Canine Medical Intuitives
A Primer on Dognosis
Timeline
The first published account of canine medical intuition dates back to 1989, where two London doctors Hywel Williams1 and Andres Pembroke relate the curious anecdote of a woman learning about her malignant tumour from the incessant pawing of her dog. From their <300 word piece titled ‘Sniffer dogs in the melanoma clinic’ published in the Lancet -
The patient first became aware of the lesion because her dog (a cross between a border collie and a doberman) would constantly sniff at it. The dog (a bitch) showed no interest in the other moles on the patient’s body but frequently spent several minutes a day sniffing intently at the lesion, even through the patient’s trousers. As a consequence the patient became increasingly suspicious. This ritual continued for several months and culminated in the dog trying to bite off the lesion when the patient wore shorts. This prompted the patient to seek further medical advice. This dog may have saved her owner’s life by prompting her to seek treatment when the lesion was still at a thin and curable stage.
The authors trot out a hypothesis that malignant tumours and their “aberrant protein synthesis emit unique odours which, though undetectable to man, are easily detected by dogs”. In other words, dogs can sniff out the stench of cancer.
Four years later, the letter happened upon a dermatologist in Florida by the name of Armand Cognetta, who had been rueing that 1 in 5 melanomas went undetected. He decided to set out to work with veteran K-9 trainer Duane Pickel and his schnauzer George to see if a dog could be trained to detect melanomas. George’s training began by searching for test tubes with melanoma samples hidden around the house, progressed to identifying the right tube amongst 10 holes in a cardboard box, and ultimately to samples hidden under bandages on a willing nurse’s body. George’s capabilities were then ad-hoc tested with a grand total of 7 patients who were suspected, but not confirmed, to have melanoma and identified cancer in four, possibly five, depending on how you go about the tricky business of evaluating what is the ground truth. Hywel and John Church2 relayed George’s exploits and another anecdote of melanoma sniffing in yet another letter to the Lancet ‘Another sniffer dog for the clinic?’ in 2001.
It was finally in 2004, 15 years after the first Lancet letter and 18 years ago, that disease detection dogs officially marked itself as a scientific field of inquiry and the ‘dog scan’ was put under the experimental scanner. A team of UK doctors investigated the bladder cancer sniffing abilities of 6 dogs and how well they did compared to chance in identifying which of 7 samples contained a cancer sample (expected value of 1/7 - 14%). After 7 months of training, the dogs achieved a mean performance of 41%, selecting the right sample on 22/54 occasions. While this wasn’t the exemplary performance one would hope for (and which would soon follow), it was striking evidence that something was indeed apaw.
Pounce of Nuance
A quick note on terminology. Canine medical intuition comes in two breeds - biomedical detection dogs and medical assistance dogs. Biomedical dogs detect the presence of disease by either directly smelling the presence of an infective agent (virus/bacteria/helminth) or more commonly, by identifying changes in the volatilome (more on that later). Medical assistance dogs learn to identify changes in either smell, behaviour, or other mysterious processes, in individuals with diabetes or epilepsy (and other rarer conditions), and warn their human guardians of an imminent hypoglycemic episode/epileptic seizure.
State-of-the-bark
Since then, canine diagnosis has been tested in at least 16 different types of diseases in humans, animals, and even trees. A handy table from a paper published in 2021 summarizes twenty papers that investigated dognosis and included sensitivity and specificity (frustratingly the table displays no discernable order) -
Dognosis saw a meteoric rise with the arrival of the pandemic times. The same paper summarizes the twenty-two papers that used double blind randomised controlled trials to assess the efficacy of dognosis for covid19 -
Something the tables don’t include that is worth pointing out is that the vast majority of studies (likely all though I didn’t check every one) had very few dogs involved - in the single-digits.
The proliferation of scientific studies has occurred alongside the rise of commercial and non-profit organizations deploying dognosis in the real world. Medical Detection Dogs in the UK is the top dog with £4M+ in funds that has ongoing studies in Covid19, prostate cancer, Parkinson’s and malaria, among others, as well as training diabetes and epilepsy alert dogs. Cancer Dogs in Canada are testing firefighters for lung cancer in over a 100 departments in the United States. Dognosis was deployed in countries other than the 13 mentioned in the Covid table, including by the Indian army, Thailand, and Rwanda. Unsurprisingly, numerous reports wagged their way into the press over the past few years - including pieces in the New York Times, WIRED, Nature and The Scientist.
Dognosis By Dog Noses
The hypothesis offered by Williams and Pembroke in 1989 that dogs are smelling out disease has since been formalized. In a way, the stench of disease is by no means a modern theory, dating back to Hippocrates in 400 BC who used the art of smell to diagnose ailments. Indeed, the miasma theory of disease, that predominated before germ theory took over, went too far and reversed the causality in claiming odours caused diseases.
Our current understanding goes something like this. Every organism undergoes cellular metabolic processes that produce molecular by-products. Some of these by-products are low molecular weight compounds that exit the body via the air, skin, blood, or liquid/solid waste products. These volatile organic compounds (VOC) are collectively called the volatilome and changes in a disease-contingent manner. With their remarkable olfactory acuity, dogs are capable of detecting these shifts and can be trained to identify them across individuals.
In the studies reviewed above, disease detection dogs generally underwent a training and a test phase that involved positive reinforcement and clicker shaping to train dogs to learn to identify target volatilomes and indicate their presence to handlers using behavioral cues such as staring, sitting, or pressing their snout against the odor container. More recent studies have moved far beyond hiding cancer test tubes around the house and use training aids such as the German ‘Detection Dog Training System’ (DDTS) by Kynoscience that automates the training process for around $17,000.
Nosey Machines
The success of dognosis of cancer prompted researchers to look for uncanine ways to detect cancer VOCs, with a 2010 Nature paper finding some success using nanosensors. More recently, an ongoing collaborative effort between MIT and Medical Detection Dogs and others, are working to develop an electronic nose (eNose) that rivals the capabilities of dognosis. In a paper published in September 2021, dognosis was leveraged to train an Artificial Neural Net on data from a gas chromatography-mass spectroscopy and microbiota analysis of urine samples from prostate cancer individuals. The results were interesting and worth diving into in a whole other piece but the punchline was -
it appears that no unique (set of) molecule(s) can be found as unambiguous biomarkers and instead it is the scent character, as opposed to the molecular composition of scent that communicates the cancer signature to canines. GC-MS produces a list of VOCs that carry the information for each experimental case but such lists themselves unfortunately do not generalize to provide a “master biomarker template” in the form of molecular species identifications and concentrations
Parting Barks
Medicine is essentially two things - diagnosis (prediction) and treatment (causal inference). 3
To me, one of the most fascinating aspects of canine medical intuition is the fundamental questions it raises about the nature of medical diagnoses.4
This is best illustrated by the the UAE study (n=3290) that performed a Bayesian latent analysis model to compare the efficacy of dognosis vs RT-PCR. They found that dognosis was actually more sensitive than the ostensible gold-standard RT-PCR, which was being used in all the other studies and real-world settings to validate the ‘ground truth’ on which to gauge the accuracy of dognosis.
Difficulties in standardization and regulation, cost of training, lack of anonymization and the infrastructure needed to care for canines are commonly cited reasons for why dognosis is impractical on a mass-scale. A big part of the challenge is the black-box nature of dognosis that provokes incredulity and the ‘can-we-really-trust-a-dog-to-tell-someone-they-have-cancer’ disbelief that forestalls wide-spread adoption.
The current popular approach appears to avoid depending directly on the carbon-based black-box dognosis and adopt silicon-based ‘we-totally-understanding-how-large-neural-nets-work’ eNoses instead. One can potentially imagine another approach that synergistically combines carbon and silicon in a manner that augments both. It remains to be actualized what this would look like. Until then, we’re stuck with good ol’ unfurry diagnoses.
Hyles just happens to be “the most cited dermatology researcher in the world on Google Scholar (137,000 citations September 2021) and H-index of 142 and i10-index of 717”.
Church was apparently known for being “involved with developing the use of maggots to clean infected wounds” before deciding to go after furrier pastures.
Taken from the UC Berkeley class PB196 : AI and Public Health taught by the brilliant Ziad Obermeyer.
I bet you can’t help but notice the noses in diagnoses now ;)