Home

Search

Collections

Journals

About

Contact us

My IOPscience

Study of the art: canine olfaction used for cancer detection on the basis of breath odour. Perspectives and limitations

This content has been downloaded from IOPscience. Please scroll down to see the full text. 2015 J. Breath Res. 9 027001 (http://iopscience.iop.org/1752-7163/9/2/027001) View the table of contents for this issue, or go to the journal homepage for more

Download details: IP Address: 130.133.8.114 This content was downloaded on 07/07/2015 at 19:47

Please note that terms and conditions apply.

J. Breath Res. 9 (2015) 027001

doi:10.1088/1752-7155/9/2/027001

topical review

received

27 November 2014

Study of the art: canine olfaction used for cancer detection on the basis of breath odour. Perspectives and limitations

re vised

14 February 2015 accep ted for publication

5 March 2015 published

6 May 2015

Tadeusz Jezierski 1, Marta Walczak1, Tomasz Ligor2, Joanna Rudnicka2 and Bogusław Buszewski2 1

Department of Animal Behaviour, Institute of Genetics and Animal Breeding of Polish Academy of Sciences, Jastrzębiec, O5-552 Magdalenka, Poland 2 Chair of Environmental Chemistry and Bioanalytics, Faculty of Chemistry, Nicolaus Copernicus University, Gagarina Str.7, 87–100 Toruń, Poland E-mail: [email protected] Keywords: canine olfaction, cancer detection, breath odour

Abstract Experimental studies using trained dogs to identify breath odour markers of human cancer, published in the recent decade, have been analyzed and compared with the authors’ own results. Particular published studies differ as regards the experimental setup, kind of odour samples (breath, urine, tumor tissue, serum), sample collection methods, dogs’ characteristics and dog training methods as well as in results presented in terms of detection sensitivity and specificity. Generally it can be stated that trained dogs are able to distinguish breath odour samples typical for patients with lung cancer and other cancers from samples typical for healthy humans at a ‘better than by chance’ rate. Dogs’ indications were positively correlated with content of 2-pentanone and ethyl acetate (r = 0.97 and r = 0.85 respectively) and negatively correlated with 1-propanol and propanal in breath samples (r = −0.98 and −0.87 respectively). The canine method has some advantages as a potential cancer-screening method, due to its non-invasiveness, simplicity of odour sampling and storage, ease of testing and interpretation of results and relatively low costs. Disadvantages and limitations of this method are related to the fact that it is still not known exactly to which chemical compounds and/or their combinations the dogs react. So far it could not be confirmed that dogs are able to sniff out early preclinical cancer stages with approximately the same accuracy as already diagnosed cases. The detection accuracy may vary due to failure in conditioning of dogs, decreasing motivation or confounding factors. The dogs’ performance should be systematically checked in rigorous double-blind procedures. Recommendations for methodological standardization have been proposed.

1. Introduction Cancer is one of the most common causes of disease death worldwide. The commonest form of cancer is lung cancer, which currently accounts for approximately 5.9% of all deaths in high-income countries and 2.4% of deaths globally [1]. High cancer mortality rates are primarily due to late diagnosis. This primarily concerns lung cancer, for which no good screening methods exist so far. Late diagnosis of cancer is mainly due to the fact that early stages are accompanied by no characteristic symptoms, or symptoms are unspecific and thus easily ignored or overlooked. Therefore a regular screening for early cancer diagnosis that facilitates a successful therapy is a promising way of reducing the mortality rate. © 2015 IOP Publishing Ltd

Low-dose CT scanning has been demonstrated to be a more sensitive tool than chest radiography for the detection of early stage lung cancer [2–4] and the ability of low-dose CT scanning to detect lung cancer at an early stage has been demonstrated in a large study [5]. This has led many experts to advocate widespread lung-cancer screening with this technique, despite the fact that the expense and disadvantages of the method. In high-resolution imaging methods, false-positive diagnosis due to unsatisfactory specificity can result in unnecessary fear for the patient and necessitate other invasive procedures (e.g. biopsies). In addition, a negative impact of multiple CT scans on human health from large doses of radiation cannot be excluded [6, 7]. Although screening methods are being improved, the ultimate diagnosis of cancer and the precondition

J. Breath Res. 9 (2015) 027001

T Jezierski et al

for starting therapy is based on invasive biopsy and histopathological examination, mostly done after a positive outcome from one or more imaging screening methods. Finding new, non-invasive, painless, easily accessible screening techniques that would be complementary to or could replace the imaging methods, and facilitate early diagnosis, is a desirable goal. In the last two decades, a new approach to cancer screening is being studied. This approach is based on an assumption that aberrant protein synthesis and changed metabolisms in cancer cells produce volatile organic compounds (VOC) that are likely to have distinctive odours detectable by highly sensitive analytical devices [8]. Numerous studies to assess VOC characteristic for cancer have been conducted [9]. During the past 10 years, more than 100 volatile biomarkers have been suggested as being related to cancer [10]. Applying gas chromatography and mass spectrometry (GC/MS) for cancer screening in oncology practice may be problematic, not only because of the sampling procedure but also due to difficulties in interpretation of the results. A variable number of VOCs may be identified in breath samples from patients who have been diagnosed with cancer. Certain compounds may be present in different combinations and quantities. Therefore, employing a single VOC as a cancer marker seems useless [11–13]. Also, while so-called ‘electronic noses’ based on a array of nanosensors are intensively being developed, they have yet to meet clinical expectations [14–16]. In parallel, a low-tech approach using canines’ sense of smell to detect cancer odours was reported in scientific literature. Sniffer dogs’ employment would have some advantages when compared with contemporary analytical VOC-identification methods such as chromatography or mass spectrometry (GC/MS). Dogs’ mobility enables the detection in different sites outside a laboratory. A trained dog’s response to a detected odour (usually sitting or laying down) is of a binary character—i.e. is a clear-cut yes/no response, which makes interpretation of results much easier. In order to make use of canines for practical cancer screening in human populations, the validity and reliability of canine response towards odour samples taken from patients with cancer disease would have to be assessed experimentally. It must be stressed, however, that assessing the ability of dogs to distinguish odour samples of fully diagnosed cancer cases from those of healthy people would not be an ultimate goal. Using sniffer dogs for cancer screening and prophylaxes would have good prospects, if it could be confirmed that dogs are able to alert early pre-clinical stages of cancer. Studies using canines for cancer detection were initially triggered by a reports published in the medical journal The Lancet involving cases of spontaneous detection of malignant melanoma by untrained dogs [17]. In the first report a case was described of a patient who first became aware of the lesion on her 2

thigh because her dog constantly sniffed at it, frequently spending several minutes a day over several months sniffing intently at the lesion, even through the patient’s trousers—and on one occasion trying to bite off the lesion. A similar scenario with another dog was described 12 years later [18], adding a detail that the dog had shown no interest in the area on the owner’s leg since excision of the lesion. In 2003 a conference was organized by Dr John Church in Saunderton (UK) to discuss studies of using dogs for cancer detection, and training of such dogs. This conference gave rise to more systematic experimental studies. In the period 2004–14 at least 12 peer-reviewed papers were published in medical and behavioural scientific journals presenting experimental results of cancer detection by trained dogs [19–30]. Aside from this, five reviews [31–35] and one theoretical evaluation of hypothesis [36] have also been published. It is difficult to directly compare the results of experimental studies on canine detection of cancer conducted so far, since these studies differed methodically in many aspects. Different forms of cancer investigated in canine training studies have included: melanoma [19, 27, 37], bladder cancer [20], ovarian cancer [23, 28], breast cancer [21, 22, 27, 37], prostate cancer [22, 25, 30], colorectal cancer [24] and lung cancer [21, 26, 27, 29, 37]. Accordingly, different kinds of odour samples were used, for example: cancer tissue scraps [19, 23], urine [20, 22, 25, 29, 30], breath and feces [24], breath and urine [29], blood plasma [28] or only breath [21, 26, 27, 37]. In this review we focus only on cancer detection by trained canines on the base of breath odour, however, studies on other kinds of odour samples also have been mentioned.

2. Dogs Individual characteristics of dogs have to be considered as an important factor, playing a role in dogs’ trainability and performance. Several different breeds and crossbreds have been used for cancer detection on the base of breath odour (table 1), with the most common breed being Labrador retrievers. The number of dogs used in reviewed studies ranged from one to five [21, 24]. Only in two publications [21, 27] was information given on the pre-selection of dogs for cancer detection training—most studies did not give any details of how the dogs were selected (table 1). Generally, it could be supposed that the dogs represented individuals that were available, or were preferred by the trainers. Few details are given in the literature as to the methods of the dog training, except for one study on the impact of individual training parameters [27]. In some studies no data was given on the duration of dog training (table 1). When such information was given, there was substantial variation in the duration of the

J. Breath Res. 9 (2015) 027001

T Jezierski et al

Table 1.  Dog number, dog breeds and training details for lung cancer or breast/colorectal cancer detection on the base of breath odour.

Authors

Number of dogs

Pre-selection of dogs for the training

McCulloch et al [21]

5

out of 13

Sonoda et al [24]

1

Labrador retriever water rescue training + cancer training for X months

Ehmann et al [26]

4

German shepherd, Australian shepherd, Labrador retriever cancer training by a professional trainer for X months

Walczak et al [27]

3

Amundsen et al [29]

4

Dog breeds Training details Labrador retriever, Portuguese water dog cancer training 4–5 times/week for 3 months

out of 6

Labrador retriever, mix cancer training for 9 months Belgian Shepherd, Border collie, Dachshund, Rottweiler scent tracking + cancer training 2–4 times / week for X months

training—ranging from 3 months [21] to 9 months [27, 37]. Some dogs underwent other special training prior to the cancer detection training—e.g. water rescue training [24] or scent tracking [29]. Most dogs were naïve at the beginning of cancer training (table 1). In some cases significant individual differences in dogs’ performance at cancer detection were reported [27, 37], both in correct and false responses. On the other hand McCulloch [21] found no statistically significant differences in detection accuracy between five dogs. Four dogs used by Amundsen et al [29] showed similar detection accuracy, and no information about significance of individual differences was given. In the study of Ehmann et al [26] the hit ratio for four individual dogs ranged from 68 to 84%—no information was given whether these differences were statistically significant.

3.  Experimental setup Out of five studies on lung cancer detection by dogs, in three studies an ‘odour lineup’ has been applied and in two studies the odour samples were arranged in a circle (table 2). The number of odour samples in the lineup or circle varied between five and six—in most cases there were five odours, out of which one sample was a target (cancer) odour and the others were healthy controls. In one study there was a variable number (zero to six) of cancer and control odour samples (table 2). Trials consisted in sniffing of all samples and ‘alerting’ the sample containing target odour that matched the training target odour. Dogs’ alerting behavior depended on the training and dog’s preference but usually consisted of sitting or lying down in front of the target sample. During work in a lineup or circle, the following reactions of the dog are possible: a correct indication of the target sample; a false alert to a decoy (control sample); a ‘miss’, i.e. not indicating the target sample; and a ‘hesitation’, i.e. an incomplete indication either of the target or the control sample. Two criteria could be used for recording correct indications by dogs during a trial: Criterion I: a yes/no response toward each sniffed sample in the lineup. In a single trial a dog could make 3

Table 2.  Experimental setup applied in studies on cancer detection on the base of breath samples. Authors

Experimental setup

McCulloch et al [21]

Lineup of 5 samples (1 cancer + 4 controls)

Sonoda et al [24]

Lineup of 6 samples (1 cancer + 5 controls)

Ehmann et al [26]

Circle of 5 samples (1 cancer + 4 controls or COPD)

Walczak et al [27]

Lineup of 5 samples (1 cancer + 4 controls)

Amundsen et al [29]

Circle 6 samples (variable number: 0–6 cancer + 6–0 controls)

both a correct indication and a false alert. For this criterion the probability of a correct response by chance in any single trial is 50% and does not depend on the number of scent samples presented or sniffed by the dog in the lineup. Criterion II: the dog choosing exclusively the one target sample (or more, if more than one target sample was available in the lineup) out of all sniffed samples, without any false alerts or hesitations. In this approach the number of odour samples in the lineup or circle plays a role for the probability of correct indications by chance—i.e. the more samples sniffed before indication the lower probability of correct indication by chance alone. For example, if five samples in the lineup had been sniffed, and only one is a target, the probability of correct indication in one trial is 20%.

4.  Exemplification of the dog training procedure The training of dogs to identify target odour samples (cancer odourous markers) in a lineup is reward-based operant conditioning—i.e. producing an association between performing a behavioral response (in this case sitting down in front of the target sample) and a reward in form of a piece of food or a favorite toy to play with. To produce such association in a naïve dog, the training was divided into three phases [27, 37]. During training phase I the dogs were trained to indicate by sitting down in front of the target sample

J. Breath Res. 9 (2015) 027001

T Jezierski et al

that contained breath odour of a cancer patient coupled with food scent. The target sample was placed randomly in the lineup amidst four other samples that contained blank (sterile) insertions from the sampling tubes. In phase II the food scent was removed from the target sample, which now contained cancer pattern odour only. To pass from this to the next training phase, a dog needed to perform a minimum of 40 faultless indications in at least 100 trials. In training phase III the four blank samples were gradually replaced by control samples taken from healthy volunteers—otherwise the training procedure remained the same as in phase I and II. The majority of trails were single blind—i.e. the dog handler who rewarded the dog was blind to the location of the target sample in the lineup. The experimenter, who was not visible to the handler and dog, and observed the trial on a monitor, knew the location of the target sample in the lineup. Immediately after a correct indication by the dog, the experimenter gave an acoustic signal for the handler by activating a clicker and upon this signal the handler rewarded the dog with a small piece of food. The following responses by the dog were recorded: (a) Correct positive indications (CP), when a dog indicated the target sample (b) Correct negative response (CN), when a dog refrained from indicating a control sample (c) False positive alert (FA), an indication of a blank or control sample (d) False negative response (a miss, MI), not indicating the pattern sample From the above mentioned responses of the dog, two parameters of detection accuracy were calculated: Sensitivity  =  CP /(CP  +  MI) (1) Specificity  =  CN /(CN  +  FA) (2)

Additionally, to check the detection reliability on each testing day, a ‘zero’ trail was conducted in which no cancer-pattern sample was placed in the lineup and the dog had to refrain from any indication. Every second or third test day, two–three double-blind trials were performed, in which neither the dog handler nor the experimenter knew the status of a test sample (i.e. whether it was cancer or healthy control) and the assessment of the sample status was made exclusively on the base of dogs’ indications. In the double-blind trials the dogs were not rewarded, whatever they indicated, since it was not known to the experimenter whether the dog’s response was correct or false. The correctness of the double-blind trials was verified after the test, based on information from a third person who knew the status of the test sample but was not present during testing. A high rate of false alerts towards a breath sample from a control person who is presumed to be healthy at the time of testing, may prospectively be used to 4

assess dogs’ ability to detect early, preclinical stages of cancer if such person proves to have cancer in near future. The double-blind trials, although appropriate from an experimental point of view, cannot be performed too often. If dogs are unrewarded several times in turn, they might change their strategy in an endeavor to get a reward. As a consequence, they may make more false alerts.

5.  Breath odour samples From a practical point of view breath odour samples are easier to collect and to handle than urine or tumor tissue samples. They require no special handling or storage conditions. It was demonstrated in some studies [21, 37] that dogs are able to distinguish breath samples of cancer patients from those of healthy volunteers after storage up to several weeks. Because of a low concentration of VOCs ranging from ppt(v) to ppb(v) level, and high humidity, the exhaled breath samples to be analyzed by GCMS require an enrichment step to achieve the LOD of detectors used in gas chromatography. The most common technique is sorption on solid sorbents following by thermal desorption. The effect of storage time on the recovery of the analytes from sorbent tubes was investigated [38, 39] and the data show that losses of particular compounds are related to the storage time of loaded tubes. In particular, significant losses of reactive compounds such as isoprene, aldehydes and volatile sulfur compounds (VSC) were observed. Therefore, an immediate desorption and analysis of the sorbent tubes was recommended [39, 40]. Due to variations in recruitment rates and the limited window of time between diagnosis and chemotherapy, sample storage time varied between 1 and 60 d [21]. In published studies there was no standardization of sample collection. Ehmann et al [26] used cylindrical glass tubes closed at both ends by removable caps and filled with polypropylene fleeces that were impregnated with a silicone oil to have either hydrophilic or hydrophobic absorbing properties. Subjects investigated by Sonoda et al [24] exhaled 100–200 ml into ­breath-sampling bags (Otsuka Pharmaceutical Company, Tokyo, Japan) that were then fitted with their end caps and sealed in Ziplock-style bags at 4 °C until presentation to the dog. In the study of Amundsen et al [29] donors exhaled three forced expirations through sterile exhalation filters (Breathing Filter Allegro ID 547468, Allegro Medical Inc. US) and the filters were transferred to cool storage in cleaned and sealed containers. For the olfactory test the filters were placed into these containers at room temperature one hour before test. McCulloch et al [21] and Walczak [27, 37] used cylindrical, polypropylene, organic vapor-sampling tubes (Defencetek, Pretoria, South Africa) open at either end for sampling and closed with removable end caps for storage. Within the tubes there were removable inserts of silicone, oil-coated polypropylene ‘wool’ that captured volatile organic

J. Breath Res. 9 (2015) 027001

T Jezierski et al

Table 3.  Detection accuracy for cancer and melanoma using breath and other odour samples.

Authors

Cancer type

Odour samples

Detection sensitivity %

Detection specificity %

Percent of correct indications by chance

McCulloch et al [21]

Lung

Breath

99

99

50

Breast

Breath

88

98

50

Colorectal

Breath

91

99

Not indicated

Colorectal

Feces

97

99

Not indicated

Ehmann et al [26]

Lung

Breath

71

93

Walczak et al [37]

Lung

Breath

83.4–83.6a

77.9–80.5a

50

53.6–54.0a

b

20

89.2–90.5a

83.9–88.8a

50

b

20

66.3–80.2a

77.2–87.4a

50

32.2–58.6a

b

20

Sonoda et al [24]

Breast

Breath

66.8–67.8 Melanoma Amundsen et al [29]

a b

Breath

a

Lung

Breath

56–64

8–33

Lung

Urine

64–74

25–29

Range for individual dogs. Not calculated due to small number of ‘zero’ trials.

compounds in exhaled breath (two–three exhalations) as breath passed through the tube.

6.  Detection sensitivity and specificity Detection sensitivity of odour markers of lung cancer by trained canines varied from 56 to 99% and the specificity ranged 32% (for an individual dog, at 20% correct by chance) to 90% (average for 50% correct by chance) (table 3). Some authors indicated the probability of correct responses by dogs by chance (table 3). The majority of authors, however, did not mention the probability of correct indications by chance. Over all papers published so far on canine detection of lung cancer in humans on the basis of breath odour, the mean and median sensitivity was 78 and 78.5% respectively, whereas the mean specificity was 71.5% and median specificity 89.5%. For comparison, the results of some studies regarding breath-sampled detection accuracy for lung cancer using another type of odour samples, and for other types of cancers, were given in table 3. Using three dogs trained to detect odour markers of lung and breast cancer as well as melanoma [37], the study found the dogs performed significantly better in indicating breath samples taken from patients with breast cancer than from patients with lung cancer. The study also found that melanoma samples were indicated at a significantly lower rate (tables 3 and 4). Since dogs are trained using pattern breath odour samples from donors with biopsy-confirmed cancer disease, and the ultimate goal of using dogs for cancer screening would be achieved if dogs could be found to discriminate early, pre-clinical stages of cancer could be proven, the stage of cancer can be regarded as an important factor. Some authors [21, 29] classified breath odour samples according to the cancer stage. For example McCulloch et al [21] classified their non 5

small-cell lung cancer (NSCLC) breath samples into stages one–four of adenocarcinoma and stages two– four of squamous type, and found no significant differences in detection sensitivity and specificity across all four stages of the disease. Amundsen [29], in an interim analysis of the first 46 patients, found canine 70 and 55.6% sensitivity for NSCLC and for small cell lung cancer (SCLC) respectively, whereas the specificity for both lung cancer types was 8.3%. After intensive training of the dogs, Amundsen et al [29] could ascertain that the sensitivity for detection of NSCLC decreased to 60% and specificity increased to 33.3%, while for SCLC the sensitivity increased to 100% and specificity increased to 33.3%. Classifying cancers and melanoma roughly into more and less advanced stages, Walczak [37] found that the odds ratio for detection of more-advanced cancer stages was not significantly higher than for less advanced stages. No influence of cancer stage on canine outcome was also confirmed in the study of Ehmann et al [26], as the accuracy of dogs’ indications did not favor advanced tumor stages in this study.

7.  Confounding factors A number of confounding factors may play a role in cancer detection by trained canines. For dogs that are trained to detect a specific chemical compound, in this case a hypothetical compound(s) emitted by cancer cells, all other ­compounds or changes in their quantitative proportions may be confusing because of problems with generalization of a characteristic odour signature of cancer disease. Influence of some confounding factors in breath samples were investigated by Walczak [37] (table 4). One of important confounding factors, which may occur frequently, is tobacco smoking. Whereas is some studies no significant influence was found of tobacco smoking on detection accuracy by

J. Breath Res. 9 (2015) 027001

T Jezierski et al

Table 4.  Impact of some factors on indication of breaths samples of patients with cancer disease and healthy volunteers [37 modified]. Odds ratio

Factor

Patients with cancer correct indications

P

Healthy volunteers false alerts

Patients with cancer correct indications

95% Confidence interval Healthy volunteers false alerts

Patients with cancer correct indications

Healthy volunteers false alerts

Cancer: Lung

1

Breast

1.54

0.016

1.08–2.20

Melanoma

0.32

0.000

0.24–0.43

Dog: COU

1

1

CYG

1.75

0.74

0.000

0.000

1.35–2.27

0.65–0.84

GRO

1.30

0.76

0.280

0.000

0.90–1.43

0.68–0.86

Non-smokers

1

1

Smokers

0.63

1.01

0.000

0.740

0.49–0.80

0.91–1.13

Women

1

1

Men

0.60

1.06

0.000

0.213

0.46–0.78

0.96–1.18

Non-dairy

1

1

Dairy

0.58

1.25

0.000

0.001

0.45–0.75

1.09–1.44

Non-meat

1

1

Meat

0.64

1.26

0.000

0.000

0.51–0.83

1.10–1.43

Non-spicy

1

1

Spicy

0.58

1.42

0.345

0.000

0.19–1.78

1.24–1.63

After meal

1

1

Before meal

1.11

1.51

0.596

0.000

0.74–1.67

1.31–1.73

1.000

1.001

0.344

0.000

0.99–1.00

0.47–1.98

Outdoor

0.92

1

0.041

0.000

0.85–0.99

3.71–7.27

In hospital

1

5.19

Smoking:

Sex:

Last meal:

Taking samples:

Sample storage: Each day Sampling site:

canines [21], in another studies [37] smoking negatively influenced detection accuracy in cancer patients since the odds ratio for detection of breath samples from smokers was significantly lower (table 4). On the other hand no significant influence of smoking was found on the rate of false indications towards healthy, control samples (table 4). Pattern breath samples used for the dog training are mostly collected from cancer-diagnosed patients in oncology hospitals. This is because the patients who are donors of the pattern samples should have had cancer confirmed by histopathological examination, but not yet have undergone chemotherapy, which would substantially change the breath odour. Because such donors of pattern samples are mostly available in hospitals, a characteristic ‘hospital odour’, derived from substances used for disinfection, should be taken as a potential confounder. The study of Walczak [37] revealed that collecting breath samples from cancer patients outside hospital rooms significantly decreased the odds ratio for indication by dogs (P = 0.041, table 4) and significantly increased the odds 6

ratio for false indications of healthy donors (P = 0.000, table 4). In follow-up study [27] the ‘hospital odour’ as a confounding factor was confirmed. However, in the latter study, breath samples of lung-cancer patients taken from the same donors in hospital and outside hospital (outdoors) were indicated by dogs at the same rate, taking into account criterion I, where the probability of correct indications by chance was 50%. Taking into account criterion II, at 20% correct indications by chance, the detection rate of breath pattern samples was significantly higher when the samples were taken outdoors (table 5). Although the number of lung-cancer patients that were sampled outdoors and the number of healthy volunteers’ samples in hospital rooms were low, the trials were repeated up to 10–12 times, so the percentage of correct indications and false alerts could be calculated. The dogs made significantly more false alerts towards breath samples taken from healthy volunteers if the samples were collected in hospital, compared with taking samples ­outdoor (table 5), but this difference was significant

J. Breath Res. 9 (2015) 027001

T Jezierski et al

Table 5.  Impact of taking breath samples inside versus outside hospitals on correct indications by dogs [27 modified]. Percent of indications

Character of indications

50% correct by chance

20% correct by chance

Samples taken inside hospitals (1) and (2) (n = 60)

Correct

79.1 a

47.2 c

Samples taken inside hospital (3) (n = 32)

Correct

80.4 a

50.1 c

Correct

78.4 a

54.4

Samples taken inside hospitals (1) and (2), (n = 6)

False alerts

34.0b

4.0d

Samples taken inside hospital (3) (n = 4)

False alerts

27.9b

11.0d

False alerts

20.3

2.0d

Samples taken inside hospitals (1) and (2) (n = 4)

False alerts

34.5b

2.9d

Samples taken inside hospital (3) (n = 3)

False alerts

30.6b

3.0d

Odour samples Breath from patients with lung cancer:

Breath from patients with lung cancer: Samples taken outside hospitals (n = 4) Breath from healthy control persons:

Breath from healthy control persons: Samples taken outside hospitals (n = 301) Ambient air from hospital room

Differences between sample groups significant at least at Chi2 > 6.64, d.f. = 1, P 80% of the variance were essential for classification in the patient group. Positive signs of loading coefficients were given in Factor 1 for ethanol, 2-methylpropane, butane, isoprene, pentane and benzene, whereas negative signs were found for carbon disulfide, 2-butanone and toluene (table 6). Factor 2,

J. Breath Res. 9 (2015) 027001

T Jezierski et al

Table 6.  Correlation coefficients between selected compounds and the dogs’ indications [43 modified].

Volatile compound

Pearson’s correlation coefficient with dogs’ indications

% Indications

Correlations between breath VOCs and dog’s indications as resulted from factor analysis (FA) Factor 1

Factor 2

0.22

0.97

2-Pentanone

0.97

0.45

0.89

Ethyl acetate

0.85

0.35

0.86

Hexane

0.50

0.37

0.55

Pentane

0.47

0.76

0.39

Toluene

0.30

 − 0.86

0.49

2-Methylpropane

0.29

0.74

0.04

Ethanol

0.27

0.97

0.09

Butane

0.26

0.92

0.00

2-Methylpentane

0.14

0.59

0.12

Benzene

0.11

0.86

 − 0.15

Isoprene

0.07

0.96

 − 0.11

2-Butanone

 − 0.51

 − 0.95

 − 0.30

2-Propanol

 − 0.60

0.58

 − 0.78

Carbon disulfide

 − 0.67

 − 0.73

 − 0.45

Acetone

 − 0.77

0.32

 − 0.79

Acetonitrile

 − 0.78

0.16

 − 0.91

Propanal

 − 0.87

0.15

 − 0.86

1-Propanol

 − 0.98

 − 0.28

 − 0.91

being associated with the dogs’ indications, showed positive loading for compounds that were not classified by Factor 2. In principal component analysis (PCA) the sign of loading explained the role of the volatile compounds in dog’s indications. In studies of Buszewski et al [42, 44] the first factor could be related to indifference by the dog and the second factor was related with dogs’ indications. Since for the healthy controls the two principal factors explained only 51% of the variance, it could not be definitely assessed which compound or combination of compounds in the breath sample were critical in whether a sample was indicated or not indicated by dogs. Another problem in relating results of GC-MS analysis to canine indications may be caused by the fact that the results of GC-MS do not explain how a mixture of volatile compounds is perceived by canine olfaction [42]. Qualitative and quantitative olfactory impression of a mixture of compounds as perceived by olfactory organ of animals and humans cannot be assessed by current analytical methods.

9.  Potential advantages and perspectives of using canines for cancer screening Using trained dogs as bio-sensors for cancer screening would have some clear advantages. First, collecting odour samples is entirely noninvasive, safe and easy—both to the patient and to everyone—and requires no special qualifications or skills. Sampling tubes are not expensive and could be purchased in bulk from several manufacturers. Storage of samples requires no special conditions and odour 8

samples for dogs could be stored up to several weeks/ months after sampling. This is an important advantage, because it is not always possible to test samples shortly after collection. When properly handled, odour samples can be tested several times, by several dogs. A variety of dog breeds used in published studies allow us to presume that no special requirements as to the dogs have be fulfilled. However, probably only a segment of canine populations is able to successfully complete the training and work efficiently for a longer period. So far no data have been published regarding how long the utilization period of trained cancerdetecting dogs could last. Training of dogs and testing procedure is relatively easy for skilled staff taking some rules into consideration. Dogs’ indications in the odour lineup are of binary character (yes/no) and are therefore easy to interpret. Hesitations of dogs during work in the lineup were practically never observed. Although some authors [22] could not confirm that trained dogs are able to distinguish patients with cancer from healthy controls on the base of odour samples at a rate better than pure chance, the majority of other authors found relatively high detection sensitivity and specificity in relation to pattern odour samples [21, 23, 24, 26–28]. Despite differences in methods and experimental set-up, the majority of results show that dogs are able to distinguish characteristic cancer odour at a rate better than by chance. Due to the relative ease, simplicity and low cost of using canines for cancer screening, this method would have good prospects, especially in low-income

J. Breath Res. 9 (2015) 027001

T Jezierski et al

c­ ountries where common access to high-tech screening methods is still problematic. Recently a new research approach has been proposed [34], involving combining highly sophisticated e-sensing technologies in the analysis of breath samples of carefully selected and defined patient populations that were positively or negatively indicated by sniffer dogs. The authors of this concept [34] are of the opinion that using this approach the goal to identify lung-cancer specific volatile compounds may be achieved. Of all papers regarding cancer detection on the base of emitted breath or urine odours using sniffer dogs, the vast majority of authors concluded that this method is a promising, practical screening methods, if rigorous double-blind tests are used to avoid confounding effects. The authors concluded that deeper understanding and refinement would be achieved through further studies [21–27, 29, 30, 34, 42].

10. Disadvantages and limitations of using canines as ‘biosensors’ for detection of volatile compounds It should be mentioned that after a decade of research, trained canines have not been acknowledged by oncologists as reliable cancer screeners. Proof that dogs were able to distinguish patients with fully diagnosed cancer from healthy controls, would be only the first step—scientifically interesting, but not very useful for oncologists who need screening methods that can identify persons with very early, preclinical stages of neoplastic diseases. There has been no confirmed proof that dogs are actually able to distinguish such early, preclinical cancer stages from healthy controls, or from other diseases. People take for granted that the canine sense of smell is extraordinarily efficient, and that operantconditioned response training—such as was used to train dogs in the trials studied – never fails. Both these presumptions may be untrue. The olfactory threshold may be different for different compounds, and there are significant individual differences between dogs, and between different estimation methods [41]. The true acuity of the canine sense of smell is difficult to assess experimentally, since it involves dogs’ ability to perform well in operant conditioning. Other methods of estimation of olfactory thresholds (e.g. electroencephalographic olfactometry), are impracticable since they require immobilization of animals [44]. Lack of a learned reaction to an odour does not mean that a dog does not distinguish it. The dog may successfully detect a particular odour but not perform a trained behavioral reaction because the odour is not interesting, or the learned operant-conditioned reaction fails for any reasons. In the work of sniffer dogs two phenomena have to be taken into consideration: memorization of particular odours and generalization of odour signatures. Dogs have an excellent ability to memorize large ­numbers of 9

odours that are associated with pleasant (e.g. a reward) or unpleasant events. Odour memorization may have both favorable and unfavorable consequences for the outcome of scent-lineup tests. The memorization of a trained characteristic cancer-odour signature (if it exists), especially in so called long-term memory, coupled with ability to generalize a common odour signature, enables the dog to recognize this common odour signature in consecutively tested samples. On the other hand an ability to memorize large numbers of individual odour samples used in the training process, coupled with poor generalization, may cause a dog to ignore novel samples of patients with cancer, especially in double-blind trials [30]. For dogs the lineup procedure is a sort of game aimed at earning a reward in the simplest possible way and at the lowest workload. This may cause a variation in detection sensitivity and specificity depending on dogs’ motivation, operant conditioning, olfactory memory, odour contamination and other factors that may be difficult to control. Therefore a systematic check of the dogs’ performance in double blind trials is necessary. Despite some promising results concerning chemical compounds that may play a role in dogs’ indications [9, 42], the use of canines for cancer screening remains a ‘black-box technology’ since it is still not known exactly which chemical compound(s) or which qualitative or quantitative combination of compounds make the cancer-odour signature the sniffer dogs react to.

11.  Recommendations for methodological standardization The variability of results of experimental studies on cancer detection by trained canines published to date could be partly attributed to a lack of methodological standardization. For a methodology to be generally accepted and applied as a standard, it must be proven that this standard would produce the most reliable and valid results and would be applicable under a variety of conditions and circumstances. A relative scarcity of published experiments conducted with scientific scrutiny does not allow us to assess which standard would be the best. The authors evidently preferred to use own methodology, based on similar principles but differing in details. Some summary recommendations as to future research on cancer detector dogs, based on their own results, have been made by Elliker et al [30]. These recommendations cover several aspects of the research and therefore should be further systematized, discussed and expanded. Procedures of standardization concern at least three aspects: (1) quality of dogs and training methods, (2) taking and handling breath samples, and (3) experimental setup and assessment of dogs’ indications. No detailed comparison of dog breeds trained for cancer detection has been done but it could be assumed that individual characteristics of a dog such

J. Breath Res. 9 (2015) 027001

T Jezierski et al

as ­trainability, motivation and acuity of the sense of smell are more important than dog’s breed. There is no clear answer to the question which breed would be most suitable for the cancer detection, although the potential to find a suitable candidate dog for cancer detection training among typical working breeds—e.g. German shepherd or Labrador retriever—seems greater than among typical pet breeds. The greater potential of the two mentioned breeds as sniffer dogs has been also suggested based on genetic diversity of olfactory receptors [33, 45]. Some authors suggested that it may be useful to breed dogs specifically for cancer odour to increase the proportion of dogs suitable for this task [30] but taking into account necessary performance tests and selection procedure in a sufficiently large dog population, this idea seems to be less realistic. No standards exist as to the type of disposable breath sampling tubes, and no special tubes for this purpose have been manufactured so far. Researchers adopted different types of tubes originally used for other purposes— e.g. for land mines detection or a heat moisture exchanger used in respiratory and anesthesia procedures. Sampling tubes should be simple and handy to be used by breath donors without special training. They should absorb well different types of VOCs without reacting with them or changing their odour properties. The construction and material the sampling tubes are made of should enable presentation of odour samples to dogs without contamination by ambient odours—i.e. should be impermeable for odours from outside and inside and prevent direct contact with the dog’s nose, saliva, etc. Lacking sufficient comparative research results, no unequivocal recommendations can be made as to collecting samples and training to indicate one type of cancer versus several types of cancer. It is still not known whether different types of cancer have a common odour signature, so that dogs could be encouraged to generalize on that common signature. If a common cancer odour signature does exist, training involving several different types of cancer would help amass sufficient pattern samples for training [30] as well as making dogs more versatile detectors. If, however, such a common odour signature does not exist, multi-cancer training may encourage the dogs to generalize on the odour of particular type of cancer and thus would increase the potential impact of confounding odours. Generally, for training and testing, a lineup or circle of four–seven odour samples can be recommended. The more samples in a lineup or circle, the lower probability of correct indications by pure chance. However, too many samples in the lineup cannot be recommended because dogs can miss, or not sniff, some samples and thus the response to those samples would be unknown. A risk of memorization of particular odour samples used for training, and poor generalization, should be taken into consideration. This is why Elliker et al [30] recommendation that repeated presentation of samples from the same donors should be minimized as far as possible is to be fully supported. 10

In some studies, training to indicate cancer odour in isolation before introducing control samples that have to be disregarded by dogs, was successful [20, 27], whereas other authors recommended training dogs to disregard control samples from the outset of the training, and never at any stage to present cancer samples without control samples [30]. However, there are no experimental data to support the latter suggestion. All tests, except for the very first training stages, should be conducted using double-blind protocol—i.e. the dog handler and the experimenter should be blind to the location of the target sample (cancer pattern odour) in the lineup to prevent unconscious signaling to the dog which sample the dog is expected to indicate. However the lineup procedure is reward-based operant conditioning, involving reinforcement of correct indications by the dog to avoid extinction of the learned reaction and confusing the dog, therefore dogs should be rewarded only for a correct indication of the target sample. To produce and sustain the learned reaction, a reinforcement should immediately follow the correct reaction. Forms of reinforcement schedules suggested by Elliker et al [30] for cancer-detection dogs include delayed reinforcement, or variable reinforcement schedule (rewarding only for a proportion of correct indications). These reinforcement schedules should first be validated experimentally, to demonstrate that they would work in all dogs regardless of differentiated temperament and learning ability. Alternative training and testing paradigms—for example habituation–dishabituation, suggested to be superior [30] to search-based discrimination tasks were applied so far— should not only be proven experimentally but a testing protocol should be proposed: how to use dogs trained by habituation–dishabituation for practical screening of healthy donors. Some authors [34] concluded that several studies had confirmed the existence of a stable marker or breath odour pattern associated with lung cancer that was independent from other diseases, especially COPD, and not influenced by tobacco smoke, food odours and drug metabolities. However, the dogs should be systematically checked and training should be sustained to ensure the dogs continue to disregard any odours that are not related with cancer. To test this, lineups should include samples from COPD patients and from patients with other diseases. Occasionally a ‘zero’ trial with no target odour sample should be conducted, to teach the dogs to refrain from a trained reaction if a target odour is not available. In conclusion, while some open questions remain that are worth investigating, it is acknowledged that trained dogs can discriminate breath samples of patients diagnosed with cancer from those of ‘healthy’ people, at a ‘better than by chance’ rate. However, it is still not confirmed that dogs can sniff out cancer odour markers in early stages of the disease, before medical diagnosis. By using an odour lineup containing a cancer odour pattern sample, one–two decoys (COPD and/or other

J. Breath Res. 9 (2015) 027001

T Jezierski et al

diseases) and two–three control odours (breath samples of healthy people), we propose a testing protocol for practical cancer screening. Under our proposal, dogs would be systematically checked for their accuracy of detection of cancer pattern odour and subjected to sustaining training that prevents extinction of the learned behavior. Also, more importantly, a larger number of healthy people should be systematically screened for cancer. If a healthy donor has been indicated by dogs statistically more often than other controls, there would be a suspicion of a undiagnosed (‘early’) cancer, and the person would be advised to undergo medical screening. This seems to be a way to demonstrate that trained dogs would actually be useful as cancer screeners.

Acknowledgments This work was supported by project SensorMed (NCBiR) PBS/A3/7/2012 (2012–2015) and the Polish Ministry of Science and Higher Education # 2POZ6204127 (2004–2007).

References [1] World Health Organization 2013 The top 10 causes of death. WHO, Media Centre, Fact Sheet 310 (updated June 2011) Available at www.who.int/mediacentre/factsheets/fs310/en/ index.html (accessed 26th January 2013) [2] Henschke C I et al 1999 Early lung cancer action project: overall design and findings from baseline screening Lancet 354 99–105 [3] Kaneko M et al 1996 Peripheral lung cancer: screening and detection with low-dose spiral CT versus radiography Radiology201 798–802 [4] Sone S et al 1998 Mass screening for lung cancer with mobile spiral computed tomography scanner Lancet351 1242–5 [5] Henschke C I, Yankelevitz D F, Libby D M, Pasmantier M W, Smith J P and Miettinen O S 2006 Survival of patients with stage 1 lung cancer detected on CT screening New Engl. J. Med. 355 1763–71 [6] Perneger T V, Cullati S, Schiesari L and Charvet-Berard A 2010 Impact of information about risks and benefits of cancer screening on intended participation Eur. J. Cancer 46 2267–74 [7] Smith-Bindman R 2010 Is computed tomography safe? New Engl. J. Med. 363 1–4 [8] Phillips M, Cataneo R N, Ditkoff B A, Fisher P, Greenberg J, Gunawardena R, Kwon C S, Rahbari-Oskoui F and Wong C 2003 Volatile markers of breath cancer in the breath Breast J. 9 184–91 [9] Amann A, Miekisch W, Schubert J, Buszewski B, Ligor T, Jezierski T, Pleil J and Risby T 2014 Analysis of exhaled breath for disease detection Annu. Rev. Anal. Chem. 7 455–82 [10] Haick H, Broza Y, Mochalski P, Ruzsanyi V and Amann A 2014 Assessment, origin, and implementation of breath volatile cancer markers Chem. Soc. Rev. 43 1423–49 [11] Phillips M, Gleeson K, Hughes J M B, Greenberg J, Cataneo R N, Baker L and McVay W P 1999 Volatile organic compounds in breath as markers of lung cancer: a cross-sectional study Lancet 353 1930–3 [12] Phillips M et al 2007 Prediction of lung cancer using volatile biomarkers in breath Cancer Biomark. 3 95–109 [13] Ligor M et al 2009 Determination of volatile organic compounds in exhaled breath of patients with lung cancer using solid phase microextraction and gas chromatography mass spectrometry Clin. Chem. Lab. Med. 47 550–60

11

[14] Peng G et al 2010 Detection of lung, breats, colorectal and prostate cancers from exhaled breath using a single array of nanosensors Br. J. Cancer103 542–51 [15] Horvath I, Lazar Z, Gyulai N, Kollai M and Losonczy G 2009 Exhaled biomarkers in lung cancer Eur. Respir. J.34 261–75 [16] D’Amico A et al 2010 An investigation on electronic nose diagnosis of lung cancer Lung Cancer68 170–76 [17] Williams H and Pembroke A 1989 Sniffer dogs in the melanoma clinic? Lancet 1 734 [18] Church J and Williams H 2001 Another sniffer dog for the clinic? Lancet 358 930 [19] Pickel D, Manucy G P, Walker D B, Hall S B and Walker J C 2004 Evidence for canine olfactory detection of melanoma Appl. Anim. Behav. Sci. 89 107–16 [20] Willis C M, Church S M, Guest C M, Cook W A, McCarthy N, Bransbury A J, Church M R T and Church J C T 2004 Olfactory detection of human bladder cancer by dogs: proof of principle study Br. Med. J. 329 712–5 [21] McCulloch M, Jezierski T, Broffman M, Hubbard A, Turner K and Janecki T 2006 Diagnostic accuracy of canine scent detection in early- and late-stage lung and breast cancers Integr. Cancer Ther. 5 30–9 [22] Gordon R T et al 2008 The use of canines in the detection of human cancers J. Altern. Complement. Med. 14 61–7 [23] Horvath G, Järverud G K, Järverud S and Horvath I 2008 Human ovarian carcinomas detected by specific odour Integr. Cancer Ther. 7 76–80 [24] Sonoda H et al 2011 Colorectal cancer screening with odour material by canine scent detection Gut60 814–19 [25] Cornu J N, Cancel-Tassin G, Ondet V and Girarget C 2011 Olfactory Detection of prostate cancer by dogs sniffing urine: a step forward in early detection Eur.Urol. 59 197–201 [26] Ehmann R, Boedeker E, Friedrich U, Sagert J, Dippon J, Friedel G and Walles T 2012 Canine scent detection in the diagnosis of lung cancer: revisiting a puzzling phenomenon Eur. Respir. J. 39 669–76 [27] Walczak M, Jezierski T, Górecka-Bruzda A, Sobczyńska M and Ensminger J 2012 Impact of individual training parameters and manner of taking breath samples on the reliability of canines as cancer screeners J. Vet. Behav. Clin. Appl. Res. 7 283–94 [28] Horvath G, Andersson H and Nemes S 2013 Cancer odour in the blood of ovarian cancer patients: a retrospective study of detection by dogs during treatment, 3 and 6 months afterward BMC Cancer13 396–402 [29] Amundsen T, Sundstrǿm S, Buvik T, Gederaas O A and Haarverstad R 2014 Can dogs smell lung cancer? First study using exhaled breath and urine screening in unselected patients with suspected lung cancer Acta Oncol. 53 307–15 [30] Elliker K, Sommerville B A, Broom D M, Neal D E and Armstrong S 2014 Key considerations for the experimental training and evaluation of cancer odour detection dogs: lesson learnt from a double-blind, controlled trial of prostate cancer detection BMC Urol. 14 22–30 [31] Moser E and McCulloch M 2010 Canine scent detection of human cancers: a review J. Vet. Behav. Clin. Appl. Res.5 145–52 [32] Bhadra A 2011 Woof! Smells like cancer Curr. Sci. 101 480–3 [33] Lippi G and Cervellin G 2012 Canine olfactory detection of cancer versus laboratory testing: myth or opportunity? Clin. Chem. Lab. Med. 50 435–9 [34] Boedeker E, Friedel G and Walles T 2012 Sniffer dogs as part of a bimodal research approach to develop a lung cancer screening Interact. Cardiovasc. Thorac. Surg. 14 511–5 [35] McCulloch M, Turner K and Broffman M 2012 Lung cancer detection by canine scent: will there be a lab in lab? Eur. Respir. J. 39 511–12 [36] Balseiro S C and Correia H R 2006 Is olfactory detection of human cancer by dogs based on major histocompatibility complex-dependent odour components?—A possible cure and precocious diagnosis of cancer Med. Hypotheses. 66 270–2 [37] Walczak M 2009 Operant conditioning of dogs for detection of odour markers of cancer diseases (in Polish) PhD thesis.

J. Breath Res. 9 (2015) 027001

T Jezierski et al

Institute of Genetics and Animal Breeding of Polish Academy of Sciences [38] Dettmer K, Knobloch T and Engewald W 2000 Stability of reactive low boiling hydrocarbons on carbon based adsorbents typically used for adsorptive enrichment and thermal desorption Fresenius J. Anal. Chem. 366 70–8 [39] Dettmer K and Engewald W 2002 Adsorbent materials commonly used in air analysis for adsorptive enrichment and thermal desorption of volatile organic compounds Anal. Bioanal. Chem.373 490–500 [40] Palluau F, Mirabel P and Millet M 2007 Influence of relative humidity and ozone on the sampling of volatile organic compounds on carbotrap/carbosieve adsorbents Environ. Monit. Assess.127 177–87 [41] Walker D B, Walker J C, Cavnar P J, Taylor J L, Pickel D H, Hall S B and Suarez J C 2006 Naturalistic quantification of canine olfactory sensitivity Appl. Anim. Behav. Sci. 97 241–54

12

[42] Buszewski B, Ligor T, Jezierski T, Wenda-Piesik A, Walczak M and Rudnicka J 2012 Identification of volatile lung cancer markers by gas chromatography—mass spectrometry: comparison with discrimination by canines Anal. Bioanal. Chem. 404 141–6 [43] Buszewski B, Rudnicka J, Ligor T, Jezierski T, Walczak M and Wenda-Piesik A 2013 Analysis of cancer biomarkers in exhaled breath and comparison with sensory indications by dogs Volatile Biomarkers: Non-invasive Diagnosis in Physiology and Medicine ed A Amann and D Smith (Amsterdam: Elsevier) pp 177–92 [44] Hirano Y, Oosawa T and Tonosaki K 2000 Electroencephalographic olfaktometry (EEGO) analysis of odour responses in dogs Res. Vet. Sci. 69 263–5 [45] Robin S, Tacher S, Rimbault M, Vaysse A, Dréano S, André C, Hitte C and Galibert F 2009 Genetic diversity of canine olfactory receptors BMC Genom. 10 21