AINEWS
Japan's Nodoca camera screens throats for flu and Covid in 10 seconds, and 2,000 clinics use it with national insurance coverage. Three years of consented training data is what made the model work.
On a remote Japanese island, the only medical equipment Sho Okuyama could reliably use was a stethoscope. A decade later, his answer to that constraint is a clinic camera that screens throats for flu and Covid in roughly 10 seconds, replacing the nasal swab at more than 2,000 Japanese medical institutions.
The device is called Nodoca, made by Iris, the company Okuyama founded in 2017. A patient opens wide; a compact camera at the end of a wand captures pharyngeal images (pictures of the back of the throat). Nodoca combines those images with answers from a short medical interview and returns an influenza assessment in a little over 10 seconds.
The mechanics are simple from the patient's chair. From the clinician's, the system is a long-overdue replacement for a diagnostic step that has barely changed in a century. The nasal swab works, but it hurts, and the discomfort pushes people to delay testing until symptoms are advanced. A throat photograph works earlier in the course of an illness, which is when treatment is most useful.
Japan's regulators agreed. In 2022, Nodoca became the first AI-equipped medical device in Japan to be approved as a "new medical device", a regulatory category for novel medical hardware that is separate from software-only approvals, and to win national health insurance coverage on the same pathway. The device is now in use at more than 2,000 institutions. In October 2025, Japan's regulators cleared an additional Covid-19 diagnostic function on the same hardware.
The interesting question is not whether the swab is annoying. It is why this device works at all, given that dozens of medical AI projects do not.
Okuyama's answer puts the bottleneck in an unfamiliar place. The conventional story about clinical AI is a model story: a foundation model, a benchmark, a fine-tune. Okuyama frames his company's differentiation around the sensing layer, not the diagnostic model. As he told Wired, "the real differentiating factor lies in sensing — how the data is acquired."
Three engineering problems had to be solved before that quote could be true. First, the hardware: a phone-grade camera does not see the back of a throat consistently across clinics and patients. Second, the training data: a useful pharyngeal image dataset did not exist, so Iris lent dedicated cameras to around 100 medical institutions and collected pharyngeal images with patient consent over roughly three years. Third, the diagnosis itself: even with the right images, the model had to learn to map visual features of inflamed tissue to a specific viral infection rather than a generic "looks sick" reading.
The hardware problem is unglamorous and underwritten by AI venture capital. It is also the reason the company's pitch is closer to a sensor company than a model company. Iris owns the camera, the imaging protocol, and the consented data pipeline. Anyone can run a vision model; few can build the dataset that lets the model be right at the point of care.
That asymmetry maps onto a structural claim about clinical AI more broadly. The model layer is increasingly commoditized. The data-acquisition moat, the durable advantage that comes from controlling how real-world clinical images are captured, consented, and labeled at scale, is not. Nodoca's three-year buildout of consented pharyngeal images across a hundred loaned cameras is the kind of asset a foundation-model lab cannot buy off the shelf.
The limits are real. The 2,000-institution figure and the 10-second assessment are company-reported, not independently verified. Nodoca's regulatory clearance and reimbursement are Japan-specific; U.S. FDA and EU CE pathways are not addressed in the source. Clinical sensitivity and specificity, the rate at which the device correctly flags people who do have flu and correctly clears people who do not, are not in the public reporting. The "sensing > diagnosis" framing is Okuyama's argument, even if the three engineering problems and the rollout make it a structural claim rather than a slogan.
What to watch: whether Iris extends the same camera-and-data pipeline to other pharyngeal conditions such as strep, RSV, and mononucleosis, and whether any non-Japanese regulator clears the device. If the sensing-layer thesis holds, the next decade of clinical AI will look less like a model race and more like a slow, camera-laden buildout. It will happen one clinic at a time, the way Okuyama started.