A team of dogs smelling the breath of 1502 participants across six Karnataka hospitals, combined with AI, detected cancer with 90+% accuracy. AI can see, hear and reason, but it cannot smell. So here the dog does the detecting, and the AI reads the dog.
Our guests, Itamar Bitan and Akash Kulgod, are the co-founders of Dognosis. Itamar leads product, and has spent a decade reading detection dogs, and together they are working to turn dogs’ ability to detect disease-associated scents into a repeatable system. Their explanation starts with a useful distinction: the dog senses something; the machine helps make sense of the dog’s response.
We ask how they train dogs and models, combine several dogs’ judgments using Bayes, and cope with a colleague who might have a bad day. The conversation moves through clinical research, the practical advantages of building in India, and the hardware needed to increase throughput. It ends with a question that stays with us: if a future AI saw humans as capable but difficult to understand, what would we want it to learn from the way we work with dogs? Along the way: a 45-minute working day, a sniff gym in every neighbourhood, and why there is no Reddit of smells.
Highlights
Who is detecting what? Akash separates the dog’s sensing from the interpretation of its response, and explains the role human trainers still play.
Learning the individual dog: Itamar describes why body language, temperament and context matter—and why dogs need not all signal in the same way. One scratches the sample, one freezes, one runs to the feeder.
Two kinds of training: Training the dog and training a model to understand the dog are connected tasks, with their own tests and feedback.
Bayes, with a pack of dogs: How the team combines a person’s risk factors with several dogs’ indications, weighted by each dog’s past performance. They call it wisdom of the pack: one dog can have a bad day and the result holds.
A good result is the beginning: What it takes to move from a study result to reliable performance across dogs, people and time.
Giving machines a sense of smell: From a simulated fruit fly to ARIA’s olfaction programme, we explore the data, sensors and representations this field needs. As Akash puts it, there is no Reddit of smells to download. But the dogs can smell cancer now.
A score for a clinician: The guests explain how they think about a triage tool and the medical judgment that follows it. The cautionary tale is Covid: sniffer dogs were trained in 60 countries, and deployment stalled because nobody knew whose job it was to approve them.
Why build in India? Akash and Itamar discuss assembling a team across clinical research, hardware, software, behaviour and canine training—and Bengaluru’s daily friction.
Sniffs, samples and scale: The guests work through utilization, automated sample presentation and the difference between possible capacity and actual demand. The sums: 30 dogs, about 45 minutes of sniffing a day each, a million samples a year. After that, a sniff gym where the best pet dogs get “drafted to the professional leagues”.
Dogs as colleagues: Routines, rest and purposeful work lead into a bigger conversation about collaboration between different kinds of intelligence. Akash’s closing thought: dogs may be the only species to have aligned a more intelligent one.
Notable quotes
You can go halfway to the dog’s way of communicating. I don’t force them to communicate in human language.
— Itamar
You don’t have, you know, a Reddit of smells you can just download and train your smell bot on.
— Akash
With a pack of 30 dogs, we can do a million samples a year with each dog having excellent work-life balance of about 45 minutes of sniffing a day.
— Akash
Dogs are one of the few species to be able to claim, maybe the only species, to have been able to align a quote-unquote more intelligent being. Now I don’t know necessarily if a dog like the pug is happy about the alignment. They ended up without noses.
— Akash
Links & resources
The published Phase II study in the Journal of Clinical Oncology, with the abstract available on PubMed.
ARIA’s Olfactory Perception programme, discussed in the conversation.
SmellNet research, for further context on the growing field of machine olfaction and its datasets.
A note on the study: The published case-control study reports 90.8% sensitivity and 91.3% specificity in a test cohort of 1,502 participants across six Karnataka hospitals. It establishes analytical validity for the combined canine-and-Bayesian system; prospective evaluation in true screening populations is the next step.
A note on smell datasets: Existing datasets include SmellNet. The broader challenge discussed here is building shared infrastructure and general-purpose capabilities comparable to those that accelerated computer vision.
