What are the most common uses of AI in Healthcare? I have seen a lot of papers and stories about systems that do impressive-looking things, but which are not widely used. What is actually being used?
Usage data
I have not seen a single paper which gives comprehensive data on usage of AI in different health contexts, but I have seen work on different aspects of this, some of which I describe below. I welcome additions to this list!
AI Medical Devices: Colangelo gives data on which medical devices have been approved by FDA in US. This is not usage, but it is closer to real-world than one-off demos. Anyways, Colangelo points out that 76% of approved devices are for radiology, which is clearly a “hot” area for medical AI. She also discusses products from GE Healthcare (which is a leader in this space in US); these help operators correctly position patients, enhance image quality and reduce noise, provide training and support to operators, and support workflows. It is noticable that all of the products she highlights are intended to support radiologists in useful ways that do *not* attempt to directly diagnose images. This does not surprise me, it matches what radiologists in Aberdeen have told me that they want, and what radiology companies have told me they can sell.
Usage of AI in Scottish NHS: Another data point was a presentation I went to last year about the use of AI in the Scottish NHS (National Health Service). Basically there are AI tools used in specific hospitals, but (at least in 2024) nothing that is used in hospitals across the Scottish NHS. As far as I can tell, the most widely used medical AI system in Scotland is Heidihealth, which is a medical scribing tool that is primarily aimed at GPs (not hospitals).
Popular consumer-facing health apps: I looked at some lists of most popular health apps in UK, and they were dominated by fitness apps such as Strava and Fitbit; these may use a bit of AI but it is not their focus. For example, the nutrition app MyFitnessPal has a “Meal Scan” feature which uses AI and computer vision to analyse pictures of meals and determine what food is in the meal. But this feature does not seem to be emphasised by MyFitnessPal, and users have mixed opinions about its usefulness.
ChatGPT to answer patient’s health queries: A lot of people use ChatGPT (and similar tools) for health queries. Ayre et al (2025) report that in Australia, 10% of people surveyed used ChatGPT to obtain healthcare related information in the previous 6 months. An additional 40% of the people surveyed had not yet used ChatGPT for health information, but said they might do this in the future. Anecdotally, I know people who in past used Google search for health information, who now use ChatGPT as well (or even intead of) Google search.
My thoughts
The most common use of AI in healthcare today is probably the general public asking chatGPT (etc) health questions. Of couse LLMs can mistakes, including saying things which are factually incorrect and also things that are inappropriate (eg, presenting marketing material pushing a specific intervention without explicitly saying this is marketing). However they also have the huge advantage (compared to Google search) that they can personalise results to individuals. So this use case is here to stay and will grow, and the challenge is to make LLMs safer and more useful in patient information contexts (which is something we are working on in Aberdeen).
Among clinicians, the most common use of AI is probably writing scribes. Which makes sense, clinicians want help with writing and LLMs are good at providing such help. Indeed, Chatterji et al (2025) say that “Documenting/Recording Information” is the most common use of chatGPT in healthcare settings and that writing tasks are the most common usage in work settings more generally. LLM medical scribes are more effective if UI/UX is tailored for user and task (Knoll et al 2022), so a range of scribes are probably need for different clinical contexts.
There are also many products which use AI to help clinicians address other pain points, such as image quality enhancement, training tools, and supporting workflows. Again this seems very sensible.
What we do not see is widespread use of AI for clinical decision support. AI researchers love this, perhaps because its easy to frame as a classification task (predict diagnosis from symptoms), but what we have seen for *70 years* is that it is difficult to get such systems used, not least because they do not address a major “pain point” for clinicians and the health system more generally.
Final thoughts
I think the market is telling us that the biggest demand for AI in healthcare in 2025 is (A) apps that give consumers and patients useful personalised health information and (B) apps that help clinicians write documents. It is a shame that researchers do not focus more on these areas…
Ehud Reiter's Blog 
From Keith Trnka (he agreed to let me share his very interesting comments)
I led the applied ML team at 98point6 from around 2018-2022. We were a US-based primary care telehealth startup.
The most promising areas we saw:
Smart intake / triage: The hard part of diagnosis is getting all the relevant information from the patient. We had great success in automating it. That would be much easier to do with LLMs than the models we used back then.
Documentation: Writing the clinical notes, filling out the right diagnosis codes, etc. That said, clinicians were very sensitive about the way notes were written because clinical notes could be used as evidence in malpractice lawsuits. Sadly it wasn’t as simple as providing continuity across visits.
Forecasting demand and scheduling clinicians: We had a unique model in which patients could show up any time without scheduling an appointment, which meant we needed reasonable forecasts. There are many challenging and interesting goals like setting the clinic schedule far enough ahead, providing consistency, reacting to sick time or mistakes, fairness, state-specific licensing, etc.
When I’ve talked to friends working in healthcare tech in 2025, OpenEvidence comes up often as a major use of AI for clinicians. AI scribes are common too.
CDS is one of those areas where I completely changed my perspective after working in industry. When I first started out, I thought CDS would be a great application of AI/ML. Diagnosis prediction is a very natural-looking ML problem and is frequently used as an example of ML in textbooks and literature.
Once I was working directly with clinicians, it was apparent that diagnosis was typically quick and easy once they had all the relevant information. (Though with the caveat that diagnosis itself was quick and finding the proper ICD-10 code was not always quick)
Beyond that, if the goal is to improve patient outcomes there are typically much more impactful opportunities, like broadening the information in the medical interview or making extra sure that we know the patient’s active medications. Last time I checked, unreliable medication history was still one of the biggest factors in adverse events.
LikeLike