can be deleted. Speaking the same language is an important element of
database construction and a prerequisite to merging existing databases. The authors
2004 by Robin C. McKellar and Xuewen Lu
~V
1237_C12.fm Page 333 Wednesday, November 12, 2003 1:14 PM
of Chapter 6 have expended considerable effort to make two major databases —
Pathogen Modeling Program and Food MicroModel — compatible and to merge
these into COMBASE. This initiative will also guide researchers toward the best
way to collect and collate new information so that it can be easily integrated into
existing databases. Properly supported, the COMBASE initiative will be a watershed
in the evolution of predictive modeling and its widespread application.
Much of what underpins a predictive models database is regarded by many as
difficult science. However, this need not unduly concern potential users of models
if there is confidence that the underpinning science is sound. What comes after this
point, in the form of user-friendly interpretative devices, will determine whether or
not the potential of predictive microbiology is realized.
Many specific applications have been suggested for predictive models and some
are mentioned earlier in this chapter. However, in a more general sense, the value of
predictive models and databases lies in their support of food safety management
strategies such as HACCP, Quantitative Risk Assessment, and Food Safety Objectives.
In particular, the role of quantitative information in empowering the HACCP
concept must not be undersold. HACCP is a simple, but elegant, concept by which
safety is built into a process. It works well where critical control points, e.g., a lethal
heat process, can be identified. Its value is less obvious where critical control points
do not stand out, e.g., in the conversion of muscle to meat. Its value is compromised
where the HACCP concept is virtual rather than real, i.e., where generic criteria are
applied to satisfy mandatory requirements that a process is HACCP-based rather
than using criteria derived from a knowledge of microbial population behavior.
The challenge now for predictive microbiology is to move from the phase of
model development and model validation (evaluation) to process validation and
verification studies that will enable models to be used with confidence in HACCP
systems as demonstrated by Jericho et al. 10
The "hard yakka" (Australian vernacular for hard work) has been done in
developing the databases and models and now there are tantalizing prospects for in-
line control of processes, appropriate corrective actions, and sound food safety
management decisions arising from the concept of predictive microbiology.
ACKNOWLEDGMENTS
The author is indebted to Dr. June Olley for stimulating discussions in the preparation
of this essay and to Meat and Livestock Australia (MLA) for continuing support of
predictive modeling research.
REFERENCES
1. Baranyi, J., Robinson, T.P., Kaloti, A., and Mackey, B.M., Predicting growth of
Brochothrix thermosphacta at changing temperature. Int. J. Food Microbiol. 27, 61,
1995.
2. Belehradek, J., Temperature coefficients in biology. Biol. Rev. Biol. Proc. Camb.
Philos. Soc. 5, 30, 1930.
2004 by Robin C. McKellar and Xuewen Lu
1237_C12.fm Page 334 Wednesday, November 12, 2003 1:14 PM
3. Bridson, E.Y. and Gould, G.W., Quantal microbiology. Lett. Appl. Microbiol. 30, 95,
2000.
4. Chang, V.P., Mills, E.W., and Cutter, C.N., Reduction of bacteria on pork carcasses
associated with chilling method. J. Food Prot. 66, 1019, 2003.
5. Dalgaard, P., Mejlholm, O., and Huss, H.H., Application of an iterative approach for
development of a microbial model predicting the shelf- life of packed fish. Int. J. Food
Microbiol. 38, 169, 1997.
6. den Aantrekker, E.D., Boon, R.M., Zwietering, M.H., and van Schothorst, M., Quan-
tifying recontamination through factory environments: a review. Int. J. Food Micro-
biol. 80, 117, 2002.
7. Esty, J.R. and Meyer, K.F., The heat resistance of the spore of B. botulinum and allied
anaerobes, XL J. Infect. Dis. 31, 650, 1922.
8. Fu, L. and Friere, E., On the origin of the enthalpy and entropy convergence temper-
atures in protein folding. Proc. Natl. Acad. Sci. 89, 9335, 1992.
9. Geeraerd, A.H., Valdramidis, V.P., Devlighere, E, Bernaerts, H., Debevere, J., and
Van Impe, J.F., Development of a novel modelling methodology by incorporating a
priori microbiological knowledge in a black box approach. Proc. 18th International
ICFMH Symposium, Lillehammer, Norway, 2002, p. 135.
10. Jericho, K.W.F., O'Laney, G., and Kozub, G.C., Verification of the hygienic adequacy
of beef carcass cooling processes by microbiological culture and the temperature-
function integration technique. J. Food Prot. 61, 1347, 1998.
11. Knight, G., Nicol, R.S., and McMeekin, T.A., Temperature step changes: a novel
approach to control biofilms of thermoduric streptococci in a pilot plant-scale cheese-
milk pasteuriser. Int. J. Food Microbiol. In press.
12. Lee, M.W., Growth Rate and Detachment Rate of a Biofilm of Thermoduric Bacteria,
M.Sc. thesis, Monash University, Clayton, VI, 1998.
13. McMeekin, T.A., Olley, J., Ratkowsky, D.A., and Ross, T. Predictive Microbiology:
Theory and Application. Research Studies Press Ltd., Taunton, U.K., 1993.
14. McMeekin, T.A. and Ross, T., Predictive microbiology: providing a knowledge-based
framework for change management. Int. J. Food Microbiol. 78, 133, 2002.
15. Mellefont, L.A., McMeekin, T.A., and Ross, T., The effect of abrupt osmotic shifts
on the lag phase duration of food borne bacteria. Int. J. Food Microbiol. 83, 281, 2003.
16. Mellefont, L.A. and Ross, T., The effect of abrupt shifts in temperature on the lag
phase duration of Escherichia coli and Klebsiella oxytoca. Int. J. Food Microbiol.
83, 295, 2003.
17. Monod, J., The growth of bacterial cultures. Ann. Rev. Microbiol. 3, 371, 1949.
18. Ratkowsky, D.A., Ross, T., Macario, N., Dommett, T.W., and Kamperman, L., Choos-
ing probability distributions for modelling generation time variability. J. Appl. Bac-
teriol. 80, 131, 1996.
19. Robertson, A.D. and Murphy, K.P., Protein structure and the energetics of protein
stability. Chem. Rev. 97, 1251, 1997.
20. Robinson, T.P, Ocio, M.J., Kaloti, A., and Mackey, B.M., The effect of growth
environment on the lag phase of Listeria monocytogenes. Int. J. Food Microbiol. 44,
83, 1998.
21. Ross, T., Indices for performance evaluation of predictive models in food microbiol-
ogy. /. Appl. Bacteriol. 81, 501, 1996.
22. Ross, T., Assessment of a theoretical model for the effects of temperature on bacterial
growth rate. Predictive microbiology applied to chill food preservation. Proc. Conf.
No. 1997/2, Comm. C2, Quimper, France, International Institute Refrigeration, Syd-
ney, 1999, p. 64.
2004 by Robin C. McKellar and Xuewen Lu
1237_C12.fm Page 335 Wednesday, November 12, 2003 1:14 PM
23. Ross, T., Predictive Microbiology for the Meat Industry, Meat and Livestock Australia,
Sydney, 1999, 196 pp.
24. Scott, W.J., The growth of microorganisms on ox muscle. I. The influence of water
content of substrate on rate of growth at -1°C. J. Counc. Sci. Ind. Res. (Aust.) 9, 177,
1936.
25. Scott, W.J., The growth of microorganisms on ox muscle. II. The influence of tem-
perature. /. Counc. Sci. Ind. Res. (Aust.) 10, 338, 1937.
26. Whiting, R.C. and Buchanan, R.L., A classification of models for predictive micro-
biology. Food Microbiol. 10, 175, 1993.
2004 by Robin C. McKellar and Xuewen Lu