Clara Wilson, University of Pennsylvania, United States
This is an open-access article distributed under the terms of the
Post-traumatic stress disorder (PTSD) is an impairing mental health condition with high prevalence among military and general populations alike. PTSD service dogs are a complementary and alternative intervention needing scientific validation. We investigated whether dogs can detect putative stress-related volatile organic compounds (VOCs) in the breath of people with trauma histories (54% with PTSD) exposed to personalized trauma cues.
Breath samples were collected from 26 humans over 40 experimental sessions during a calm (control breath sample) and stressed state induced by trauma cue exposure (target breath sample). Two scent detection canines were presented with the samples in a two alternative forced choice (2AFC) discrimination and yes/no detection task. The 2AFC task assessed the dogs' ability to discriminate between the two states within the breath samples of one individual. The detection task determined their ability to generalize the target odour across different individuals and different stressful events of one individual. Signal Detection Theory was applied to assess dogs' sensitivity, specificity, precision, and response bias.
The dogs performed at ∼90% accuracy across all sample sets in the discrimination experiment, and at 74% and 81% accuracy, respectively, in the detection experiment. Further analysis of dog olfactory performance in relation to human donor self-reported emotional responses to trauma cue exposure suggested the dogs may have been detecting distinct endocrine stress markers. One dog's performance correlated with the human donors' self-reported fear responses and the other dog's performance correlated with the human donors' self-reported shame responses. Based on these correlations between dog performance and donor self-report measures, we speculate that the VOCs each dog was detecting likely originated from the sympathetico-adreno-medullary axis (SAM; adrenaline, noradrenaline) in the case of the first dog and the hypothalamo-pituitary-adrenal axis (HPA; glucocorticoids) in the case of the second dog.
Our proof-of-concept study is the first to demonstrate that some dogs can detect putative VOCs emitted by people with trauma histories when experiencing distress theoretically associated with the intrusion and arousal/reactivity symptoms of PTSD. Results have potential to improve the effectiveness and training protocol of PTSD service dogs with a focus on enhancing their alert function.
香京julia种子在线播放
Post-traumatic stress disorder (PTSD) is a trauma- and stress-related disorder in the
PTSD is prevalent [7.8% lifetime prevalence in the U.S. general population (
One complementary/alternative treatment for PTSD involves psychiatric service dogs–assistance dogs permanently placed with a patient and trained to help them (
Interrupting or alerting to episodes of anxiety/distress (e.g., flashbacks, nightmares) is reported as within the top three most appreciated and frequently-used trained tasks by veterans with a PTSD service dog (
Canines' sense of smell is 10,000–100,000 times more sensitive than humans' (
Three categories of individual human odours are: the stable “primary odour” based on the individual's genetics, age, and sex; the changing “secondary odour” based on endogenous dietary and environmental factors, including pathological status; and the “tertiary odour” based on exogenous factors like personal hygiene products (
Certain medical conditions alter the VOCs cells release into the respiratory system (
There is preliminary evidence of dogs’ ability to detect VOCs associated with elevated human stress levels. One study (
Another study (
The above-described studies suggest human stress responses involve VOCs that at least some dogs can detect, although the specific stress biomarker(s) dogs rely on remain unclear. Technological advancements have also successfully detected human stress response biomarkers in breath (
Another study (
Dogs might have an advantage as they likely possess a “sensor” capable of a more comprehensive and less specialised perception of the array of stress volatiles. There are several reasons why dogs could be sensitive to human stress VOCs. First, dogs might have developed olfactory recognition of human emotions during domestication (
Another possible mechanism is emotional contagion, which occurs when “another's emotional state triggers a similar emotional response in an observer” (
The main endocrinological factor characterising PTSD/anxiety disorders is chronic amygdalar and stress response over-activity (
For PTSD, anxiety, and depression, there is an imbalance between amygdalar activity and prefrontal cortical activity–the amygdala is markedly over-active and the activation of the medial prefrontal cortex is to some extent impaired (
The SAM axis reacts instantaneously: within milliseconds of an exposure to a stressor, the sympathetic nervous system engages the adrenal medulla to produce adrenaline and noradrenaline (
The function of the SAM axis response is to prepare the organism for a sudden increase in energy demands, i.e., immediate fight-or-flight. Oxygen intake is increased by elevating respiration rate; heart rate is elevated; blood flow to the muscles and blood glucose levels are boosted to prepare for movement; processes not immediately necessary for survival (e.g., digestion, reproduction, growth) are inhibited; pain perception is blunted; and alertness and sensory/learning/memory functions are enhanced (
We investigated whether dogs could be trained to alert to the early onset of PTSD intrusion/hyperarousal symptoms by relying on olfactory cues. The goal was to determine whether dogs can detect and discriminate between breath samples of donors with a trauma exposure history and varying levels of PTSD symptoms, collected immediately before and during exposure to cues related to their personal traumatic experiences. This is the first study to investigate canine ability to detect stress volatiles theoretically involved in PTSD intrusion/hyperarousal symptoms.
We proposed three hypotheses. First, we hypothesised that at least some dogs can discriminate between breath samples collected from the same human donors during a personalised trauma cue vs. resting baseline. This requires that, when presented with the two samples, the canines can detect VOCs associated with stress-induced PTSD intrusion/arousal symptoms.
Second, we hypothesised that at least some dogs can detect the trauma cue breath samples across different individuals (or the same individual in different contexts) when presented with one sample (baseline or trauma cue) at a time. In real-life, dogs would not have simultaneous breath samples to compare but would be presented with one “sample”. Confirmation of this second hypothesis would provide considerably stronger evidence for training PTSD alert dogs.
Finally, we hypothesised a positive correlation between dogs' performance and donors' self-reported negative affect during the trauma cue exposure (indices of human distress when confronting trauma reminders).
Ethical approvals were obtained from the Dalhousie University Committee on Laboratory Animals and Nova Scotia Health Authority Research Ethics Board.
Human donors were recruited from a study on neurocognitive mechanisms underlying trauma–cannabis use links (
Breath samples were collected from 26 donors aged 20–53 years (mean age = 31.2). Eight were male, 18 female. The self-report PTSD Checklist (PCL-5) (
The mean PCL-5 score was 45.08 (median = 44, range = 25–73)—above the cut-point of 38 for a likely PTSD diagnosis (
To assess for cannabis' possible influence on dog performance, the self-report Cannabis Use Disorder Identification Test-Revised (CUDIT-R) (
Initially, 25 pet dogs of different breeds were enrolled and commenced scent-detection training at Dalhousie's Canine Olfaction (CO) Lab. Dogs were recruited through personal contacts, e-mails to guardians who had expressed interest in volunteering their dogs, and word-of-mouth. Four dogs who passed the training criteria (
Ivy is a spayed female working-line Red Golden Retriever who was 5–6 years old during this study. Callie is a spayed female mix of German Shepherd and Belgian Malinois who was 3–4 years old during this study. Callie's parents came from a working-line pedigree but were not active working dogs themselves. Both Ivy and Callie are companion dogs who live with their guardians and were brought to the CO Lab, one dog at a time, 1–3 day(s)/week. Callie had no experience in scent-detection work prior to this study. Ivy had participated in other CO Lab studies since she was 2 years old: discriminating ketogenic from normal cow breath and searching for endangered wood turtles. She had not previously worked with human breath samples.
The small number of dogs does not pose a limitation to this study: the goal was not to demonstrate that all dogs can detect the stress VOCs emanating from people with trauma histories in response to trauma cues but to provide evidence that some dogs can, at high accuracy. An increased number of dogs would not meaningfully contribute to testing this study's hypotheses. High attrition rates based on selection criteria and performance are typical in scent-dog training (
The parent study involved two stress-induction experimental sessions with the donors–the interview session (
In each session, two breath samples were collected via disposable medical-grade face masks. The experimenter was trained in sample collection procedures, wore a medical grade mask themselves, and used a new pair of disposable powder-free latex gloves while handling each mask. The experimenter recorded the duration each mask was worn, and this variable was used in subsequent analyses. After collection, masks were stored in separate marked 4 oz glass jars with plastic caps (
Face masks in the glass sample collection jars (left) and pieces of the masks in the glass vials (right).
All 26 humans participated in the interview session. A first mask (resting baseline) was supplied and worn at the session's start, while providing informed consent and completing questionnaires (
Breath samples were collected from 14 imaging sessions (from 13 donors
Unlike in the interview session, the baseline and trauma masks of the imaging sessions were not identical. All were medical-grade disposable masks of light blue colour and visually identical; however, the baseline mask for the imaging session was provided by the hospital and could have been manufactured by another company than the masks provided by our team during scanning. In addition, a single experimenter removed the metal wires from all imaging masks (for safety reasons during scanning), wearing latex gloves and a medical-grade mask. Dogs were only presented with fabric mask pieces (never the nose wire or ear loops). In every other respect, the sample collection/handling/storage procedures for the imaging session were identical to the interview session.
For both sessions, donors were instructed to: stay abstinent from alcohol/drugs for 12 h and caffeine for 2 h prior to the sessions (verified via self-reports and urine tests); not eat/drink anything but still water during each session; not wear make-up, perfume, or scented products on their face; and refrain from touching the mask. Donors mouth-rinsed with still water before donning the interview session baseline mask (mouth-rinsing at the hospital entrance was not practical for the imaging session). To minimise sample contamination, donors donned and removed masks by touching only the elastic ear loops.
In total, 40 breath sample sets from 26 individuals were collected. Of these 26, 13 donated one set (from the interview session), 12 donated two (from both sessions on non-consecutive days), and one donated three (one interview and two imaging session sets, all on separate days). The interviews allowed collection of individualised information to create standardised cues for the imaging session (
Samples were prepared for presentation to the dogs in the CO Lab by LK, wearing powder-free latex gloves and a medical-grade mask. The sample preparation began 15–30 min prior to dog testing. LK first disinfected the jars. Having changed gloves, she then procured two sterilised 20 ml glass vials (
After another glove change, LK removed/unfolded the baseline face mask from the jar and cut a small piece of fabric using sterilised scissors. The piece was inserted in the glass vial marked with “S−” using sterilised metal tweezers. She then returned the rest of the mask to the jar, changed gloves again, capped the vial and jar, and repeated the procedure with the trauma cue mask using a different set of sterilised scissors/tweezers. The two vials were sealed together in a Ziploc bag until testing began. The jars containing the masks were preserved in the fridge within their original bag immediately after sample preparation, for future use.
As no gold standard is currently available to determine whether a breath sample contains sufficient stress volatiles to confirm its “stressed” status, one can only indirectly assess the sample's validity by measuring the donor's physiological (e.g., heart rate, blood pressure, stress hormones), behavioural, or psychological stress indicators during sample collection. We measured the latter. During the interview session, the well-validated (
In the imaging session, a 9-item visual analogue scale (VAS) (
To assess for any putative influence of craving-related VOCs impacting dogs' performance, donors' cannabis craving during the baseline and trauma cue was evaluated using the well-validated Marijuana Craving Questionnaire-Short Form (MCQ-SF) (
Dog training/testing took place in the CO Lab. Dogs were trained and handled by LK under the supervision of SG (qualifications in
Details of the early scent-discrimination training of all the 25 dogs by following a Low Saliency Training (LST) protocol (
Experiment 1 training was conducted in a two alternative forced choice (2AFC) task, where the dog was presented with a line-up of two odours: the positive stimulus (S+) and the negative stimulus (S−). The dog had to find and signal the S+ by holding their nose for 5 s on the respective container (
Ivy signalling the S+ in the 2AFC task.
Human breath samples used for training were prepared in the manner described earlier (
The glass vials containing samples next to the metal funnels (left), and the glass vials placed underneath the funnels as the experiment set-up (right).
During between-session breaks, samples were capped to preserve the odour volatiles. Each set was used for 3–5 sessions (on a single day) with one dog only. Funnels were cleaned of dog saliva with a single dry paper towel between trials. To avoid olfactory confounding, both funnels were cleaned using the same paper towel even when only one had saliva on it. After each session, funnels were cleaned with a single paper towel infused with 70% isopropylic alcohol solution and let dry completely before the next session. After the dog finished working for the day, funnels were washed with clean water and disinfected with the isopropylic alcohol solution. The metal board holding the funnels was also disinfected.
To aid the dog in learning the target odour, the handler guided the dog to sniff both samples and then signal the S+ (only before the first session). Once the dog began to exhibit signs of identifying the target odour independently (usually after 1–3 mock trials), a session was commenced. To minimise handler influence, we used a double-blind protocol. The handler and the dog waited behind a closed door between each trial, while the assistant experimenter changed (or went through the motions of changing) the odour locations according to a predetermined randomised placement order. When the trial was set up, the assistant went behind a visual barrier and called, “Ready!” The dog handler then opened the door and directed the dog to sniff the samples by saying, “Go find”. The assistant did not see the dog and the handler did not know the placement of the target odour. When the dog made a choice, the handler called the position signalled (left or right). If the dog made the correct choice, the assistant experimenter pressed the clicker, and the handler gave the dog verbal praise and a treat. If the dog made an incorrect choice, the assistant experimenter said “Thank you” in a neutral but friendly voice. This functioned as a non-reward marker to signal the trial was over and the dog could stop signalling. Without giving a reward, and in a friendly manner, the handler then called the dog behind the door to await the next trial.
During training, error correction was sometimes allowed when the dog signalled an incorrect sample, by allowing the dog to figure out the correct answer themselves or by guiding them to it. The dog received a reward after an error correction, but the trial was recorded as incorrect. As the difficulty level was very high in training, error corrections were to reduce the dog's frustration/confusion and maintain motivation (
Dogs were considered trained and ready for testing once they began to regularly meet the passing-level criteria (80% accuracy in three consecutive sessions or 90% accuracy in two consecutive sessions) with novel donor samples and double-blind conditions. Only Ivy and Callie successfully completed this training phase. Ivy's training was complete with 12/40 sample sets (eight interview session, four imaging session) in 13 days over four months. Callie's training was complete with 16/40 sample sets (ten interview session, six imaging session) in 17 days over five months. The training period was long due to slow sample collection.
Training for Experiment 2 (yes/no detection) commenced after Experiment 1 was completed; only Ivy and Callie were trained for this task. At this point, both dogs were visiting the CO Lab 2–3 days/week. In Experiment 2, we presented one sample at a time and asked dogs to signal whether the presented sample was S+ or S− (
Callie signalling the S− in experiment 2 (handler in the background).
The signal for S+ (i.e., “yes”) remained the same: a 5-s nose-hold on the funnel covering the target odour. To signal S− (i.e., “no”), we trained dogs to sit next to the sample (
At first, sessions were not double-blind. The handler guided the dog towards the correct answer if the dog appeared confused; error corrections were used more liberally. To make the task easier for the dogs in the beginning, only one donor's sample set was used over the whole session and the vial containing the S− was placed on the side, while the S+ vial was always presented upright. This difference was not visible to the dogs (the samples were covered by the funnel) but placing the S− vial on the side lowered its saliency. Once the dog was able to meet the passing-level criteria for such sessions in double-blind conditions, both the S+ and S− vials were placed upright. When the dog also passed that level in double-blind conditions for at least one sample set, the next step involved sessions with multiple individuals' sets. One session always had an equal number of S− and S+ samples. Dogs were considered ready for testing when their performance was above chance in double-blind conditions. Ivy was trained for the detection task using seven sample sets (two presented only in multiple-sample-set sessions), and Callie was trained on 11 sets (seven presented only in multiple-sample-set sessions).
The testing equipment was identical to the training equipment. The testing always followed rigorous double-blind methods and the sample position/condition was always randomised. All samples presented to dogs were newly prepared–mask pieces and vials used during the training were not re-used for testing. Thus, although the chemical composition of some of the samples was the same as the dogs had encountered during training (see
Experiment 1 targets the first hypothesis. In a 2AFC task, we tested whether dogs could discriminate between breath samples collected from the same donor during the trauma cue and baseline conditions. During this experiment, both dogs were tested on all 40 sample sets. To get sufficient data to adequately evaluate their olfactory discrimination abilities for every individual set, each dog performed three to five sessions (i.e., 30–50 trials) per set. The samples were collected over the course of 10 months. The testing was determined by the pace of sample arrival. After the initial months of training, the testing period for Experiment 1 lasted six months for Ivy and five months for Callie. As Ivy was trained using 12 sample sets, the rest of the 28 sets were completely novel to her during testing. For Callie, 24 sets were completely novel during testing.
All the samples used for training and most of the samples used for testing in Experiment 1 were presented to the dogs either on the same day or within three days of collection. However, the samples used in training (i.e., samples collected first) were presented for testing last–not knowing their expiration date, new samples were prioritised over the old ones during the testing phase. Some of these last-tested samples were thus up to eight months old. During the testing phase of the study, the dogs visited the CO Lab 2–3 times a week for a 2 h workday. Both dogs had occasional days where they did not feel like working which often presented in the dogs' poor performance. To rule out the possibility that poor performance on any particular sample set was caused by a “bad day” or sample contamination instead of the samples' olfactory properties, all sets with which the dog exhibited inconsistency (sessions with very high accuracy followed by sessions with poor accuracy) and low motivation to work (i.e., expressing dissent
The testing procedures for this task were identical to the double-blind training phase procedures described in
The second nuance different from the training procedures involved recording the dogs' sample sniffing order during each testing trial as a double measure to ensure that they regularly checked both samples (although checking both samples was not a requirement to complete a trial) and to avoid sample contamination caused by one sample getting more exposure to the dog's breath than the other. In case saliva was spotted in a vial, both samples were discarded, and a whole new set of the same donor's masks was prepared to continue testing. To minimise handler influence, it was opted to not guide the dogs into checking both samples. The dogs were allowed to re-check samples before committing to an answer.
Experiment 2 targets the second hypothesis–in the yes/no task, we tested the dogs' ability to generalise, i.e., whether they could tell apart the trauma cue and baseline breath samples across different individuals when presented with one sample at a time. Furthermore, it was a confirmatory test of the dogs' sensitivity (detectability of the putative volatiles, as opposed to discriminability) with the same sample set. Each dog was tested on 21/40 sample sets in Experiment 2. As there was no objective measure to confirm the existence of stress volatiles in the trauma cue mask, the sets used in this experiment had to be confirmed as differentiable by the dogs in Experiment 1. Of all the differentiable sets, 26 with the highest performance accuracy were selected. Of these 26 sets, each dog was presented with 21 sets: 16 were common to both Callie and Ivy and five were selected individually. The reason why Experiment 2 did not involve all 40 sample sets pertains to dog welfare. By Experiment 2, Ivy began expressing sustained dissent to participating in further testing (even though she had enjoyed working throughout all the prior phases of this study). To respect her sustained dissent, it was decided to cease testing with her after roughly half of the samples had been tested. To keep the conditions equivalent for the dogs (and allow for adequate comparison of their results), Callie was tested on the same number of samples.
The testing procedures remained identical to the double-blind training procedures (all samples placed upright). No cue was used. Testing for Experiment 2 lasted two days for Ivy and one day for Callie. The testing was conducted in four sessions. As there were 21 sample sets, three sessions included five sets (five S+ and five S− samples) and one session included six sets (six S+ and six S− samples). Each sample was presented once during the session. Thus, three sessions required 10 trials and one session required 12 trials. By the time of Experiment 2, the dogs had already encountered the masks of all 40 sample sets in the previous experiment but, again, new pieces of these masks were prepared for the yes/no task for each dog. In Experiment 2, the sample age varied from 2.5 weeks to 10 months. The last encounter with these samples in Experiment 1 varied from two weeks to seven months, whereas among the samples not used during the training for Experiment 2, the longest time since the last encounter was six months.
Experiment 1 results were analysed by calculating each dog's overall accuracy and conducting binomial tests. Overall accuracy is the percentage of correct responses throughout all trials with all donor samples. The binomial test was based on the number of successful trials within the total number of trials, while the probability of success in each trial was 0.5, and the alpha-level, 0.05.
Experiment 2 gave additional information. Besides calculating the dogs' accuracy and conducting binomial tests, the yes/no detection task design enabled implementing Signal Detection Theory (SDT) (
For hypothesis 3, dogs' accuracy across the trials of each sample set in Experiment 1 was correlated with donors' distress scores during sample collection (interview session: trauma cue PANAS negative affect, State Anxiety, and each item score; imaging session: trauma cue VAS negative affect, and each item score). Correlations were calculated both using Pearson's r and the non-parametric Spearman's rho, which has less statistical power than Pearson's but is also less sensitive to the outliers and, thus, more reliable. Both were calculated using Jamovi (version 2.3.16.0). Supplementary analyses were conducted to identify any other potential variables affecting dogs' ability to discriminate between baseline and trauma samples (e.g., PTSD symptom severity, PTSD diagnostic status;
In Experiment 1, both Ivy and Callie successfully discriminated between trauma and baseline breath samples of 36 sets. Both dogs had four sets where they performed at (or below) chance. For Callie, one of the non-discriminable sample sets belonged to her guardian (nevertheless, Callie is a regular companion dog, not a trained service dog). However, the guardian's samples represented a confound for Callie: firstly, Ivy's accuracy on that sample set was 100%, indicating the samples could have been discriminable; secondly, the preliminary analysis between Callie's performance and donor outcome measures revealed some inconsistencies in the case of her guardian's samples; and thirdly, Callie's poor performance contradicted the guardian's testimony about Callie's recent receptiveness (seemingly developed as a result of participating in this study) to the guardian's PTSD-related arousal at home. Thus, the guardian's samples were removed from Callie's data. The final number of sample sets in Experiment 1 was 39 for Callie and 40 for Ivy; the number of non-discriminable sets was three for Callie (two from the interview session, one from the imaging session) and four for Ivy (two from the interview session, two from the imaging session). For both dogs, the non-discriminable interview sample sets were all from different people. In case of the non-discriminable imaging samples, the dogs had been able to discriminate the same donor's interview samples. Only one non-discriminable set (from the imaging session) was common to both dogs.
Both dogs discriminated with 100% accuracy for nine sample sets (two common to both dogs). For Ivy, two out of these nine sets had been used during training (using different mask pieces/vials). For Callie, five out of the nine sets had been used during training. To rule out sample contamination or the “bad day” effect, Ivy was re-tested on seven sets (using new mask pieces/vials). Ivy had five “bad days” during the study, while two of the re-tested sets proved to be truly non-discriminable for her. Two sets had to be re-tested with Callie. Callie had only one “bad day”, while the other set turned out to be non-discriminable for her.
The dogs' performance in the 2AFC task is shown in
The dogs’ performance in the 2AFC discrimination task of Experiment 1.
| Dog | No. of sample sets presented | No. of trials | No. of success | Binomial test |
Overall accuracy | Accuracy across discriminable sample sets ( |
|---|---|---|---|---|---|---|
| Interview samples only | ||||||
| Ivy | 26 | 930 | 841 | 90.81% | 93.66% (24) | |
| Callie | 25 | 990 | 863 | 88.19% | 90.86% (23) | |
| Imaging samples only | ||||||
| Ivy | 14 | 500 | 426 | 86.98% | 93.64% (12) | |
| Callie | 14 | 550 | 497 | 91.86% | 94.62% (13) | |
| Interview and imaging samples combined | ||||||
| Ivy | 40 | 1,430 | 1,267 | 89.47% | 93.65% (36) | |
| Callie | 39 | 1,540 | 1,360 | 89.51% | 92.22% (36) | |
| Total | 79 | 2,970 | 2,627 | 89.49% | 92.94% (72) | |
Ivy performed 1,430 trials and Callie 1,540 trials in Experiment 1 (
Overall, the trauma cue exposure via interview (interview session) increased the donors' experienced negative affect generally and State Anxiety specifically, and the brief audiovisual trauma cue exposure in the scanner (imaging session) increased their negative affect (
To test hypothesis 3, each dog's performance was correlated with the donors' emotional responses to the trauma cues separately for the interview and imaging sessions. Given directional predictions were made
Ivy's accuracy % (left) and Callie's accuracy % (right) in the 2AFC task in relation to the donors’ self-reported State Anxiety scores during the interview session's trauma condition. The maximum value of the Y-axis is 100. The scale on the graph accommodates the standard error bar.
Ivy's accuracy % (left) and Callie's accuracy % (right) in the 2AFC task in relation to the donors’ self-reported PANAS negative affect scores during the interview session's trauma condition. The maximum value of the Y-axis is 100. The scale on the graph accommodates the standard error bar.
The dogs’ accuracy in the 2AFC task in relation to the donors’ self-reported negative affect (total and items) after the trauma cue.
| Ivy | Callie | |||||||
|---|---|---|---|---|---|---|---|---|
| Outcome measures | Pearson's r | Spearman's rho | Pearson's r | Spearman's rho | ||||
| Interview | ||||||||
| PANAS NA | 0.444 |
0.011 | 0.151 | 0.230 | 0.124 | 0.278 | 0.037 | 0.430 |
| State Anxiety | 0.617 |
<.001 | 0.466 |
0.008 | −0.021 | 0.540 | −0.202 | 0.833 |
| Distressed | 0.422 |
0.016 | 0.320 | 0.055 | −0.258 | 0.894 | −0.317 | 0.939 |
| Scared | 0.390 |
0.024 | 0.349 |
0.040 | 0.287 | 0.082 | 0.114 | 0.294 |
| Afraid | 0.182 | 0.186 | 0.140 | 0.247 | −0.238 | 0.874 | −0.337 | 0.951 |
| Nervous | 0.376 |
0.029 | 0.321 | 0.055 | 0.179 | 0.196 | 0.209 | 0.158 |
| Jittery | 0.514 |
0.004 | 0.282 | 0.082 | 0.026 | 0.451 | −0.073 | 0.635 |
| Upset | 0.548 |
0.002 | 0.419 |
0.017 | −0.023 | 0.544 | −0.028 | 0.553 |
| Guilty | −0.051 | 0.597 | −0.256 | 0.897 | 0.087 | 0.339 | −0.026 | 0.549 |
| Ashamed | 0.113 | 0.291 | −0.166 | 0.792 | 0.356 |
0.040 | 0.263 | 0.102 |
| Irritable | −0.069 | 0.631 | −0.154 | 0.773 | 0.070 | 0.370 | 0.002 | 0.497 |
| Hostile | 0.337 |
0.046 | 0.251 | 0.108 | 0.141 | 0.250 | 0.099 | 0.318 |
| Imaging | ||||||||
| VAS NA | 0.544 |
0.042 | 0.531 |
0.046 | 0.140 | 0.341 | −0.282 | 0.799 |
| Anxious | 0.409 | 0.106 | 0.398 | 0.113 | 0.224 | 0.253 | −0.180 | 0.701 |
| Upset | −0.060 | 0.569 | −0.025 | 0.530 | 0.454 | 0.080 | 0.187 | 0.291 |
| Irritable | 0.225 | 0.253 | 0.440 | 0.088 | 0.413 | 0.104 | 0.016 | 0.481 |
| Down | −0.008 | 0.510 | −0.105 | 0.620 | −0.020 | 0.524 | −0.108 | 0.624 |
| Uncomfortable | 0.453 | 0.081 | 0.266 | 0.214 | −0.386 | 0.879 | −0.605 | 0.976 |
Due to technical errors, trauma cue VAS negative affect and its items were recorded for only 11 of the 14 donors.
Ivy's accuracy % (left) and Callie's accuracy % (right) in the 2AFC task in relation to the donors’ self-reported VAS negative affect scores during the imaging session's trauma condition. The maximum value of the Y-axis is 100. The scale on the graph accommodates the standard error bar.
Ivy's accuracy % (left) and Callie's accuracy % (right) in the 2AFC task in relation to the donors’ self-reported feeling of shame during the interview session's trauma condition. The maximum value of the Y-axis is 100. The scale on the graph accommodates the standard error bar.
In supplementary analyses, we examined whether the dogs' performance was correlated with PTSD symptom severity or PTSD diagnostic measures. To evaluate whether the dogs responded to putative cannabis-related VOCs, correlations were run with donors' CUD symptom severity and cannabis craving to trauma cue exposure across the interview and imaging sessions. No significant correlations were detected (
Finally, a post-hoc analysis examined whether donors' PTSD symptom severity might have indirectly influenced dog performance via donors' negative emotional reactions to the trauma cues. Donors’ interview-based PTSD symptom count was significantly correlated with their trauma cue State Anxiety score (
The dogs' performance in the yes/no detection task is shown in
The dogs’ performance in the yes/no detection task of Experiment 2.
| Ivy | Callie | |
|---|---|---|
| No. of samples presented | 42 | 42 |
| Trauma samples | 21 | 21 |
| Baseline samples | 21 | 21 |
| Correct responses | 31 | 34 |
| Hits | 16 | 15 |
| Misses | 5 | 6 |
| False alarms | 6 | 2 |
| Correct rejections | 15 | 19 |
| d' | 1.278 | 1.875 |
| C | −0.073 | 0.372 |
| Accuracy | 73.81% | 80.95% |
| Sensitivity | 76.19% | 71.43% |
| Specificity | 71.43% | 90.48% |
| Precision | 72.73% | 88.24% |
| Binomial test |
Ivy's overall accuracy was 73.81% (range = 66.67%–90%) and Callie's was 80.95% (range = 70%–91.67%). Ivy's sensitivity and specificity were similar: she correctly identified the positive samples in 76.19% of their presentations and the negative samples in 71.43% of their presentations. Ivy's very low bias (C) suggested she is an “ideal observer”. Callie demonstrated higher specificity than sensitivity: she correctly identified the negative samples in 90.48% of their presentations, but the positive samples in 71.43% of their presentations. Her positive response bias (C) indicates a conservative decision-making strategy: when in doubt, Callie was more inclined to signal the stimulus was negative.
On no occasion was either dog able to correctly identify one person's samples collected from one session but completely unable to identify their samples from a different session. When analysing the dogs' performance with samples collected from the same donor on different stressful events, out of 20 presentations (i.e., five donors with two sets or four samples), Ivy correctly identified 3/4 samples of four donors and 4/4 samples of one donor. In Callie's case, out of 26 presentations (i.e., five donors with two sets or four samples; and one donor with three sets or six samples), she correctly identified 4/4 samples for two donors, 3/4 samples for another two donors, 2/4 samples for one donor (from different sessions), and 5/6 samples for the one donor who participated in two imaging sessions.
Experiment 1 confirmed our first hypothesis by providing evidence that some dogs can learn to discriminate between human breath samples collected during a relatively relaxed state and during induced stress designed as an experiential analogue to a PTSD state of intrusion or hyperarousal. The dogs' performance did not depend on the length of time the donor spent wearing the mask nor on the type of experimental session (interview or imaging) from which the samples were collected. The sample collection, preparation, and handling procedures generally did not allow for any systematic differences between the baseline and trauma cue samples besides the target odour (but see
Experiment 2 confirmed our second hypothesis by providing evidence that some dogs can detect the trauma cue samples from the pool of trauma cue and baseline breath samples of different individuals when presented with samples independently. Furthermore, our design where the donors underwent trauma cue exposure in two settings (interview vs. in scanner) using two cue exposure methods (interview vs. brief audiovisual cue) allowed demonstration that the dogs can perform this detection for the same individual across different stressors.
Depending on the stressor and its context, stress responses can vary in speed and magnitude of the released hormones, whether they induce adrenaline and/or noradrenaline, and which other hormones are activated (
The third hypothesis of a positive correlation between the dogs' performance and the donors' self-reported negative affect during the trauma cue condition was partially confirmed. One dog's (Ivy's) ability to discriminate between the baseline and trauma cue samples in Experiment 1 was correlated with the donors' overall PANAS and VAS negative affect scores during the trauma cue conditions of the interview and imaging sessions, respectively, demonstrating Ivy's consistency in what she was detecting across the two stress-induction environments.
Both dogs' Experiment 1 performance with the interview session samples had correlations with some of the PANAS individual negative affect items. Ivy's performance showed strong positive correlations with most PANAS items reflecting humans' anxiety (i.e., distress, scared, nervous, upset, jittery), and with the State Anxiety composite, during the trauma cue exposure. Although Callie showed equally high overall accuracy as Ivy, the only human negative affect item to which she seemed sensitive was shame–one of the self-conscious emotions (
Despite the dogs' success in generalising the target odour in Experiment 2, results do not support the assumption that PTSD as a diagnosis has a disease-specific signature VOC profile as PTSD symptom count, severity, and diagnoses showed no significant relation with either dog's Experiment 1 performance. The commonality of the donors in this study was trauma exposure rather than PTSD, as 12 of the 26 donors had symptoms below the diagnostic level (CAPS-5 clinical interview). In the context of PTSD service dogs, the dogs' inability to discriminate between trauma-exposed people with and without a current PTSD diagnosis is irrelevant as the dogs would not be utilised for diagnostic purposes but, instead, as alert dogs to facilitate management of PTSD symptoms which can occur in both people with clinical PTSD and, to a lesser (but still potentially disruptive) extent, among trauma survivors with subthreshold PTSD symptoms. It is the volatiles associated with the emotional reactions to the trauma cue that the dogs should pick up on. Our results support this proposition.
However, PTSD symptoms were still a distal predictor of dog performance. For both dogs, donors' greater PTSD symptoms were indirectly associated with greater dog performance by way of donors' greater levels of specific negative affective responses to the trauma cue (State Anxiety for Ivy, shame for Callie). This suggests alert dogs may perform better with individuals with PTSD given their greater emotional responses (and presumably higher emission of VOCs) to trauma reminders.
It is often hypothesised that scent-detection dogs search for specific components or concentrations of a target odour rather than its complete signature (
Breath samples collected in the current study contain putative VOCs possibly emanating from any of the involved endocrine subsystems (
The strong positive correlation between Ivy's performance and the donors' State Anxiety indicates her sensitivity primarily to the SAM axis volatiles. This is consistent with a prior study (
A number of papers have suggested using a true detection (yes/no) task (
The yes/no procedure allows testing for generalisation without the possibility of the dogs relying on working memory or odour comparison. Generalisation is a result of training when the dog spontaneously learns to respond to target odour variations by learning their common properties (
As we were investigating PTSD alert dog functions, ideally, we would have focused on testing dogs' ability to alert to one person's different episodes of distress/arousal with more sample sets from each donor than we had available. However, given the high interpersonal variability of human hormonal stress response, generalising the odour across individuals should be a more arduous task than generalising across different events of one individual. For example, detecting hypoglycemia in the breath samples of different people is a harder task for dogs than detecting separate hypoglycemic events in the same individual (
Experiment 1 also provides evidence for generalisation: Ivy was able to successfully find the trauma cue sample in the sets of 24/26 donors and Callie in 23/25 donors. Moreover, with one exception, the dogs were able to successfully discriminate between the stress and baseline samples of the same person collected during different stressful events. Hence, the combined evidence of generalisation from both Experiment 1 and 2 endorses the possibility of training PTSD service dogs to use their acute olfaction to alert to the patient's episodes of trauma cue-related distress.
In Experiment 1, theoretically the dogs could have memorised the positive sample from each set in the first session and based their performance during the rest of the sessions on that learning–a significant confound in the field of scent-detection (
One of the main disadvantages of the mAFC procedure is dogs could learn to perform by comparing samples rather than evaluating each independently (
Another advantage of the yes/no procedure is its suitability for SDT, offering information about dogs' sensory sensitivity, specificity, and response bias (
Ivy's 74% and Callie's 81% overall accuracy in Experiment 2 (although lower than their 90%+ accuracy in Experiment 1) demonstrates that, despite only two weeks of training for this significantly harder task, the dogs still performed at levels comparable to the capabilities of current technology (
Ivy had four and Callie three non-discriminable sets (Experiment 1). Myriad reasons independent of the samples' odour properties may explain dogs' low performance in scent-detection tasks. Both internal [e.g., medications, hydration, changes in blood flow (
First, in the case of non-discriminable samples, the dog may have found the person's scent attractive, aversive, or otherwise distracting (
Unfortunately, the pool of non-discriminable samples in this study was too small to provide any firm evidence to support the above-mentioned propositions. Interestingly though, in all but one case, the non-discriminable samples for one dog were discriminable for the other. Reasons why one imaging session set was non-discriminable to both dogs could vary from the donor's diet and cosmetic products to them having high anticipatory anxiety prior to entering the scanner leading to indistinguishable olfactory properties of the baseline and trauma cue masks.
The methodology of canine scent-detection studies ought to be scrutinised for any systematic differences between the target and non-target odours to avoid biased or confounded results. Even though great effort was put into conducting the experiments in a way that would minimise the risks of bias or confounds, several potential sources remain. When scrutinising sample collection methods, one could argue that, in case of the imaging sets, the smell of the scanner, and the fact that the imaging session baseline and trauma cue masks were not identical (possible different manufacturers; lack of metal wire in trauma mask) could have caused confounding. Furthermore, during the imaging session, the donors' experienced stress induced by the cue exposure vs. the scanner could not be differentiated. Moreover, since donors had the choice to opt out from the imaging session (
However, there was no indication of dogs being able to discriminate the imaging session samples significantly better/worse than interview session samples. Thus, it is plausible to assume that the dogs learned to ignore any systematic differences between the imaging session trauma cue and baseline masks and focused on the target odour instead. Any donor-specific differences and the added stress due to the brain scanning appears to not have affected the dogs' ability to discriminate between the trauma cue and baseline samples when compared to their performance with the interview samples.
Another potential methodological criticism involves the trauma cue mask also being worn during the neutral and cannabis cues. It is doubtful that the neutral cue could have presented a confound. However, if there are distinct putative VOCs emanating during cannabis craving, the cannabis cue could have presented a confound. As mentioned, the dogs’ performance was neither correlated with the donors' CUD symptom severity nor with their cannabis craving combined across sessions.
Some procedural aspects may have confounded results. Although measures were taken to avoid sample contamination with dogs' breath and saliva (
Another example is the repeated use of samples within one testing day in Experiment 1. Various biomedical scent-detection studies [e.g., (
Lastly, the use of self-report to measure human response to the trauma cue exposure could have minimised correlations between the dogs' performance and donors' response. Although we selected measures with good psychometric properties, and tapped subjective experience to which only the individual has access (
Several limitations should be considered when interpreting study results. First, although the psychological distress measures proved valuable and informative, they have a relatively high risk of bias. Future studies should corroborate subjective self-report findings with more objective physiological stress measures and blinded behavioural observations to confirm each sample's status as target or non-target. A predetermined sample status would facilitate dog training, provide further insight into dogs' performance, and eliminate mistaking the dogs' correct responses as false (e.g., when the baseline is equally or more “stressed” than the experimental condition). However, the endocrinology of human stress response varies so widely that a stress sample can include the target odour or not depending on the VOCs to which the dog is sensitive. To get better insight into which stress-related VOCs dogs are detecting, future studies should measure the changes in the human donors' endocrinological profile (including SAM and HPA axis hormones) during the trauma cue vs. baseline conditions.
Another limitation is that most samples likely included both early and late stress VOCs as the donors wore the trauma mask for a relatively long duration. Results provided indirect evidence for canine olfactory ability to detect early-onset stress experienced by PTSD patients in response to trauma cue exposure: strong correlation of Ivy's performance with the donors' State Anxiety (associated with SAM axis hormones); and both dogs' abilities to detect stress VOCs in the breath of donors whose self-reported negative affect during the trauma cue was equal to or lower than at baseline. Given PTSD alert dogs must detect early-onset stress VOCs, future studies should test dogs' olfactory acuity on samples collected very shortly (mere minutes) after the trauma cue.
This study may also be criticised for using donors who were regular cannabis users. Although cannabis use was a constant across donors and conditions, it is still possible that our results only generalise to donors who are chronic cannabis users due to alterations in their endocrine stress response. However, chronic cannabis use has been associated with a blunted HPA axis response and heightened baseline cortisol (
Further limitations include the repeated use of training samples during testing and using the discrimination task samples again in the detection task. Ideally, training samples would not be reused for testing (
There is no concrete standard on how many unique samples are required for training scent-detection dogs for biomedical tasks (
This study is the first to demonstrate canine olfactory ability to detect putative VOCs emitted by people with trauma histories when experiencing distress theoretically associated with PTSD intrusion/hyperarousal symptoms. These results are consistent with other studies that have provided preliminary evidence of dogs' ability to detect stress in humans by using their olfaction. We add to this evidence-base by expanding from stress to trauma (all donors reported trauma histories and more than half had PTSD) and employing well-validated trauma-related stress-induction measures. Our results do not suggest that PTSD has a disease-specific signature VOC profile. Instead, the results are aligned with the premise that the endocrinology of the acute stress response to exposure to personalised trauma cues can be detected by dogs. Moreover, the results demonstrated that dogs can generalise the olfactory biomarkers of this stress response, making the idea of PTSD alert dogs plausible.
However, the olfactory biomarkers that dogs generalise may vary, with different dogs detecting different olfactory biomarkers in human trauma cue breath samples. While both dogs performed at high accuracy in the 2AFC odour discrimination task of Experiment 1, the differences in correlation of the dogs' performance with the donors' specific distress measures indicate dogs may have been attuned to separate endocrine stress markers. Although changes in endocrine parameter levels in donors were not recorded, we hypothesise that one dog was detecting the SAM axis hormones, while the other was sensitive to the HPA axis hormones. Future studies should test this hypothesis by measuring levels of the respective hormones. As the concept of PTSD alert dogs requires detecting the early onset of stressful arousal, canine ability to detect the SAM axis hormones is particularly important to confirm.
The yes/no detection task in Experiment 2 brought the experimental conditions closer to real-world settings for PTSD alert dogs. Furthermore, Experiment 2 was the first to examine dogs’ sensitivity and specificity in detecting putative human stress VOCs. Results showed that dogs possess a comparable or higher sensitivity and specificity than current technology (e.g., solid-phase microextraction, gas chromatography-mass spectrometry, and “electronic noses”). Dogs have other advantages over technology in assisting people with PTSD as dog functions extend far beyond alerting to and distracting from upcoming distress episodes [e.g., offering unconditional love and support (
In conclusion, this study provides information on mechanisms underlying the alert function of PTSD service dogs, as well as evidence for the possibility of training dogs to detect upcoming distress episodes through the person's breath. This skill would enhance dogs' alert function by enabling even earlier distraction (i.e., before the onset of observable cues), potentially leading to the person's enhanced comprehension of their symptoms, more efficient application of coping skills learned in therapy, and prevention of the episode from “spiralling”. Nevertheless, as this is a proof-of-concept study, validation studies are required to confirm these promising results.
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.
The studies involving humans were approved by Dalhousie University REB (Research Ethics Board). The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study. The animal studies were approved by Dalhousie University UCLA (University Committee on Laboratory Animals). The studies were conducted in accordance with the local legislation and institutional requirements. Written informed consent was obtained from the owners for the participation of their animals in this study.
LK: Conceptualization, Formal Analysis, Investigation, Methodology, Project administration, Supervision, Writing – original draft, Writing – review & editing. SS: Conceptualization, Formal Analysis, Investigation, Methodology, Supervision, Writing – review & editing. SG: Conceptualization, Investigation, Methodology, Project administration, Supervision, Writing – review & editing.
The author(s) declare that no financial support was received for the research, authorship, and/or publication of this article.
The authors want to thank Pars Atasoy, Pamela Collins, Tessa Cosman, Sarah DeGrace, Mikaela Ethier-Gagnon, and Jane Girgulis for their work as research assistants, and Dr. Pablo Romero-Sanchiz as an invaluable research associate. We also want to thank the volunteers in the Canine Olfaction Lab: Sofia Baidón de Rodriguez, Tessa Cosman, Owen Johnstone, Samantha Pratt, Starr Lawlor, Kate MacDiarmid, and Fanny Vain. We obviously want to thank the dogs, Callie and Ivy, and their guardians. The authors would also like to extend their thanks to the many volunteers, dogs and their guardians who participated in the initial selection and dog-training process for this study.
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
The author(s) declared that they were an editorial board member of Frontiers, at the time of submission. This had no impact on the peer review process and the final decision.
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
The Supplementary Material for this article can be found online at:
1This works the other way around as well: an aroused sympathetic nervous system stimulates the amygdala (
2Due to technical errors with the imaging equipment in one donor's first imaging session, a significant amount of parent study data (
3Examples of how the two dogs in our study expressed dissent when asked to go sniff the samples include the following behaviours: remaining in the waiting area and sitting/lying down and/or staring at the handler; going to the assistant experimenter instead; sniffing the floors/walls/doors of the experiment room instead of samples; pawing at the door leading to the rest area; initiating play; mounting the handler and thrusting; restlessness (motor); sweaty paws (leaving paw prints on the floor); whining/whimpering; yawning; lip-licking; barking. The presence of two or more such behaviours in the beginning of testing day (already during the first session) was interpreted as indicative of the dog's low motivation to work on a given day.
4Accuracy % = (hits + correct rejections)/(hits + misses + correct rejections + false alarms) × 100
Sensitivity % = hits/(hits + misses) × 100
Specificity % = correct rejections/(correct rejections + false alarms) × 100
Precision % = hits/(hits + false alarms) × 100
C = –(Zhits + Zfalse alarms)/2
d’ = Zhits – Zfalse alarms
The