Kayla Fratt

Peer reviewed

“Canines seem to detect coronavirus infections with remarkable accuracy, but researchers say large-scale studies are needed before the approach is scaled up.” States an article published in Nature in November 2020.1 However, there is a massive gap between currently published research and operational use of COVID-19 sniffer dogs.

Around the world since early 2020, dog trainers and researchers have been racing to see whether or not dogs can be used to screen people for COVID-19 infections. As of January 2021, very few studies have actually been published on the efficacy of COVID sniffing dogs. The two referenced most heavily here are listed as a “pilot study” and “proof-of-concept,” respectively. The abstracts of two pre-print articles are available online but do not provide sufficient detail to delve into here.


Detection dogs have been used for a wide variety of medical tasks, including diabetes,2 prostate cancer,3 malaria4 and Clostridium difficile.5 Despite promising results in several studies, medical detection dogs are far from widely implemented. This is likely because of similar hurdles and logistical challenges facing the widespread implementation of COVID-sniffing dogs. The stated goal of many programs is to deploy COVID-sniffing dogs in airports, sport stadiums, and/or concert venues.

Pathogens and other target odors are identified by dogs thanks to specific patterns and compositions of volatile organic compounds (VOCs). Since viruses like coronavirus do not have their own metabolism, it is suspected that the most salient VOCs are actually related to the host’s body’s response to the infection rather than from the virus itself.6

While the current research is promising, tightly controlled laboratory conditions are far from real-world applications. Studies must be at least single-blind (where the handler and dog are unaware of the location of the positive sample) and ideally double-blind (where any other staff present are also unaware). Studies also should aim to present a given sample to each dog one time only to avoid the dog memorizing the odor profiles rather than actually recognizing the VOCs from COVID-19. All possible efforts should also be made to ensure that COVID-19 testing status is the only variable that differs between samples. If, for example, all positive COVID-19 cases are also from hospitalized males over 80, the negative samples should also be from hospitalized males over 80.7

Logistical Barriers to Implementation of COVID-Sniffing Dogs

  • There are potential risks to the dogs and human researchers when collecting and handling positive or unknown COVID-19 samples.
  • Current “gold-standard” tests for COVID-19 have varying degrees of accuracy, making it difficult to confirm the dogs with 100% certainty
  • Training for novice dogs will generally take 3-6 months and will take 6-8 weeks for an experienced detection dog. Many detection programs have washout rates exceeding 50%. This means that once a protocol is identified, scaling up in the number of dogs will still be challenging.
  • All of the studies referenced so far provide the dog with at 1 positive sample per 3-7 negative samples. A field application will have rates of find/reward much lower than this. Maintaining motivation in a field setting will therefore be challenging.
  • It is unknown how dogs may react to mild cases of COVID-19, asymptomatic cases, or pre-symptomatic. Both studies referenced below utilized hospitalized symptomatic cases for positive samples. If there is a difference in odor from hospitalized (and therefore severe) cases versus asymptomatic carriers, results may change. Given that detecting severe, hospitalized patients is far less useful than detecting asymptomatic or pre-symptomatic or even mild cases, upcoming studies must investigate these cases as well.
  • It is also unknown how dogs may react to a person who previously had COVID-19 but has since recovered.
  • Further testing with larger sample sizes of both dogs and people will ensure that the dogs are not just memorizing odor of given patients and reduce the likelihood that the dogs are actually learning VOCs that correlate with factors aside from COVID-19 status (such as age, other medical conditions, or hospital settings).

Much of the current information and news available about COVID-sniffing dogs are based on unpublished data reported by the same researchers and trainers involved in training the dogs. Rather than delving too deeply into those sources, today we are turning our attention to the published research on COVID-sniffing dogs.

Scent dog identification of samples from COVID-19 patients – a pilot study

By Jendrny et al, 2020, in BMC Infectious Diseases8

In this paper, researchers describe the results of training 8 Labradors for one week on saliva or tracheobronchial swabs from patients infected with SARS-CoV-2. The study was randomized and double-blind over the testing of 1012 randomized samples presented 10,338 times total. The dogs were able to discriminate between infected and non-infected individuals at an average diagnostic sensitivity of 82.63% and specificity of 96.35%. Sensitivity refers to the ability to correctly identify positive (infected) samples, while specificity refers to the ability to correctly identify negative (non-infected) samples. In other words, the dogs were ~82% accurate identifying positive samples and ~96% accurate identifying negative samples.

The dogs were trained using saliva and tracheobronchial samples from hospitalized patients infected with SARS-CoV-2 confirmed through nasopharangyal swabs; negative samples were obtained from people with a negative PCR test and no history of COVID-19 symptoms and without other cold or infection at the time. Samples were neutralized to reduce risk to human trainers and the dogs. The dogs were trained using 100μl of sample pipetted onto a cotton swab and placed in a glass tube.

The dogs were trained using a system that presents the dog seven scent holes. The dogs were trained for two weeks to habituate to the system, then trained for five days to get their results above chance for detection (though further detail on how this training was done was not provided). The dogs were automatically rewarded with a toy or ball after a sustained nose-touch response (1-2s) and the handler (standing behind the dog) was unaware of which scent hole contained the target. Only one positive sample was presented at a time. The study does not state if novel samples were reserved for testing, though novel samples were reserved for later in training to test if the dogs had successfully generalized in their detection of positive samples.

The dogs showed no difference in ability to identify saliva versus tracheobronchial samples. All of the dogs showed a high specificity (they almost never identified a negative sample as positive) but there was variability in sensitivity (some dogs were much more likely than others to incorrectly identify a positive sample as negative). The study points out that the “gold standard” test of RT-PCR for COVID-19 can have a false detection rate of up to 25%, even in trained hands – so the dogs still outperformed this test.


Training dogs to sniff out samples in a lab environment is very different from deploying them in a sports stadium or airport. One of the main reasons to be excited about using dogs for sniffing out COVID is their speed and the potential to use dogs to screen large groups of people at events or transportation hubs. However, this study did not test dogs on their ability to sniff COVID from patients or their vapor wakes. A dropoff in the dog’s accuracy is to be expected when transitioning from carefully controlled samples to a chaotic “real-world” application. The dogs were only trained and tested using hospitalized COVID patients, so it is still unknown how they may respond to asymptomatic or mild infections.

The study also did not mention providing the dog with blank rounds, which is another cause for concern. The dogs were essentially trained to expect a rate of reward (and rate of find) of 1 find per 7 holes, and 1 reward per repetition. This is far from real-world detection rates and the transition could lead to decreases in the dog’s performance, frustration on the dog’s part, false positives (the dog alerting to a non-target sample), or any number of other issues.

In the hypothetical ideal of having dogs screen large groups of people, then using RT-PCR or other tests to confirm the dog’s findings, it may be better to have dogs that are less specific and more sensitive. While it’s highly disruptive for an individual to be falsely identified as infected by a canine, it is arguably worse for a canine to miss a carrier and then precipitate a superspreader event. The oft-stated application for COVID-sniffing dogs is to screen large groups of people quickly and then for those who are identified to be further screened; it therefore would likely be preferable to have dogs with extremely high sensitivity and slightly lower specificity.

Considerations like this are not new to working detection dog programs. Similar considerations are used when training and deploying explosive detection dogs (where a miss can be truly catastrophic) versus narcotic detection dogs (where false positives can unravel a case in court and it’s better to have extremely high specificity) (based on discussions from professionals on the K9s Talking Scents Podcast).

Work by Simon Gadbois and Catherine Reeve has suggested that signal detection theory can be used to identify a dog’s particular tendencies and make adjustments in training for research goals.9 The Gadbois paper also suggests that, based on past psychophysic data, that the alternative choice paradigms used in studies like the Jendrny paper may misrepresent the accuracy of canines. The Gadbois team suggests instead using a yes/no or go/no go setup where the dog is presented with a single sample at a time and the dog must make a decision on that sample, rather than “shopping around” to decide out of a menu of options.

Can the detection dog alert on COVID-19 positive persons by sniffing axillary sweat samples? A proof-of-concept study

By Grandjean et al, 2020, in PLOS One7

This study works with six experienced detection dogs (three explosives detection dogs, one search and rescue dog, and a two colon cancer detection dogs) spread between Beruit, Lebanon and Paris, France. The training originally started with 14 dogs, but eight were not considered ready at the time of testing and testing “could not have been postponed.” The dogs were trained and tested using 177 samples – 95 symptomatic positive cases and 82 asymptomatic negative cases. No asymptomatic positive cases were used. COVID-19 cases that had received long-term medical treatment for more than 36 hours prior to the sample were excluded to avoid VOCs associated with long-term medical treatment. The samples were collected from the underarm region and confirmed in positive or negative status using the same RT-PCR test used in the Jendrny paper. Sample collection was identical between positive and negative cases – the paper even mentions using the same brand of gloves between samples. Negative cases were recruited from the same hospitals as the positive cases. Samples were neutralized before being presented to the dogs. The average age of COVID-19 positive cases was higher than the average age for negative cases for the France location (70 versus 42 years old), but was similar in the Beirut location (48 years for positive and 42 for negative).

The dogs were trained to sit (alert) at a positive sample in a lineup. The researchers used a systematic approach that was not described thoroughly in Jendrny et al. The dogs were initially trained to use the lineup by being rewarded for sniffing the cones (1), then being rewarded for sniffing a COVID-19 positive sample when all other cones were empty (2), then being rewarded for the positive sample when all other cones held testing material without samples (3), then finally for the positive sample in relation to negative samples. Total training for dogs took one to three weeks, though each dog moved from step 1 to 2 in the first day of training.

For testing the dogs were presented with 3-4 cones. A data collector placed the samples and recorded responses. This lower-technology approach compared to Jendrny resulted in only single-blind results, although the data recorder was out of sight for the dog and handler. For each test run, the dog was presented with one positive sample, at least one negative, and between zero and two mock samples without any sweat at all. This approach varied based on how the dog normally worked in its “regular job.” All samples in a given test were from the same hospital and used the same sample collection materials. The dog and handler left the room after each test for the data recorder to re-set. If the dog alerted to a given cone, the handler then instructed the dog to search the remainder of the cones before the test was considered completed. This is termed a simultaneous lineup rather than sequential.

This study did not calculate specificity or sensitivity due to the line-up being simultaneous rather than sequential. Instead, the researchers looked at the random choice proportion for each test and removed mock samples for calculations. So if a 4-cone lineup included 1 mock, the dog’s success rate was 4-1=3, 1/3=33% chance of getting the sample randomly correct. Because each dog was presented the same sample more than once, the researchers marked samples that were presented repeatedly for separate reporting. The dogs success ranged from 76% to 100%, depending on the dog. The four Paris dogs had varying success based on whether the samples were presented the first, second, or third time – two dogs improved over time and two worsened over time. These data were not available for the Beruit dogs.

Two samples originally collected as negative cases were repeatedly marked as positive by the dogs. This suggests that dogs may be useful in detecting pre-symptomatic cases. Upon further PCR testing, the samples were marked as positive. For several incorrect tests, video shows distractions such as a horse walking by or excessive noise adjacent to the testing area. However, other tests did not show evidence of obvious distraction.


This paper lacked sensitivity or specificity data. The use of mocks in a line-up may also make the test easier for a dog, and the lack of consistency in testing protocols with the dogs makes analysis difficult. Samples were used repeatedly and were not directly comparable across positive and negative cases due to discrepancies in age and lack of data on medical history for the cases. While this is a useful validation that dogs may be able to detect COVID-19 based on underarm sweat samples, the study overall leaves a lot to be desired.

This study has the same concerns as the Jendrny study: difficulties in transitioning from lab to real-world and a lack of completely blank rounds. Because no sensitivity or specificity data were collected, the concerns surrounding signal detection theory are even higher with this study. It is a useful first step, but it is a far cry from being ready for use. It’s interesting to note that Beirut reports currently training 20 dogs for use at the airport, although no further information is available on how they plan to take this laboratory study to an operational level.10

General discussion and conclusions

Overall, I’m impressed and excited by how much process has been made with COVID sniffer dogs in such a short amount of time. The initial studies seem very promising, but have a long way to go before the published data supports using COVID sniffer dogs to check large crowds. I am eagerly awaiting data from the airports already using dogs. It is such a huge leap from checking lineups of a few samples (where one is always positive) to searching hundreds of people per day. The studies above are interesting, but they ought to be treated as preliminary. There’s plenty of room for progress, and I’m excited to see that happen.


  1. Else, H. (11/23/2020). Can dogs smell COVID? Here’s what the science says. Retrieved January 11, 2021.
  2. Rooney, N. J., Morant, S., & Guest, C. (2013). Investigation into the value of trained glycaemia alert Dogs to clients with Type I diabetes. PLoS ONE,8:8.
  3. Cornu, J., et al (2011). Olfactory detection of prostate cancer by dogs sniffing urine: A Step Forward in Early Diagnosis. European Urology,59:2, 197-201.
  4. Guest, C. et al (2019). Trained dogs identify people with malaria parasites by their odour. The Lancet Infectious Diseases,19:6, 578-580.
  5. Bomers, M. K., et al (2014). A detection dog to identify patients with Clostridium difficile infection during a hospital outbreak. Journal of Infection,69(5), 456-461.
  6. Amann, A. et al (2014). The human volatilome: Volatile organic compounds (VOCs) in exhaled breath, skin emanations, urine, feces and saliva. Journal of Breath Research,8:3, 034001.
  7. Grandjean, D. et al (2020). Can the detection dog alert on COVID-19 positive persons by sniffing axillary sweat samples? A proof-of-concept study. Plos One,15(12).
  8. Jendrny, P. et al (2020). Scent dog identification of samples from COVID-19 patients – a pilot study. BMC Infectious Diseases,20:1.
  9. Gadbois, S., & Reeve, C. (2016). The semiotic canine: Scent processing dogs as research assistants in biomedical and environmental research. Dog Behavior,2:3, 26-32.
  10. Hu, J. (08/06/2020). Bad News About Those COVID-Sniffing Dogs. Retrieved January 11, 2021