How Dogs and AI Are Learning to Sniff Out Cancer From Breath

In a lab near Bangalore a beagle stopped mid-sniff, stepped back, circled once and then returned to the vial. The machine beside him recorded a tiny spike in breathing and a flicker of movement. To the human eye it was a moment; to the sensors it was a data point. To the team at Dognosis it was a sign.

Dognosis, an Indian startup, is marrying the extraordinary olfactory powers of trained dogs with machine learning to create a noninvasive early-screening tool for cancer. The idea is simple in concept and complex in execution: people breathe into a cotton mask for about ten minutes, the mask is sealed and sent to a lab, and trained dogs sniff the samples while sensors capture their respiration, brain activity, posture and micro-movements. Algorithms then comb those signals for patterns that might point to cancer-related chemical signatures.

So far the results are compelling enough to draw attention. In a multi-center effort involving more than 1,500 participants across six hospitals, Dognosis reports accuracy above 90% across seven cancer groups. In a phase-two trial published in Clinical Oncology, the company described roughly 90% sensitivity for detecting early-stage cancers across seven broad categories encompassing more than 20 cancer types.

These numbers do not mean the method is ready to replace biopsies or established diagnostic tests. Oncology experts caution that promising findings must be replicated in larger, independent cohorts before the approach can be deemed reliable. Still, the hybrid model—canine scent detection augmented by AI—offers advantages: the sampling is noninvasive, patients never touch the dogs, and the dogs work from sealed samples so biosafety risks remain low.

The dogs live and train at Dognosis’s farm and laboratory near Bangalore. Their days include 30 to 60 minutes on a sniffing platform, where training is gamified and treats are the usual reward. The group includes beagles, Labradors, German Shepherds and other breeds chosen for their scenting abilities and trainability. One cofounder, Akash Kolgode, says the animals sometimes surprise researchers: dogs trained on odors from ten cancer types have, in some instances, signaled correctly when presented with an eleventh type they had never been taught to detect.

That anecdote raises both excitement and questions. Can this performance hold across India's enormous genetic, dietary and environmental diversity? Can labs scale the logistics of collecting, sealing and transporting tens of thousands of breath samples while keeping sensor data consistent? And will clinicians and patients accept a screening pathway that relies on animals as the frontline detectors?

The work is being pitched against a pressing need. According to the International Agency for Research on Cancer, India saw roughly 1.5 million new cancer cases and more than 900,000 deaths in 2024. Conventional population-wide cancer screening remains limited in many regions, so a low-cost, noninvasive triage tool could change how early detection is pursued—if it proves robust in bigger, independent trials.

The path forward is clear in one sense: larger, multi-site studies that include diverse populations, blind validation and open datasets for independent analysis. There are practical hurdles too—standardizing sample collection, ensuring sensor and algorithm reliability, and building trust among physicians and patients. But the image of a dog pausing over a tiny vial, its subtle reactions translated by sensors and then read by an algorithm, offers a new kind of partnership between biology and computation. Will that partnership reshape early cancer detection? Time and rigorous science will tell.