Quantifying the Sensitivity of Scent Detection Dogs To Identify Fecal Contamination on Raw Produce

Abstract

Consumption of raw produce commodities has been associated with foodborne outbreaks in the United States. In a recent Centers for Disease Control and Prevention report outlining the incidence of food-related outbreaks from 1998 to 2008, produce of all kinds were implicated in 46% of illnesses and 23% of deaths. Methods that quickly identify fecal contamination of foods, including produce, will allow prioritization of samples for testing during investigations and perhaps decrease the time required to identify specific brands or lots. We conducted a series of trials to characterize the sensitivity and specificity of scent detection dogs to accurately identify fecal contamination on raw agricultural commodities (romaine lettuce, spinach, cilantro, and roma tomatoes). Both indirect and direct methods of detection were evaluated. For the indirect detection method, two dogs were trained to detect contamination on gauze pads previously exposed to produce contaminated with feces. For the direct detection method, two dogs were trained to identify fecal contamination on fresh produce. The indirect method did not result in acceptable levels of sensitivity except for the highest levels of fecal contamination (25 g of feces). Each dog had more difficulty detecting fecal contamination on cilantro and spinach than on roma tomatoes. For the direct detection method, the dogs exhibited >75% sensitivity for detecting ≥0.25 g of feces on leafy greens (cilantro, romaine lettuce, and spinach) and roma tomatoes, with sensitivity declining as the amount of feces dropped below 0.025 g. We determined that use of a scent detection dog to screen samples for testing can increase the probability of detecting ≥0.025 g of fecal contamination by 500 to 3,000% when samples with fecal contamination are rare (≤1%).

Full text

Quantifying the Sensitivity of Scent Detection Dogs To Identify
Fecal Contamination on Raw Produce
MELISSA L. PARTYKA,
1
RONALD F. BOND,
1
JEFF FARRAR,
2
ANDY FALCO,
3
BARBARA CASSENS,
4
ALONZA CRUSE,
5
AND EDWARD R. ATWILL
1
*
1
Western Institute for Food Safety and Security, School of Veterinary Medicine, University of California, 1089 Veterinary Medicine Drive, Davis, California
95616;
2
U.S. Food and Drug Administration, Office of Food and Veterinary Medicine, Silver Spring, Maryland 20993;
3
Falco K-9 Academy, 615 Berry
Street, Brea, California 92821;
4
U.S. Food and Drug Administration, Office of Partnerships, Rockville, Maryland 20857; and
5
U.S. Food and Drug
Administration, Office of Regulatory Affairs, Los Angeles District Office, Irvine, California 92620, USA
MS 13-249: Received 17 June 2013/Accepted 4 September 2013
ABSTRACT
Consumption of raw produce commodities has been associated with foodborne outbreaks in the United States. In a recent
Centers for Disease Control and Prevention report outlining the incidence of food-related outbreaks from 1998 to 2008, produce
of all kinds were implicated in 46%of illnesses and 23%of deaths. Methods that quickly identify fecal contamination of foods,
including produce, will allow prioritization of samples for testing during investigations and perhaps decrease the time required to
identify specific brands or lots. We conducted a series of trials to characterize the sensitivity and specificity of scent detection
dogs to accurately identify fecal contamination on raw agricultural commodities (romaine lettuce, spinach, cilantro, and roma
tomatoes). Both indirect and direct methods of detection were evaluated. For the indirect detection method, two dogs were trained
to detect contamination on gauze pads previously exposed to produce contaminated with feces. For the direct detection method,
two dogs were trained to identify fecal contamination on fresh produce. The indirect method did not result in acceptable levels of
sensitivity except for the highest levels of fecal contamination (25 g of feces). Each dog had more difficulty detecting fecal
contamination on cilantro and spinach than on roma tomatoes. For the direct detection method, the dogs exhibited .75%
sensitivity for detecting $0.25 g of feces on leafy greens (cilantro, romaine lettuce, and spinach) and roma tomatoes, with
sensitivity declining as the amount of feces dropped below 0.025 g. We determined that use of a scent detection dog to screen
samples for testing can increase the probability of detecting $0.025 g of fecal contamination by 500 to 3,000%when samples
with fecal contamination are rare (#1%).
Globalization of the world’s food supply, increased
demand for fresh and minimally processed produce, and
direct marketing of raw agricultural commodities have
contributed to new patterns of produce distribution and
foodborne illness associated with the consumption of these
commodities (8, 18). Consumption of raw or minimally
processed produce is increasingly recognized as a vehicle for
transmission of foodborne pathogens such as Escherichia coli
O157:H7, Salmonella, and Listeria (24, 25). From 1996 to
2010, approximately 131 outbreaks of foodborne illness were
associated with fresh produce commodities; 13 of these
outbreaks were linked to tomatoes and 24 were linked to leafy
greens such as lettuce and spinach (8, 31, 32). For many of
these outbreaks, identification of the originating source(s) of
the foodborne pathogen was challenging because of such
factors as a short shelf life and growing period, a long time
between consumption and identification of the outbreak, and
commingled lots of harvested produce. New technologies are
needed that can assist outbreak investigators in the rapid and
accurate identification of the specific food brands or lots
causing the outbreak and the key risk factors and contributing
elements that lead to microbial contamination of foods so that
effective preventive controls can be implemented.
State and federal agencies such as the California
Department of Public Health and the U.S. Food and Drug
Administration (FDA) currently use outbreak investigation
techniques to determine the microbial source and contrib-
uting factors responsible for outbreaks of foodborne illness
(13, 26). This process can involve grower and processor
interviews, collection of invoices and bills of lading for
traceback investigations, in-depth environmental investiga-
tions of preharvest and postharvest processing environ-
ments, and laboratory analyses of the suspect commodity
and associated environmental samples such as irrigation
water, animal scat, and soil amendments (27). When rates of
contamination are low, hundreds or thousands of samples
and resource- and time-intensive microbiological tests may
be needed to have a reasonable chance of detecting the
contaminated source(s). New tools that would enable a risk-
based prioritization of field samples are needed to expedite
the identification and confirmation of suspected brands or
lots of contaminated produce. The more rapidly the source
of contamination and affected commodity is identified, the
* Author for correspondence. Tel: 530-754-2154; Fax: 530-752-7181;
E-mail: ratwill@ucdavis.edu.
6
Journal of Food Protection, Vol. 77, No. 1, 2014, Pages 6–14
doi:10.4315/0362-028X.JFP-13-249
Copyright G, International Association for Food Protection

more quickly a recall can be initiated, thereby reducing
human exposure and the number of ill individuals.
The prevalence of pathogen contamination of produce
is typically very low during both outbreak investigations
and routine industry monitoring (11, 29). Microbial
contamination is often due to fecal contamination of
produce, and methods that can quickly identify samples
most likely to contain fecal contamination should allow
investigators to prioritize samples for preferential testing.
Prioritization can increase the probability of pathogen
detection and identification of pathogen sources. Fecal
material on produce can be difficult to identify visually
when the amount is small, e.g., feces from rodents (14),
from birds (12), and hidden between leaves or diluted in
irrigation or processing water (32). Given that feces and
fecal by-products have specific odors (19), use of olfactive
rather than visual techniques may be a sensitive and specific
method for detection of low levels of fecal contamination on
produce.
Dogs rely on olfaction to detect a wide array of
compounds (1); this fact and their demonstrated trainability
make them the ideal subjects for determining the feasibility
of using scent detection to identify fecal contamination on
produce. Scent detection dogs have been used to detect a
variety of targets, such as cancer (9, 20, 35), nematode
infection (21), gypsy moth egg masses (33), microbial
growth in buildings (15), termite infestation (6), and
pathogens (5). Although historically dogs have been used
to inspect produce for pests (10, 30), to date no information
on the use of scent dogs for detecting fecal contamination on
raw produce has been published. The purpose of this study
was (i) to determine the ability of scent dogs to detect fecal
contamination in ready-to-eat raw produce, (ii) to ascertain
whether produce type and amount of feces has an effect on
detection success and if so to ascertain the thresholds of
detection, and (iii) to clarify whether scent detection dogs
are a reliable tool for prioritizing produce samples for
analytical processing during events such as outbreak
investigations or routine monitoring by industry or regula-
tory agencies.
MATERIALS AND METHODS
Overall study design. We explored two scenarios for how a
scent detection dog might be used to prioritize samples for fecal
detection and microbial testing. The indirect detection method
minimizes possible cross-contamination between the dog and the
produce sample when samples require a high level of biosecurity.
The direct detection method allowed the dogs to directly
investigate a subsample of raw product for fecal contamination.
Dog trials were conducted under University of California, Davis
animal use protocol 15310.
Indirect detection training. Two mixed-breed, female dogs
(dog 1 and dog 2) were trained using scent detection techniques
similar to those for narcotics, bomb, and arson detection (34). The
dogs were selected for trainability and assessed for their
disposition, search and hunt capability, and play drive (34). All
training was conducted by and took place at the Falco K-9
Academy (Brea, CA). The dogs were trained to search for a target
fecal scent associated with a reward such as a chew toy or tennis
ball. The target scent consisted of a 4-ply sterile gauze pad (1 by 1
in. [2.54 by 2.54 cm]; Covidien, Mansfield, MA) saturated with a
fecal slurry (1:3 feces:water) composed of feces from eight
vertebrate species found within or near agricultural fields in the
western United States (dog, cow, horse, black-tailed deer, feral pig,
coyote, Canada goose, and human). Midway through the study,
coyote feces became difficult to obtain and were replaced with
sheep feces. All fecal samples were prescreened for Salmonella
and E. coli O157:H7 using previously described methods; positive
samples were discarded and replaced with samples that were
negative for these pathogens. To construct the composite fecal
slurry, 61 g of feces from each vertebrate species was combined
and homogenized with 1,464 ml of deionized water, making a 1:3
feces:water slurry. The slurry was refrigerated at 4uC and used
within 30 days.
To train the dogs, the saturated pads were inserted into stand-
up Whirl-Pak bags (4 oz [120 ml]; Nasco, Fort Atkinson, WI) and
placed in hollow core cinder blocks (8 by 8 in. [20.3 by 20.3 cm]).
Cinder blocks were placed approximately 1 m apart in a row. The
dogs were guided to each block, allowed to investigate the bag, and
trained using positive reinforcement (chew toy or ball) to sit (the
alert response) when they encountered the target scent. Two
handlers were used during the training, and each handler worked
with both dogs. Training lasted for 3 months. To reduce the
potential for handler bias, both handlers trained both dogs to
search, locate, and alert.
Indirect detection trials. Four types of produce were
selected for the indirect detection trials: whole romaine lettuce,
bunches of spinach, bunches of cilantro, and whole roma tomatoes.
Produce items were washed with tap water and spun dry in a
household salad spinner for 30 s. Romaine lettuce was coarsely
chopped before use to mix inner and outer leaves in each sample.
Weighed samples of romaine lettuce (n~30), spinach (n~30),
and cilantro (n~30) (0.10 kg each) and 0.20 kg of whole roma
tomatoes (n~30) were placed into separate 69-oz (2.1-liter)
Whirl-Pak bags labeled 1 to 30 for each produce type (120 total
bags). For each produce type, 8 of the 30 bags were randomly
selected for treatment. Two treatment bags were spiked with 0.025,
0.25, 2.5, or 25 g of feces so that each amount was replicated
within a produce type. The remaining sample bags (22 per produce
type) were used as negative controls (no feces). Two 4-ply cotton
gauze pads (2 by 2 in. [5.1 by 5.1 cm]; A Plus International Inc.,
Chino, CA) were suspended by string above the produce within
each Whirl-Pak bag (240 total pads). The bags were then sealed for
24 h at room temperature to enable fecal scent transfer from the
produce to the gauze pads without direct contact. The following
day, the two gauze pads were removed with sterile forceps from
each bag and separated, and each pad was placed in a separate
smaller (4-oz) stand-up Whirl-Pak bag with the corresponding
label (produce type and number 1 to 30). The pads were then split
into two equal sets of 120 gauze samples with additional labeling
to be used for trials during one morning and one afternoon session.
This process was repeated with fresh produce, fresh fecal slurry,
and fresh gauze for the second day of the trial.
Four sessions were conducted, with two sessions per day
(morning and afternoon) for 2 days. A session consisted of four
searches by each dog (dog 1 and dog 2) with a single search
involving 30 gauze samples from one produce type (8 exposed
samples and 22 negative controls). Both trials were conducted in
an empty room at the School of Veterinary Medicine (University of
California, Davis) and unfamiliar to the dogs. For each trial, the
small Whirl-Pak bags containing the gauze samples were opened
and placed in holders constructed of 4-in. (10.2-cm) diameter
J. Food Prot., Vol. 77, No. 1 PRODUCE SCENT DETECTION DOGS 7

polyvinyl chloride (PVC) pipe cut to 10-in. (25.4-cm) lengths and
set inside plastic flanges to reduce tipping. Holders were
numbered 1 to 30 and distributed in three rows of 10 pipes each;
within each row, holders were 3 ft (0.9 m) apart, and rows were
10 ft (9 m) apart (Fig. 1A). The dogs, handler, and third party data
recorder were blind to the order of fecal-positive and fecal-negative
gauze samples. The researchers distributed the gauze samples
within the holder array based on a randomly generated number
sequence and then left the room to view the search via remote
video so no visual or behavioral cues would be transmitted to the
handlers, dogs, or data recorder.
The dog was brought into the room, and the handler worked
the dog down each row making sure the dog examined every
holder. When a dog gave a positive alert response (sitting) at a
sample holder, a reward chew toy was given and the handler called
out the holder number to the data recorder. After the first dog had
completed its investigation of the holders, the researchers checked
for any visible oral or nasal secretions left on the gauze sample
bags by the first dog. When secretions were found, the gauze in the
bag was transferred to a fresh bag using sterile forceps. Between
each produce type, holders were emptied and sprayed with canned
air (Uline, Pleasant Prairie, WI) to help remove residual odors. This
process was repeated until all four produce types had been tested
by each dog (one session). Dogs were given 30-min rest periods
between produce types and a 1.5-h rest period between the
morning and afternoon sessions. A second 2-day trial was
conducted 6 weeks later using the same methodology.
Direct detection training. The second method evaluated the
dog’s ability to identify subsamples of raw product contaminated
by different amounts of fecal material directly, without the use of a
gauze intermediary. The same four varieties of mature processed
produce were selected for this portion of the study: whole romaine
lettuce, bunches of spinach and cilantro, and whole roma tomatoes.
Two female, mixed-breed dogs (dog 1 and dog 3) were trained
for 60 days at Falco K-9 Academy to identify direct fecal
contamination in multipurpose specimen storage containers (Fisher
Scientific, Suwanee, GA) filled with produce and spiked with
,2.5 g (high level of contamination) and ,0.025 g (low level of
contamination) of feces (as described above). Dogs were trained
using positive reinforcement techniques as described above. Dog 1
had been utilized during the indirect detection trials, but dog 3 had
not been trained to scent detect before this portion of the study.
Direct detection trials. For the direct detection trials,
produce was washed and spun dry in a household salad spinner
for 30 s. Romaine lettuce was coarsely chopped before use to mix
inner and outer leaves in each sample. Weighed samples (0.1 kg for
leafy greens and 0.2 kg for whole tomatoes) of each type of
produce type (n~30 for each type) were placed into clean
containers labeled with a produce code and numbered 1 to 30.
Eight of the 30 containers per produce type were randomly chosen
to receive treatment application. For the first direct detection trial,
the treatment containers were spiked with 0.0025, 0.025, 0.25, or
2.5 g of feces in duplicate (two treatment containers per fecal
amount per produce type). Following the first direct detection trial,
the largest amount of feces (2.5 g) was dropped from the trials, and
a smaller amount was added so that the treatments consisted
0.00025, 0.0025, 0.025, and 0.25 g of feces in duplicate. The
remaining containers (22 per produce type) were loaded with fresh
produce alone and functioned as negative controls. All containers
were sealed with plastic caps and stored for 12 to 18 h at room
temperature before use in the first session (morning). This process
was repeated for another set of samples using fresh produce, fresh
feces, and clean containers for the second (afternoon) session and
again the following day for the final session (morning) on the
second trial day (180 samples per trial).
A trial consisted of three sessions carried out over 2 days. A
single session consisted of both dogs investigating 30 samples
from each produce type (120 total samples). The 12 PVC holders
used during the direct trials consisted of 64-oz (1.9-liter) specimen
containers screwed 1 ft (0.3 m) apart onto a 12-ft (3.7-m) length of
wooden board (2 by 6 in. [5.1 by 15.2 cm]) and were numbered
from 1 to 12 (Fig. 1B). Feedback from the handlers following the
indirect trials suggested that the dogs could lose focus when
presented with all 30 containers at one time. To mitigate this effect,
samples were presented to the dogs in smaller subsets of 12, 12,
and 6 for each produce type.
Five minutes before the dog entered the trial space five G-in.
(1-cm) holes were drilled into each container cap. The caps were
kept in place to allow for air exchange while preventing direct
contact with the produce. After dog 1 searched the first subset of
12 samples, the containers were examined for oral and nasal
secretions. When secretions were found, the lids were aseptically
wiped down and returned to the container; then dog 3 searched the
same 12 samples. When both dogs had been through the array, the
second subset of samples was placed in the holders and searched
by both dogs, followed by the third subset until all 30 of each
produce type were searched by both dogs. This process was
repeated for each produce type until all four produce types had
been searched (n~120), constituting one session. To avoid
rewarding the dogs for incorrect responses, verbal praise without a
reward chew toy was given after an alert response (sitting).
Training reinforcement was performed between searches of each
produce type by incorporating a nonscored run where the dogs
were guaranteed success, and the handlers were allowed to verbally
FIGURE 1. Schematic layout depicting configuration and spatial
arrangement of sample holders for (A) indirect detection method
trials and (B) direct detection methods trials.
8PARTYKA ET AL. J. Food Prot., Vol. 77, No. 1

praise and reward the dog with a chew toy. Each trial included
three sessions conducted over 2 days. A second trial was conducted
4 weeks later.
Statistical analysis. The primary outcome measure was the
sensitivity, Prob(Az|Fz) and specificity, Prob(A2|F2) for each
dog, where A was the occurrence of an alert (Az) or no alert (A2)
for each sample given that the sample had been exposed to feces
(Fz) or not (F2). A McNemar exact test was performed to
contrast each dog’s performance, with the sensitivity and
specificity evaluated separately (Stata Statistical Software, release
11, StataCorp LP, College Station, TX). Mixed effects logistic
regression (Stata, release 11) was used to determine whether the
amount (grams) of feces and the type of produce were significantly
associated with the odds of detecting fecal contamination in the
produce, adjusted for the possibility of correlated data by treating
session as a group or random effect. Each dog’s performance was
analyzed separately, and the threshold for significance was P#
0.05. The positive predictive value (PPV), defined as the
proportion of true positive alerts (alert when feces were present)
out of all alerts, was calculated for a range of prevalences of fecal
contamination: PPV ~[Prob(Az|Fz)|Prob(Fz)]/[Prob
(Az|Fz)zProb(Fz)zProb(F2)|(Prob(F2)]. To compare
how scent detection dogs might improve the probability of
detecting a produce sample with fecal contamination, we calculated
the ratio of PPV to random sampling of produce for a range of
prevalences of fecal contamination. Random sampling of produce
was assumed to generate an expected probability of detecting a
sample with fecal contamination equivalent to the background
positive prevalence of fecal contamination in the field.
RESULTS
Indirect detection. Through the course of the indirect
trials, each dog was presented with 964 Whirl-Pak bags
containing gauze pads. A small portion (192) of the gauze
pads had been exposed to produce contaminated with feces
(treatment samples), allowing sensitivity to be measured.
The remaining pads (772) had been exposed to nonconta-
minated produce and functioned as negative controls,
allowing for an evaluation of specificity (Table 1). Dogs 1
and 2 had 4 to 5 higher odds of alerting when they
encountered samples exposed to feces compared with
encountering control samples (P,0.001). Dogs 1 and 2
detected the presence of fecal contamination in 30.2 and
16.1%of treatment samples, respectively; this 1.9-fold
difference in sensitivity between the dogs was significant
(P,0.001). The dogs missed detecting fecal contamination
in 70 to 84%of the samples where the gauze had been
exposed to fecal contamination. Although dog 1 was nearly
twice as effective at detecting treatment samples, it also had
a significantly higher prevalence of false-positive responses
(15.0%) than did dog 2 (7.3%)(P,0.001). The ability of
each dog to detect fecal contamination differed by produce
type and level of fecal contamination. Both dogs were
more sensitive to contamination on roma tomatoes than to
contamination on cilantro and spinach (P#0.05) (Table 2).
Using mixed effects logistic regression, the odds that a
dog would alert was significantly associated with produce
type and level of fecal contamination (amount of feces in a
given volume of a 1:3 feces:water slurry) (Table 3). Both
dogs were more likely to detect fecal contamination on roma
tomatoes than on cilantro or spinach (P#0.05; Fig. 2);
there was no significant difference in detection between all
other pairs of produce types. When the level of fecal
contamination exceeded 2.5 g of feces, both dogs were
much more likely to alert to the sample compared with the
odds of an alert response at lower level of fecal
contamination (Fig. 2).
Direct detection. Given the relatively low sensitivity
exhibited by each dog during the indirect detection trials, we
redesigned the trial to determine whether dogs could detect
fecal contamination directly on produce (feces spiked onto
produce) with improved sensitivity. During the direct
detection trials, each dog was presented with 720 containers,
TABLE 1. Relationship between the occurrence of fecal contam-
ination and dog alert response for indirect fecal exposure trials
a
Alert response
Sample treatment
FzF2Total
Dog 1
Az58 116 174
A2134 656 790
Total 192 772 964
Dog 2
Az31 56 87
A2161 716 877
Total 192 772 964
a
Fz, fecal contamination present; F2, no fecal contamination;
Az, dog alert; A2, no dog alert.
TABLE 2. Specificity and sensitivity of scent detection dogs for detecting indirect fecal contamination on produce
Produce type
Specificity (%of 0-g
samples detected)
a
Sensitivity (%of samples detected) to fecal contamination of
b
:
Mean sensitivity (%of
samples detected)
c
0.025 g 0.25 g 2.5 g 25 g
Cilantro 87.8 8.3 25.0 4.2 20.8 14.6
Romaine lettuce 87.6 25.0 8.3 12.5 50.0 24.0
Spinach 89.6 0.0 16.7 8.3 37.5 15.6
Tomato 90.4 25.0 29.2 37.5 62.5 38.5
Overall mean 88.9 14.6 19.8 15.6 42.7 23.2
a
Proportion of noncontaminated samples for which no detection alert was given (data pooled for both dogs). Values provide a measure of
mean specificity for each produce type, i.e., the proportion of false-positive results for each produce type ~12the specificity.
b
Mass of feces in a 1:3 fecal slurry (feces:water).
c
Excludes samples without fecal contamination.
J. Food Prot., Vol. 77, No. 1 PRODUCE SCENT DETECTION DOGS 9

156 of which contained produce contaminated with feces
(treatment samples); the remainder (564) contained nonconta-
minated produce to serve as negative controls (Table 4). Dogs
1 and 3 had 11.1 and 23.6 higher odds of alerting, respectively,
when encountering treatment samples compared with control
samples (P,0.001), i.e., twice to four times higher odds than
obtained for the indirect method. Dog 1 and dog 3 detected the
presence of fecal contamination in 42.3 and 38.5%of
treatment samples, respectively, although this 1.1-fold differ-
ence in sensitivity was not significant (P~0.33) (Table 4).
With respect to false-positive results, dog 1 was significantly
more likely (6.2%) to incorrectly alert in the presence of a
control sample than was dog 3 (2.7%)(P~0.002) (Table 4).
The odds that a dog would alert when directly
examining contaminated produce was significantly associ-
ated with the level of fecal contamination (Tables 5 and 6)
but not with produce type (P.0.05). As with the indirect
trials, mixed effects logistic regression was used to model
the probability that a dog would positively respond to a
sample as a function of the level of fecal contamination
(amount of feces in a given volume of a 1:3 feces:water
slurry) and produce type. As the level of fecal contamination
increased above 0.025 g, the probability of an alert quickly
increased to achieve 75%probability of detection at 0.25 g
and almost 100%at 2.5 g (Fig. 3).
Using Dog 3 as an example of alert accuracy, the PPV
(proportion of alerts that correctly detect produce contam-
inated with feces) was #25%when the background
prevalence of contaminated produce was #1%and the
level of fecal contamination was low but increased to 50 to
75%when the prevalence of contamination exceeded 5%
for samples with .0.025 g of feces (Fig. 4). Although these
PPVs are quite low for field situations that have infrequent
episodes fecal contamination, the likelihood of finding
FIGURE 2. Indirect detection method trials. Results from mixed effects
logistic regression for the ability of two dogs to detect fecal contamination
on produce as a function of amount of feces (grams), stratified by produce
type (cilantro, romaine lettuce, spinach, and tomato). (A) Dog 1; (B) dog 2.
TABLE 3. Mixed effects logistic regression models for the odds of a positive alert response for samples with indirect fecal exposure
a
Analysis Coefficient Odds ratio Pvalue 95%confidence interval
Dog 1
Intercept 21.91 0.003
Produce type
Cilantro
b
0.0 1.0
Romaine lettuce 0.71 2.02 0.31 0.52, 7.93
Spinach 20.48 0.62 0.57 0.12, 3.26
Tomato 1.47 4.36 0.05 1.01, 18.8
Feces (g)
c
0.051 1.05 0.02 1.01, 1.1
Dog 2
Intercept 23.69 0.001
Produce type
Cilantro
b
0.0 1.0
Romaine lettuce 1.26 3.54 0.19 0.52, 23.9
Spinach 0.87 2.39 0.42 0.29, 19.4
Tomato 2.68 14.64 0.009 1.96, 109.4
Feces (g) 0.081 1.08 0.003 1.03, 1.14
a
Data were adjusted for repeated measures from four trials. Data from each dog were analyzed separately.
b
Referent category (odds ratio ~1.0).
c
Mass of feces in a 1:3 fecal slurry (feces:water).
10 PARTYKA ET AL. J. Food Prot., Vol. 77, No. 1

produce samples with fecal contamination increases 5- to
30-fold over random sampling by using a dog to first screen
all field samples for the presence of feces (Fig. 5). In other
words, for samples with $0.025 g fecal contamination, the
probability of collecting samples of produce with fecal
contamination is 5- to 30-fold higher (500 to 3,000%) when
using a dog than when randomly selecting produce samples
across a field, as is sometimes done during investigations.
DISCUSSION
The diagnostic value of using scent detection dogs to
identify produce contaminated with feces relies on the ability
of these dogs to detect contamination at levels below the
limits of human olfactory and/or visual detection. Currently,
few accepted validation tests for the scent detection ability of
dogs have been developed, and none are available for use
with specialized search dogs (22). Although several studies
have documented the ability of scent detection dogs to detect
a variety of compounds at .75%accuracy at various odor
intensities (7, 15, 20), few studies have quantified the full
range of sensitivity, specificity, and PPV for a range of
contamination scenarios, as was done in this study.
A primary goal of this study was to characterize the
sensitivity of a trained scent detection dog to detect fecal
contamination on ready-to-eat produce commodities as a
function of level of contamination. In particular, we sought
to determine at what level of fecal contamination dogs
excessively fail (%20%accuracy) or adequately succeed
(&50%accuracy) and whether these thresholds are
influenced by produce type such that certain fruits or
vegetables mask or confound fecal scent detection. To
maintain sample integrity, we evaluated an indirect fecal
detection method whereby gauze pads were first exposed to
produce contaminated with feces, then these pads were
examined by a scent detection dog in a separate facility. In
this manner, the samples of produce were not exposed to the
dog, which prevented cross-contamination between the dog
and the sample. Unfortunately, this protocol did not result in
acceptable levels of sensitivity for any but the highest levels
of fecal contamination (25 g of feces); each dog had more
difficulty detecting fecal contamination on cilantro and
spinach than on roma tomatoes regardless of the level of
fecal contamination (Fig. 2). In addition, the two dogs
alerted on about 10%of the produce samples that had no
fecal contamination (presumably false-positive responses),
emphasizing the need to train dogs both to detect the odor of
feces when present and to not give an alert when the odor of
feces is not present.
Allowing dogs to directly evaluate the samples of
produce dramatically improved their ability to detect fecal
contamination in excess of 0.025 g (Fig. 3). The two dogs
exhibited 76 and 86%sensitivity, respectively, for detecting
$0.25 g of fecal contamination in 0.1-kg samples of leafy
greens (cilantro, romaine lettuce, or spinach) and 0.2-kg
samples of roma tomatoes. To put this level of fecal
contamination into perspective, a wide variety of domestic
and wild animals defecate about 2 to 8%of their body
weight per day (3, 4). For example, adult California ground
squirrels (Spermophilus beecheyi) produce about 10 to 20 g
of feces per day (2), and this species can be quite common
in produce fields of the western United States. Based on the
results of this study, scent detection dogs might be able to
detect a little as 1%of a ground squirrel’s daily fecal load if
deposited on, for example, spinach or cilantro or after foliar
irrigation when feces or scat are deposited in furrows and
along beds of leafy greens.
The PPV for correctly detecting fecal contamination is
highly dependent on the overall prevalence of fecal
contamination in these produce commodities (Fig. 4). For
example, the accuracy of fecal detection is relatively poor
when the prevalence of contamination is #1%of samples
TABLE 4. Relationship between the occurrence of fecal contam-
ination and dog alert response for direct fecal exposure trials
Alert response
Sample treatment
FzF2Total
Dog 1
Az66 35 101
A290 529 619
Total 156 564 720
Dog 3
Az60 15 75
A296 549 645
Total 156 564 720
a
Fz, fecal contamination present; F2, no fecal contamination;
Az, dog alert; A2, no dog alert.
TABLE 5. Specificity and sensitivity of scent detection dogs for detecting direct fecal contamination on produce
Produce type
Specificity (%of 0-g
samples detected)
a
Sensitivity (%of samples detected) to fecal contamination of
b
:
Mean sensitivity (%of
samples detected)
c
0.00025 g 0.0025 g 0.025 g 0.25 g 2.5 g
Cilantro 91.5 5.6 16.7 38.9 72.2 100.0 30.8
Romaine lettuce 97.2 11.1 27.8 33.3 77.8 100.0 34.6
Spinach 94.7 5.6 22.2 16.7 88.9 100.0 30.8
Tomato 98.9 5.6 0.0 55.6 94.4 83.3 35.9
Overall mean 95.6 6.9 16.7 36.1 83.3 95.8 40.4
a
Proportion of noncontaminated samples for which no detection alert was given (data pooled for both dogs). Values provide a measure of
mean specificity for each produce type, i.e., the proportion of false-positive results for each produce type ~12the specificity.
b
Mass of feces in a 1:3 fecal slurry (feces:water).
c
Excludes samples without fecal contamination.
J. Food Prot., Vol. 77, No. 1 PRODUCE SCENT DETECTION DOGS 11

regardless of the level of fecal contamination per sample.
However, as the prevalence reaches 5%then the corre-
sponding PPV can exceed 50%, depending on the level of
fecal contamination. Nevertheless, when the prevalence of
fecal contamination is very low (#1%of samples) the scent
detection dogs dramatically improve the probability of
finding such contaminated produce samples. Using random
sampling as the comparison procedure, a protocol that uses
a fecal scent detection dog to first screen all produce samples
and then test only those to which the dog alerted can increase
the probability of detecting contaminated produce by up to
3,000%(Fig. 5), depending on the background prevalence of
fecal contamination in the field. When one of the goals for
produce sampling is to identify fecal contamination in an
effort to isolate a foodborne pathogen, then the success of
traceback investigations can be substantially improved when
the adulterating pathogen is isolated and DNA fingerprinted
for comparison to other clinical and environmental isolates.
This method of direct detection of contamination eliminated
the masking or inhibition observed for cilantro and spinach
during the indirect detection trials; the sensitivity of the dogs
to fecal scent was not significantly different across the four
produce types (cilantro, romaine lettuce, spinach, and roma
tomatoes).
The accuracy of the dogs in this study was lower than
that reported by others and may have been affected by
several factors such as differences between studies regard-
ing the target odor, breed of dog, attention span, search
fatigue, training paradigm, and handler bias (16, 17, 36).
Differences among breeds and among individual dogs is
expected (27, 36), just as mechanical detection devices of
different models or from different manufacturers might
differ in performance. According to the National Detector
Dog Manual (30), scent detection dogs can search for 40 to
60 min before significant decreases in proficiency occur,
and rest periods can mitigated this decrease. However, the
dogs in this study were exposed to a search paradigm
different from that used in a majority of published studies
(16, 23, 25). Although dogs were given short rests between
produce types and longer rests between sessions, the type of
searching required making a ‘‘yes’’ or ‘‘no’’ decision at
each sample container, which might have adversely affected
the dogs’ alert accuracy. Lit and Crawford (16) and Schoon
(23) found that the training paradigm greatly affected scent
detection dog performance. Different training techniques
may be necessary to train dogs to search a large array of
samples in a systematic manner while mitigating potential
fatigue.
TABLE 6. Mixed effects logistic regression models for the odds of a positive alert response for samples with direct fecal exposure
a
Analysis Coefficient Odds ratio Pvalue 95%confidence interval
Dog 1
Intercept 3.74 ,0.001
Feces (g)
b
3.45 31.62 ,0.001 4.68, 213.5
Feces
2
(g) 0.47 1.60 0.04 1.03, 2.49
Dog 2
Intercept 1.74 ,0.001
Feces (g) 2.22 5.68 ,0.001 3.28, 9.85
a
Data were adjusted for repeated measures from six trials. Data from each dog were analyzed separately. Produce type was not a significant
variable (P.0.05).
b
Mass of feces in a 1:3 fecal slurry (feces:water).
FIGURE 3. Direct detection method trials. Results from mixed
effects logistic regression for the ability of two dogs to detect fecal
contamination on produce (cilantro, romaine lettuce, spinach,
and tomato) as a function of amount of feces (grams), stratified by
dog.
FIGURE 4. Positive predictive value associated with dog 3 for
detecting fecal contamination of produce (cilantro, romaine
lettuce, spinach, and tomato) as a function of amount of feces
(grams), modeled at different hypothetical prevalences of fecal
contamination among samples.
12 PARTYKA ET AL. J. Food Prot., Vol. 77, No. 1

The dogs exhibited a tendency to return to containers
where they had previously been rewarded and, in the case of
dog 1, to alert at the end of a row of samples. This latter
behavior frequently confounded the handlers because they
were less inclined to call out a positive alert on an end
container even though the container occasionally contained
a positive sample. This fact and other observed instances of
miscommunication between handler and dog emphasize the
importance of handler-dog relationships to the success of
this and other studies (23, 28). Lit et al. (17) pointed out the
effect of handler beliefs on dog performance, showing that
the handler’s beliefs about the presence (or absence) of the
target odor may influence their handling of the dog and their
willingness to call an alert. In our study, the handlers were
completely blinded to the location and number of treatment
samples in the sampling array.
In conclusion, scent detection dogs appear capable of
directly detecting low levels of fecal contamination on
romaine lettuce, cilantro, spinach, and roma tomatoes and
thereby elevate the probability that an investigator, grower,
or processor can successfully detect produce samples with
fecal contamination before harvest. For example, the use of
scent detection dogs to identify contaminated produce using
a Z-pattern sampling approach in the field, as is used in
pesticide residue sampling, may increase the effectiveness
of pesticide monitoring methods. If we assume that most
of the microbial contamination on raw produce is the
consequence of fecal contamination in the field, then the use
of scent detection dogs will allow us to prioritize produce
samples for analytical testing and thereby optimize the
detection of both feces and the associated microbial
pathogens that so often accompany fecal contamination.
ACKNOWLEDGMENTS
The authors extend our appreciation to Andrew Gordus (California
Department of Fish and Wildlife) and Thomas McCoy (T.J. McCoy and
Associates, Kansas City, MO) for their invaluable assistance. The authors
also thank Dr. Lisa Lit (Department of Animal Sciences, University of
California, Davis) for her valuable feedback and advice and Dr. Brenda
McCowan (Population Health and Reproduction, University of California,
Davis) for continual support and guidance. This work was funded by a
grant from the California Department of Food and Agriculture, Specialty
Crop Block Grant Program, with additional funding from the U.S. Food and
Drug Administration (project U01-003-572).
REFERENCES
1. Ashton, E. H., J. T. Eayrs, and D. G. Moulton. 1957. Olfactory acuity
in the dog. Nature 179:1069–1070.
2. Atwill, E. R., S. M. Camargo, R. Phillips, L. H. Alonso, K. W. Tate,
W. A. Jensen, J. Bennet, S. Little, and T. P. Salmon. 2001.
Quantitative shedding of two genotypes of Cryptosporidium parvum
in California ground squirrels (Spermophilus beecheyi). Appl.
Environ. Microbiol. 67:2840–2843.
3. Atwill, E. R., X. Li, D. Grace, and V. P. J. Gannon. 2012. Zoonotic
waterborne pathogen loads in livestock, p. 75–116. In J. Bartram, R.
Bos, and V. Gannon (ed.), Animal waste, water quality and human
health. World Health Organization, IWA Publishing, London.
4. Atwill, E. R., R. Phillips, and F. Rulofson. 2002. Estimating
environmental loading rates of the waterborne pathogenic protozoa,
Cryptosporidium parvum, in certain domestic and wildlife species in
California, p. 241–243. In R. M. Timm and R. H. Schmidt (ed.),
Proceedings of the 20th Vertebrate Pest Conference, University of
California, Davis.
5. Bomers, M. K., M. A. van Agtmael, H. Luik, M. C. van Veen, C. M.
J. E. Vandenbroucke-Grauls, and Y. M. Smulders. 2012. Using a
dog’s superior olfactory sensitivity to identify Clostridium difficile
in stools and patients: proof of principle study. Br. Med. J. 345:
e7396.
6. Brooks, S. E., F. M. Oi, and P. G. Koehler. 2003. Ability of canine
termite detectors to locate live termites and discriminate them from
non-termite material. J. Econ. Entomol. 96:1259–1266.
7. Cablk, M. E., and J. S. Heaton. 2006. Accuracy and reliability of dogs
in surveying for desert tortoise (Gopherus agassizii). Ecol. Appl. 16:
1926–1935.
8. Centers for Disease Control and Prevention. 2012. Foodborne
diseases active surveillance network (FoodNet): FoodNet surveillance
report for 2011 (final report). Centers for Disease Control and
Prevention, Atlanta.
9. Cornu, J.-N., G. Cancel-Tassin, V. Ondet, C. Girardet, and O.
Cussenot. 2011. Olfactory detection of prostate cancer by dogs
sniffing urine: a step forward in early diagnosis. Eur. Urol. 59:
197–201.
10. Engeman, R. M., D. V. Rodriquez, M. A. Linnell, and M. E. Pitzler.
1998. A review of the case histories of the brown tree snakes (Boiga
irregularis) located by detector dogs on Guam. Int. Biodeter.
Biodegrad. 42:161–165.
11. Gajraj, R., S. Pooransingh, J. I. Hawker, and B. Olowokure. 2012.
Multiple outbreaks of Salmonella Braenderup associated with
consumption of iceberg lettuce. Int. J. Environ. Health Res. 22:
150–155.
12. Gorski, L., C. T. Parker, A. Liang, M. B. Cooley, M. T. Jay-Russell,
A. G. Gordus, E. R. Atwill, and R. E. Mandrell. 2011. Prevalence,
distribution, and diversity of Salmonella enterica in a major produce
region of California. Appl. Environ. Microbiol. 77:2734–2748.
13. Guzewich, J. J., and P. A. Salsbury. 2000. FDA’s role in traceback
investigations for produce. Food Saf. Mag. 6:8.
14. Jay-Russell, M., J. Montfort, Y. Liu, S. Huang, L. Gorski, M. Cooley,
R. Mandrell, J. Wheeler, D. Reis, X. Li, and E. R. Atwill. 2012.
Zoonotic risks from amphibians and reptiles, p. 113–116. In R. M.
Timm (ed.), Proceedings of the 25th Vertebrate Pest Conference,
University of California, Davis.
15. Kauhanen, E., M. Harri, A. Nevalainen, and T. Nevalainen. 2002.
Validity of detection of microbial growth in buildings by trained
dogs. Environ. Int. 28:153–157.
16. Lit, L., and C. A. Crawford. 2006. Effects of training paradigms on
search dog performance. Appl. Anim. Behav. Sci. 98:277–292.
17. Lit, L., J. Schweitzer, and A. Oberbauer. 2011. Handler beliefs affect
scent detection dog outcomes. Anim. Cogn. 14:387–394.
FIGURE 5. Relative increase in the probability of correctly
identifying produce samples contaminated with feces following
initial screening by scent detection dogs compared with random
sampling of produce as a function of amount of feces (grams),
stratified by hypothetical prevalences of fecal contamination
among samples.
J. Food Prot., Vol. 77, No. 1 PRODUCE SCENT DETECTION DOGS 13

18. Lynch, M., R. Tauxe, and C. Hedberg. 2009. The growing burden of
foodborne outbreaks due to contaminated fresh produce: risks and
opportunities. Epidemiol. Infect. 137:307.
19. Mackie, R. I., P. G. Stroot, and V. H. Varel. 1998. Biochemical
identification and biological origin of key odor components in
livestock waste. J. Anim. Sci. 76:1331–1342.
20. McCulloch, M., T. Jezierski, M. Broffman, A. Hubbard, K. Turner,
and T. Janecki. 2006. Diagnostic accuracy of canine scent detection
in early- and late-stage lung and breast cancers. Integr. Cancer Ther.
5:30–39.
21. Richards, K. M., S. J. Cotton, and R. M. Sandeman. 2008. The use of
detector dogs in the diagnosis of nematode infections in sheep feces.
J. Vet. Behav. 3:25–31.
22. Rooney, N. J., S. A. Gaines, J. W. S. Bradshaw, and S. Penman.
2007. Validation of a method for assessing the ability of trainee
specialist search dogs. Appl. Anim. Behav. Sci. 103:90–104.
23. Schoon, G. A. A. 1996. Scent identification lineups by dogs (Canis
familiaris): experimental design and forensic application. Appl. Anim.
Behav. Sci. 49:257–267.
24. Sivapalasingam, S., C. R. Friedman, L. Cohen, and R. V. Tauxe.
2004. Fresh produce: a growing cause of outbreaks of foodborne
illness in the United States, 1973 through 1997. J. Food Prot. 67:
2342–2353.
25. Smith, D. A., K. Ralls, A. Hurt, B. Adams, M. Parker, B. Davenport,
M. C. Smith, and J. E. Maldonado. 2003. Detection and accuracy
rates of dogs trained to find scats of San Joaquin kit foxes (Vulpes
macrotis mutica). Anim. Conserv. 6:339–346.
26. Sobel, J., P. M. Griffin, L. Slutsker, D. L. Swerdlow, and R. V.
Tauxe. 2002. Investigation of multistate foodborne disease outbreaks.
Public Health Rep. 117:8.
27. Svartberg, K., and B. Forkman. 2002. Personality traits in the
domestic dog (Canis familiaris). Appl. Anim. Behav. Sci. 79:133–155.
28. Szetei, V., A
´. Miklo´si, J. Topa´ l, and V. Csa´nyi. 2003. When dogs
seem to lose their nose: an investigation on the use of visual and
olfactory cues in communicative context between dog and owner.
Appl. Anim. Behav. Sci. 83:141–152.
29. U.S. Department of Agriculture. 2011. Microbiological data program.
Progress update and 2009 data summary. U.S. Department of
Agriculture, Agricultural Marketing Service, Washington, DC.
30. U.S. Department of Agriculture. 2012. National detector dog manual.
U.S. Government Printing Office, Washington, DC.
31. U.S. Food and Drug Administration. 1998. Guide to minimize
microbial food safety hazards for fresh fruits and vegetables. U.S.
Government Printing Office, Washington, DC.
32. Wachtel, M. R., L. C. Whitehand, and R. E. Mandrell. 2002.
Association of Escherichia coli O157:H7 with preharvest leaf lettuce
upon exposure to contaminated irrigation water. J. Food Prot. 65:
18–25.
33. Wallner, W. E., and T. L. Ellis. 1976. Olfactory detection of gypsy
moth pheromone and egg masses by domestic canines. Environ.
Entomol. 5:183–186.
34. Wasser, S. K., B. Davenport, E. R. Ramage, K. E. Hunt, M. Parker, C.
Clarke, and G. Stenhouse. 2004. Scat detection dogs in wildlife
research and management: application to grizzly and black bears in the
Yellowhead ecosystem, Alberta, Canada. Can. J. Zool. 82:475–492.
35. Willis, C. M., S. M. Church, C. M. Guest, W. A. Cook, N. McCarthy,
A. J. Bransbury, M. R. T. Church, and J. C. T. Church. 2004.
Olfactory detection of human bladder cancer by dogs: proof of
principle study. Br. Med. J. 329:712.
36. Wilsson, E., and P.-E. Sundgren. 1997. The use of a behaviour test
for the selection of dogs for service and breeding. I. Method of testing
and evaluating test results in the adult dog, demands on different
kinds of service dogs, sex and breed differences. Appl. Anim. Behav.
Sci. 53:279–295.
14 PARTYKA ET AL. J. Food Prot., Vol. 77, No. 1