Machine Learning in Healthcare: Applications, Benefits, and What to Build

AI in healthcare is in production across medical imaging, clinical documentation, coding and risk prediction, and still experimental in diagnosis and treatment planning. The clearest measure of maturity is regulatory: the FDA’s AI-Enabled Medical Device List held 1,614 authorised devices as of its September 2026 update, and 1,230 of them, about 76%, are radiology. Adoption is real but narrower than the coverage suggests. Menlo Ventures research put domain-specific AI adoption at 22% of healthcare organisations, a sevenfold rise in a year from a small base. 

Where AI in healthcare actually stands

The gap between what AI in healthcare can do and what it is doing is wide, and most articles on the subject quietly close it.

Here’s the useful way to read the field. The FDA keeps a public list of every AI-enabled medical device it has authorised for marketing in the United States. That list is the closest thing to an honest scoreboard, because getting on it means a device cleared a regulatory bar rather than a press release.

Three things fall out of that table.

Imaging is not one use case among many. It is three quarters of the entire field. Everything else combined accounts for roughly 384 authorised devices across every other specialty. Cardiology is a distant second.

Almost nothing goes through a novel pathway. 96% clear via 510(k), which means demonstrating substantial equivalence to a device already on the market. AI does not automatically trigger a harder regulatory route, which is part of why the numbers climbed so fast.

The growth is genuine. Six devices in 2015, 335 in 2025, 181 in the first half of 2026 alone. That’s not a hype curve flattening out.

On the adoption side, the picture is more modest than vendor marketing implies. Menlo Ventures research put the share of healthcare organisations running domain-specific AI tools at 22%, up roughly sevenfold from the year before. Growing fast, but from a small base, and heavily concentrated in large systems with the integration budget to make it work.

The 18 use cases, sorted by how ready they actually are

Most lists present every use case as equally viable. They are not. These three tiers reflect what’s deployed at scale, what’s moving into production now, and what still needs evidence.

Tier 1: In production, with regulatory clearance

1. Medical imaging analysis

The most mature category by a wide margin, with 1,230 FDA-authorised devices behind it. Models detect lung nodules on CT, flag intracranial haemorrhage on head CT for priority reading, identify fractures on X-ray, screen for diabetic retinopathy from fundus images, and assist mammography interpretation. These integrate into PACS and DICOM workflows and surface findings as prioritised worklists and image overlays, not as standalone screens.

The real benefit is usually triage speed rather than raw accuracy. A stroke flagged in the queue two hours earlier changes the treatment window.

2. Ambient clinical documentation

AI that listens to the consultation and drafts the note. Nuance DAX, Abridge and DeepScribe are the names that come up most. This is the fastest-spreading category in healthcare right now, and the reason is simple: documentation burden is the single largest driver of clinician burnout, and ambient scribes attack it directly without touching clinical decisions.

Mass General Brigham reported a 21.2% drop in physician burnout prevalence after 84 days of ambient AI use. That’s a rare thing in health IT, a measured outcome rather than a projected one.

3. Medical coding assistance

NLP models read clinical documentation and suggest ICD-10 diagnosis codes and CPT procedure codes. Coders validate suggestions instead of coding from scratch. Coding specificity usually improves alongside speed, which matters because under-specific coding is itself a denial cause.

4. Claims denial prediction

Models score claims before submission against diagnosis and procedure combinations, payer rules, documentation completeness and historical denial patterns, then route likely failures for correction. With average denial rates running 15% to 17% for commercial and Medicare Advantage plans in 2026, this is one of the clearest ROI cases in the whole field. The economics of the underlying systems are covered in our breakdown of what billing software costs to build.

5. Patient scheduling optimisation

No-show prediction, intelligent overbooking, matching slot length to visit complexity. Unglamorous, well understood, and it pays back quickly because an empty slot is pure lost revenue.

6. Risk stratification

Scoring patients for hospitalisation, readmission, ED use and cost, using clinical data, claims data, social determinants and behavioural signals. More accurate than the regression models it replaces, and it drives where care management teams spend their time.

Tier 2: Moving into production now

7. Clinical decision support

Models predicting sepsis onset from vital sign trends, flagging deterioration, suggesting medication adjustments against pharmacogenomic profiles. Delivered through CDS Hooks at the point of care. The engineering is solved. The hard part is alert fatigue, which is a design problem rather than a model problem and is covered in our guide to healthcare interface design.

8. Early warning and deterioration prediction

Continuous models watching vitals, labs and remote monitoring data to predict sepsis, respiratory failure or ICU transfer hours before standard assessment would. Strong in research settings. Performance varies considerably once you move to a different patient population, which is the whole argument for local validation.

9. Digital pathology

AI reading digitised slides to detect cancer cells, grade tumours and measure biomarker expression. Held back less by the models than by the scanning infrastructure, since most labs still work with glass.

10. Prior authorisation automation

Predicting which services need authorisation, auto-populating forms from the record, estimating approval likelihood. Combined with Da Vinci PAS FHIR APIs, this cuts a process that currently wastes enormous amounts of clinical staff time.

11. Drug interaction and adverse event prediction

Models weighing medication lists, genomics, labs and comorbidities to predict interactions with more specificity than rule-based checking. The goal here is suppressing low-risk alerts, not adding more of them.

12. Clinical trial matching

NLP extracting eligibility-relevant attributes from unstructured records and matching against thousands of active trials. Finds matches manual review misses, which matters most in oncology.

13. Symptom triage and virtual assistants 

Chatbots helping patients judge urgency and route to the right level of care. These need careful design to avoid false reassurance, and they sit alongside telemedicine consult flows rather than replacing them.

14. Medication adherence and behavioural nudging

Models reading fill data, device readings, portal engagement and appointment history, then delivering personalised interventions. Works best where adherence directly determines outcomes, which is why it shows up first in pharmacy and chronic care platforms.

15. Remote monitoring analytics

Continuous glucose monitors, ECG-capable wearables and connected devices generate more data than any clinician can read. The AI job is summarising without losing the exception that matters, which is now a core requirement in health tracking platforms.

Tier 3: Real potential, not yet proven in routine care

16. Generative AI for clinical reasoning

Large language models answering clinical questions, summarising patient histories, drafting differential diagnoses. The published benchmark results are genuinely impressive. The deployment record is thin.

There’s a regulatory fact here that explains a lot, and almost nobody writing about healthcare AI mentions it: as of early 2026, no device on the FDA’s list was authorised as a generative-AI or LLM-powered medical device. Everything on that list of 1,614 is conventional machine learning.

That is why ambient scribes spread so quickly while diagnostic LLM tools did not. A scribe that drafts a note a clinician reviews and signs is not making a clinical claim, so it generally falls outside the Software as a Medical Device definition. A model that suggests a diagnosis is making one, and that’s a different regulatory conversation entirely. If you’re planning a product, that line is the single most important thing to understand before you scope it.

17. Drug discovery and repurposing

Predicting molecular interactions and identifying candidates. Reshaping pharmaceutical pipelines, with reported R&D timeline reductions in the 30% to 40% range, but it affects which treatments exist rather than how care gets delivered.

18. Autonomous diagnosis

A small number of devices are authorised for autonomous use, most notably in diabetic retinopathy screening. Beyond narrow, well-bounded screening tasks, autonomous diagnosis is not where the field is, whatever the conference keynotes suggest.

What these systems cost to build

No competitor in this space publishes cost figures. Here are working ranges for healthcare AI features, assuming you already have the underlying platform.

Two patterns hold across all of them.

The model is rarely the expensive part. Data pipelines, EHR integration and validation are. Budget 40% to 60% of an AI feature’s cost for getting data in and results back out of the clinical workflow.

Anything that meets the Software as a Medical Device definition changes category entirely. Clinical validation studies, quality management systems and regulatory submission typically cost more than the software. A tool that supports a clinician’s decision and a tool that makes one have very different price tags for the same model.

Why most healthcare AI projects stall

Four failure modes come up repeatedly, and none of them are modelling problems.

It lives outside the workflow

An AI tool needing a separate login, a separate screen or manual data transfer does not get used, no matter how good it is. Integration through SMART on FHIR, CDS Hooks or an embedded EHR view is the difference between adoption and shelfware.

It wasn’t validated locally

Published performance is not your performance. A sepsis model trained on an academic medical centre’s population behaves differently in a community hospital with different demographics, different documentation habits and different data quality. Validate against your own patients before production.

Bias wasn’t tested for

Healthcare AI has to be tested for performance disparities across race, ethnicity, age, sex and socioeconomic status. Models trained on non-representative data amplify existing disparities, and in healthcare that is a patient safety issue before it is a compliance issue. Test subgroup performance explicitly rather than relying on aggregate accuracy, which hides it.

Nobody owns it after launch

Model performance drifts as populations, coding practices and care patterns change. A model without monitoring and a retraining schedule degrades quietly. Budget for the ongoing work or don’t start.

Regulation, and not only in the US

Most writing on this subject assumes a US reader. Here’s the short version for the markets that matter.

United States

The FDA regulates AI meeting the Software as a Medical Device definition. Most authorisations go through 510(k). ONC requires algorithm transparency in certified health IT. HIPAA governs the data regardless of whether the tool is a regulated device. The current list stands at 1,614 authorised devices.

European Union

AI medical devices fall under the Medical Device Regulation, with the EU AI Act adding a second layer that classes most healthcare AI as high risk and brings its own obligations around risk management, data governance and human oversight.

United Kingdom

The MHRA regulates medical devices post-Brexit, and NICE evaluates clinical and cost effectiveness separately. Clearance and adoption are two different hurdles.

India

CDSCO handles medical device regulation, and software as a medical device is a developing area. ABDM sets the health data exchange standards that any serious platform needs to work with.

UAE

DHA and DOH set their own requirements, and eClaimLink governs the claims side. Assumptions carried over from US projects will be wrong in both directions.

A vendor quoting you an FDA pathway for a Dubai deployment has not read the brief.

How to pick your first use case

A short test that filters most bad ideas quickly.

• Is there labelled historical data already, or does it have to be created? Creating it usually costs more than the model.
• Does the output land where someone already works, or does it need a new screen?
• Does a human review it before it affects a patient? If not, you are building a regulated device.
• Can you measure the benefit in a number that finance recognises, such as denied claims, no-show rate, documentation minutes per encounter?
• Would a well-configured rules engine get 80% of the result? Sometimes the honest answer is yes.

Tier 1 use cases pass that test easily. Tier 3 use cases usually fail on the third question, which is exactly why they are tier 3.

Most teams that succeed here start with something administrative, prove value in a number, then earn the political capital to attempt something clinical. Teams that start with autonomous diagnosis tend not to get a second project.

Closing section

The honest summary of AI in healthcare in 2026 is that it works extremely well in a narrow band and is still being figured out everywhere else. Three quarters of regulatory clearances sit in one specialty. The fastest-spreading tool in the field is one that writes notes rather than one that makes diagnoses.

That’s not a reason to wait. It’s a reason to pick the right first problem. Administrative AI has clear economics, no regulatory submission and measurable outcomes, which is why it is where most successful programmes start.

Comfygen has been building healthcare and AI systems since 2019, across 550+ projects for 400+ clients in 30+ countries, with ISO 9001:2015, ISO/IEC 20000-1:2018 and ISO 27001 certification. The team scopes the regulatory position of an AI feature before the architecture, because whether something is a medical device changes everything downstream. If you’re weighing where AI fits in a healthcare platform, or want an existing system assessed, a scoping call will tell you which tier your idea sits in and what it realistically costs.

FAQ