Can Medical Equipment Think? What AI in Medical Devices Really Does
More than 1,200 AI-enabled devices are cleared in the US, most of them reading scans. The evidence is strong for narrow tasks and much weaker for the bigger claims. Here is how to tell the difference.
Summary
- The FDA's list of AI-enabled devices passed about 1,200 authorisations by mid-2025, with more than 200 added a year, and most of them read medical images.
- The evidence is strongest for narrow tasks: flagging findings on scans, reading ECGs, speeding up reports and predicting equipment faults. Claims for AI-assisted surgery and broad decision support rest on much weaker studies.
- The real risks are bias, thin disclosure and clinicians over-trusting the machine. A hospital buying AI should ask for validation on patients like its own; an investor should ask who pays for the output.
A radiologist may read thousands of chest X-rays a week. A cardiology team watches continuous ECG streams from dozens of patients. A hospital's biomedical engineer tries to guess when a scanner costing crores will fail. All three jobs demand sustained attention to subtle patterns in large volumes of data, and all three are where AI inside medical equipment is doing real work today.
Whether that counts as the equipment thinking is a matter of definition. It does not reason or exercise judgment. It does recognise specific patterns, in specific data, at a speed and consistency no team can match, and in some tasks as accurately as experts. The useful question is which tasks, on what evidence.
How much AI is already inside devices
The US Food and Drug Administration keeps a public list of AI-enabled devices it has authorised. It passed about 1,200 entries by mid-2025, after several years in which more than 200 were added annually. Radiology dominates the list by a wide margin, followed at a distance by cardiology and neurology.
The large majority reach the market through the 510(k) route, by showing they are substantially equivalent to a device already cleared, rather than through the slower De Novo route for genuinely new devices. That keeps approvals quick. It also means many AI products have been cleared on comparison with a predecessor rather than on fresh clinical trials, a point worth remembering when reading a vendor's accuracy claims.
What AI in medical equipment actually covers
- Imaging and detection. Software that reads X-rays, CT, MRI and ultrasound to flag abnormalities, measure lesions and push urgent cases such as a brain bleed or pulmonary embolism to the top of the worklist.
- Cardiac monitoring. Algorithms that read ECGs from hospital monitors, portable devices and watches. AliveCor's Kardia 12L, cleared in 2024, reads a portable 12-lead ECG for 35 cardiac conditions including heart attack.
- Clinical decision support. Tools inside bedside monitors and records that combine vital signs, labs and medications to warn of deterioration or guide dosing, especially in intensive care.
- Equipment health. Systems that watch the machine itself and predict component failure before it causes downtime.
- Surgical assistance. Guidance and control features in surgical robots and navigation systems.
Where the evidence is good
Imaging. A study of 6,716 patients from the US National Lung Screening Trial, reviewed in iRadiology in 2025, reported an AI model predicting lung cancer risk with sensitivity of 94.4% and specificity of 93.8%, and cutting false positives against six radiologists. Meta-analyses of fracture detection on X-rays published between 2020 and 2025 report pooled sensitivity and specificity above 90%, with AI assistance lifting radiologists' own sensitivity by several points.
Workflow. The most convincing real-world result may be the least dramatic. Northwestern Medicine ran a generative AI reporting tool across 11 hospitals and about 24,000 radiology reports in 2024 and reported an average 15.5% gain in report-completion efficiency, with no loss of accuracy.
Uptime. Predictive maintenance is where AI pays most directly. Siemens Healthineers says its Guardian programme monitors more than 80 critical components of its Atellica lab platform and schedules service before failure. A four-hospital case study in Malaysia reported MRI downtime falling about 20% with condition monitoring. For a hospital, an idle MRI is lost revenue by the hour; our piece on recurring revenue in medical equipment explains why manufacturers are selling this as a service.
Time saved per report is money a hospital can measure; a claimed 40% gain in surgical precision is not.
Where the claims run ahead
Reviews of AI-assisted robotic surgery circulate impressive figures: a quarter less operating time, 30% fewer complications, 40% better precision. They come from a narrative synthesis of 25 studies with different designs, procedures and definitions, not from randomised trials. Treat them as a signal that the area is worth watching, not as numbers to put in a business case.
The same caution applies to market forecasts. Published projections for the AI medical device market vary several-fold between research firms, which tells you more about definitions than about demand.
The real risks
- Bias. A 2024 scoping review of 692 FDA-authorised AI devices found only 3.6% reported the race or ethnicity of their validation patients, and very few linked to peer-reviewed performance studies. A tool validated on one population may perform worse on another.
- Thin disclosure. Many authorisations say little about training data, algorithm type or both sensitivity and specificity, which makes independent assessment hard.
- Over-trust. Research in colonoscopy found clinicians' own detection rates fell after they became used to AI assistance. A tool that makes clinicians less attentive can do net harm.
- Changing software. The FDA's December 2024 guidance on predetermined change control plans lets makers pre-agree how an algorithm may be updated after clearance. In Europe, the AI Act sits on top of the device regulations, adding a second compliance layer for data governance and fairness.
What this means in India
Indian hospitals are buying AI through their medical equipment: new scanners, monitors and lab analysers increasingly arrive with it built in, often as a separately priced licence. The problem of validation is sharper here, because many tools were trained and tested on American or European patients.
- For hospital buyers: ask for performance data on Indian or comparable populations, a local pilot with your own cases before signing, the licence cost after year one, and who is accountable when the software is wrong.
- For founders building AI devices: accuracy on a benchmark does not sell. A measured saving does: reporting time, avoided downtime, fewer missed findings. Plan regulatory clearance with CDSCO, and with the FDA or EU if you intend to export, from the start.
- For investors: check whether the product has a buyer who pays for its output, whether its validation holds outside the original dataset, and whether the clearance route leaves room for updates. Those questions belong in any due diligence on a healthtech company.
The same data question runs through continuous monitoring, where wearables produce far more signal than clinicians can read; we cover that in the move from checkups to continuous care. Can medical equipment think? No. But it can already do several narrow jobs very well, and the hospitals and companies that match the tool to the task will get the value.
How useful was this article?
One tap. It tells us what to write more of.
About the author
SRF Capital Studio
The next one
Get what we publish next, by email.
Working notes on raising, borrowing, protecting, growing and structuring capital in India. One email a week at most, and you can leave any time.
We use your address only to send this. See our privacy policy.
We store your address to send you these emails and nothing else. See our privacy policy.
Related reading
