Dr Kola Tytler (MBBS MBA MRCGP) | 15 July 2026 | 11 min read
As preventive health platforms collect more physiological and imaging data per person, the central challenge in the category is shifting from a technical one, can the system detect an abnormality, towards a genuinely clinical one, does that abnormality actually matter for this person, and what should happen next. Neko Health's own published outcome data is a useful, and unusually transparent, case study for exploring where that line sits.
The promise of detection at scale
Modern AI systems can process millions of individual measurements per assessment, identify abnormalities considerably faster than a manual review workflow, and compare an individual's results across repeated visits over time, potentially flagging combinations of small changes that would be difficult for a human reviewer to notice working through the data manually. That capability is genuinely new, and it is reasonable to expect it to catch some things that conventional, less frequent screening would miss.
The problem detection alone does not solve
Not every technologically detectable abnormality is clinically important, and a system built to be highly sensitive will, by construction, surface more incidental findings and borderline results than one calibrated more conservatively. This matters more in an asymptomatic, low pre-test-probability population, which is precisely the population most preventive AI platforms are screening, than it would in a symptomatic clinical population where the prior probability of significant disease is already higher. A finding that would be unremarkable, or simply monitored, in a patient presenting with relevant symptoms can generate disproportionate anxiety and unnecessary downstream investigation when it appears unprompted in an asymptomatic person who came in for a routine preventive check.
What Neko's own published data actually shows
Neko has published outcome data from its clinics with unusual transparency for the category, and it is worth reading in full rather than only as a headline statistic. In its year-two data story, covering 4,362 members scanned in Stockholm, 81.3 percent were in good health and required no further follow-up. Of the remainder, 18.7 percent were referred for further assessment; within that group, 1.2 percent were found to have a previously unknown, life-threatening condition requiring immediate care, and a further 6.4 percent had medically significant findings requiring treatment but not immediate emergency intervention. Roughly 90 percent of those found to have a significant or life-threatening condition reported no prior awareness of it or symptoms before their visit. Neko itself is explicit that this is not a controlled study, and that these figures describe outcomes within its own self-selected, largely self-funded member base rather than a randomised or population-representative sample.
Why these figures are encouraging but not sufficient on their own
Detection rates of this kind are genuinely useful signals, and a company publishing its own outcome data at all, rather than only marketing anecdotes, deserves some credit relative to much of the sector. But detection rate alone does not establish the things that actually determine whether a screening programme is net beneficial: whether earlier detection in this population changes long-term outcomes such as mortality or serious morbidity, what the true rate of false positives and overdiagnosis is once longer follow-up data accumulates, whether the anxiety and downstream investigation generated by borderline findings is proportionate to the benefit gained, and whether the programme is cost-effective at a health-system level rather than simply commercially viable at an individual-consumer level.
What established screening principles say about this trade-off
The UK National Screening Committee's long-standing position on any screening programme is that false positives, overdiagnosis, patient anxiety and unnecessary follow-up investigation are inherent risks that have to be weighed explicitly against the benefit of earlier detection, not treated as an acceptable cost of doing more testing. The American College of Radiology has previously stated, in the specific context of total-body MRI screening in asymptomatic individuals, that the evidence has not yet established that such screening prolongs life or is cost-effective at scale. Neko is not a whole-body MRI service and the specific evidence base is not identical, but the general principle, that broad, low pre-test-probability screening requires its own dedicated evidence of net benefit rather than borrowing credibility from the fact that individual tests within it are individually validated, applies just as directly.
Why this is where a clinical evidence and interpretation layer becomes essential
Detection AI, however good, needs a layer that can answer a different set of questions for both the clinician and the patient in front of a result: what does this specific finding actually mean in context, does it meaningfully change this individual's risk, which current guideline applies to it, what is the appropriate and proportionate follow-up, and how urgently does the patient actually need to be seen. That is a distinct function from generating the finding in the first place, and it is where clinical decision-support and evidence-retrieval platforms become complementary to diagnostic AI rather than competitive with it. A finding without that layer risks becoming data without a decision attached to it.
Conclusion
The genuinely hard problem in this category was never really detection. Sensors and AI models are already good, and improving quickly, at finding things. The harder and more clinically consequential problem is deciding, case by case, what a detected finding actually means for the person in front of you, and that is a problem measurement alone does not solve.