Ten dogs of seven different breeds (1–11 years old, four females and six males) were initially recruited from a pool of dogs attending classes at a dog training centre. This selection was based on the professional opinion of dog trainers or behavioural scientists, who had trained or observed the dogs during activities such as agility, obedience, gun-dog work, or location and retrieval of items for their owners. The dogs had not previously been used in scientific odour discrimination work. Further down-selection based on the dogs’ abilities to detect odours was carried out in the training stages (see Training procedure). All dogs were handled by professional dog trainers or behavioural scientists during the training and testing sessions and were cared for by their owners between sessions.
Urine sample collection and preparation
Urine samples were collected and prepared, with the donors’ permission, in the Department of Oncology, Addenbrooke’s Hospital, UK. All samples were collected using the same protocols, at the same locations and by the same research team to ensure they had the same general background odour. The age, Prostate Specific Antigen (PSA) measurement, urinalysis, Gleason Test score (when obtained) and medical history were recorded.
In total, 50 prostate cancer (CaP) samples and 67 control samples from different individuals were collected and used over the course of the dog training period. CaP samples were collected from donors with prostate cancer that had been previously confirmed by biopsy but remained untreated. The degree of disease in CaP donors varied from small, relatively innocent tumours to metastasised cancer. Control samples were collected from men with Benign Prostatic Hyperplasia (BPH – a benign enlargement of the prostate), as well as 10 healthy men without clinical symptoms. Fifty-two controls had PSA levels <0.5 ng/ml, two had PSA <1.5 ng/ml and seven had PSA between 2.2 and 11.6 ng/ml. Thirteen of these controls, including all with PSA >2.2 ng/ml, had previously undergone prostate biopsy with negative results. Donors were excluded from the CaP group if they presented with frank haematuria or urinary tract infection, from the control group in the case of uncertain diagnosis and from both groups if they suffered from a current, non-prostatic cancer. CaP and control sample pools were chosen that were as similarly aged as was feasible based on the available urine donors.
For double-blind testing, 15 CaP and 45 control samples were used in Test 1 and 16 CaP and 48 controls in Tests 2 and 3 (Table 1). All control donors used in tests had a PSA <0.5 ng/ml. Overall mean donor ages in Test 1 were 64.1 years (standard deviation = 8.3) for CaP and 58.3 years (standard deviation = 6.6) for control. For Tests 2 and 3 the overall mean ages were 63.6 years (standard deviation = 6.4) for CaP and 57.7 years (standard deviation = 5.2) for control. The range of donor age differences between the CaP and control samples presented to the dogs in each array was: 34.5% of control samples within ± 5 years of the CaP, 28.6% within ± 10 years, 32.1% within ± 20 years and 4.8% ± 20 years or more. All samples were taken from different individuals, had not previously been presented to the dog and were presented only once during each test.
Urine samples were collected in 50 ml polypropylene screw-cap tubes and frozen at -20°C within 10 minutes. Samples were transported to the testing centre on dry ice, defrosted in a 37°C water bath, aliquoted into 1.5 ml polypropylene micro-centrifuge tubes and stored in a freezer at -20°C. Samples were generally stored for 1 to 60 days prior to presentation to the dogs, though some were stored for up to 6 months. During training or testing, 1 ml aliquots were heated to 37°C in a water bath and then presented to the dogs in new open-top, polypropylene test tubes.
Dogs were trained in a 6 m × 10 m rubber-floored arena (Figure 1). Urine samples were presented in four, 90 mm-deep aluminium flasks recessed into a 3 m long floor-mounted plastic array. The urine was not visible or accessible to the dogs other than by olfaction through four, 20 mm-diameter ‘scent holes’, spaced 0.75 m apart, positioned directly above each flask. These holes were labelled above with numbers. To prevent cross-contamination, the investigator wore nitrile gloves when handling sample tubes and inserted them into the array using stainless steel forceps.
Dog training procedure
Dogs were trained using a positive reinforcement ‘clicker’ technique and food rewards/praise. Initially, food rewards were randomly hidden in one of the scent holes and the trainer rewarded the dog for sitting or lying next to and placing its nose on the hole. Once interested in the holes, Stage 1 commenced in which dogs had to indicate on single CaP urine samples placed in a random hole, with empty test tubes in the remaining holes. The criterion for a dog to move to the next stage was 9/10 successive runs correct. In Stage 2, the arrays contained one CaP sample, with the remaining holes containing different control samples. The trainer was blind to the sample positions and had to call out the number of the hole suspected to contain the CaP based on the dog’s choice. The investigator, who was visually isolated, then informed the trainer whether the choice was correct, allowing the trainer to reward the dog appropriately. When the dog was able to identify the scent hole containing the CaP more frequently than expected by chance it was moved to the formal test stage.
The number of urine samples presented to each dog during Stages 1 and 2 varied depending on their individual rate of progress. Urine samples from different donors were used to try to encourage the dogs to generalise on a common cancer odour. CaP and control samples from new donors became available in batches of 5 to 10 at intervals over the training period, and it was sometimes necessary to present urine from the same donors several times during training. Urine samples from two or three different donors were sometimes pooled in different combinations to try to vary the odour profiles. Although four scent holes were used during the testing stages, various numbers (between three and six) were trialled during initial training.
Following training, three rigorous double-blind tests were conducted for two dogs that showed ability to discriminate CaP and control samples during training Stage 2. Test 1 involved dog A, a nine year-old yellow Labrador, who had undergone approximately 5 months of training in Stages 1 and 2 prior to the test. During the test, the dog was presented with 15 arrays, each containing one CaP sample and three controls. Each of the 15 CaP and 45 control samples were from unique donors and were new to the dog. For each of the 15 arrays, the position of the CaP sample in the array (1, 2, 3 or 4) was secretly allocated using random number lists generated remotely. This sample allocation code was sent to the urine collection centre, where it was used to prepare the 15 arrays of samples. The sample allocation code was concealed from the investigators and dog handlers, who did not know the position of the CaP and controls at any point before or during the test.During Test 1, the dog handler and dog were visually isolated in room A (Figure 1), while the investigator placed samples in the array. The investigator was then visually isolated in room B and called “ready”, signalling that the handler could enter the arena and allow the dog to sniff the array. Once the dog had chosen a hole, the handler rewarded the dog and called out the position in which he/she believed the CaP sample resided (1, 2, 3 or 4). Following a run the handler and dog moved back into room A while the next array was prepared. An independent referee who did not have a vested interest in the project was present to verify the double-blinding procedure.
Tests 2 and 3 followed the same protocol except that 16 arrays of new samples were used. Test 2 again involved dog A, who had undergone approximately 8 months of Stage 2 retraining following Test 1. For Test 2, an additional independent referee was provided with the sample allocation code prior to the test. Using a live speaker-phone system this referee, who was not present at the testing centre, was able to directly respond to the position numbers called out by the dog handler and immediately inform them whether the choice was correct. Unlike Test 1, this system allowed the dog handler to reward the dog only for correct choices of the CaP samples whilst still ensuring effective double-blinding.
Test 3 involved dog B, a three-year old Border Collie with a different handler from Tests 1 and 2. This dog had undergone approximately 5 months of training in Stages 1 and 2 prior to the test. Test 3 followed the protocol for Test 2, except that it took place in a different testing venue and the visually-isolated investigator informed the handler whether the dog’s choices were correct or incorrect. Dog B was tested using 16 sets of 4 urine samples from the same urine donors who had provided the samples for Test 2, presented in a different, randomised order.
For Tests 1 to 3, the primary outcome measure was each dog’s overall success rate in identifying cancer urine samples, compared with the success rate expected by chance, i.e. 25% (four scent holes). Each test had a power of greater than 80% to detect the difference between a dog’s success rate of 70% compared with the success rate expected by chance, assuming a two-sided significance level of 1%. The sensitivity and specificity were calculated for each test and the corresponding 95% confidence intervals were calculated using a robust variance estimate to allow for between-array variance. Agreement between dogs A and B in Tests 2 and 3, in which samples from the same urine donors were presented, was assessed using Cohen’s Kappa.
The University of Cambridge, Department of Oncology local ethics committee approved the study.