This issue, we close out our three-part Signal From Noise series into the deployment of AI into health care. The first two installments were devoted to how AI is helping researchers to identify promising potential medicines and treatments, and then how AI is helping companies to determine – and prove – that the new drug candidates are both safe and effective. However, the practice of medicine and healthcare go far beyond simply popping pills or having shots administered. In fact, there are other aspects of medicine that are arguably more important.
AI is helping in these areas as well. In this issue, we discuss how, and we look at some of the companies in these areas.
To a large extent, most applications of AI in a clinical setting fall under the category of medical technology. The Food and Drug Administration (FDA) currently lists approvals for 1,614 medical devices that use machine learning or AI. The list shows 181 such approvals in 2026 as of Sept. 4, 2026, which would put us on pace to set a new record this year.
The FDA approval process for medical devices differs from the process for pharmaceuticals. Although there are three possible pathways to get a new medical device cleared or approved, it’s fair to say that each of the three features a lower standard than for pharmaceuticals.
The 510(k) standard
This is a widely used path, particularly for AI-powered medical technology. That’s because once a pioneering company has won approval for its device, any subsequent competitor need only show that its new device is “substantially equivalent” to any other already-approved device in terms of safety, functionality/technology, and accuracy. Often, human clinical trials are not even required, making this a relatively short, six-month process. The 510(k) pathway is also used to obtain approval for iterative improvements (for instance if the software behind a piece of medtech is updated or if the hardware is upgraded in a subsequent model). Notably, the medtech industry is lobbying to ease this requirement for adaptive/learning AIs.
De Novo
If a company develops a completely new technology with no pre-existing predicate, the process is somewhat more challenging, though with a typical 9- to 12-month timeframe, still arguably not terribly onerous. Basically, a De Novo standard requires that a company provide clinical data to show that the device presents relatively low or moderate risk and is effective.
Premarket Approval (PMA)
For new medtech devices that are deemed to be high risk – devices intended to sustain human life or carry significant risk of severe injury, the PMA approval process is required. This is the strictest of the three, approaching the standards seen in clinical drug trials. The time and expense required for PMA approval is similarly challenging.
Clinical efficacy standards
Partly because of the FDA’s lower standards for medtech approval, clinicians can sometimes be reluctant to buy into a new technology. Clinicians also have sometimes expressed reluctance to try AI-empowered devices because AI decisions tend to be black boxes, and providers fear that AI-driven errors will increase doctors’ potential liabilities. Thus, regulatory approval does not obviate the need for peer-reviewed research afterwards, so companies often form research partnerships with academic medical centers even after regulators have given the go-ahead to go to market.
Triage
AI-powered medical technology is now finding its way into every stage of patient care. Such devices and platforms are now used to assess patients the moment they arrive at a provider location.
In emergency rooms, privately held Aidoc and Viz.ai now offer systems that provide high-speed analyses of CT scans, enabling ER personnel to quickly identify patients likely to have time-critical conditions such as intracranial hemorrhaging or pulmonary embolisms.
Meanwhile, one of the more prominent offerings from GE HealthCare GEHC 1.44% is an AI-powered mobile X-ray system that triages chest X-rays in real time to quickly alert physicians of potentially life-threatening conditions (for instance, a collapsed lung) that require immediate attention and an AI-assisted handheld ultrasound that helps guide users on optimal probe positioning. The company has multiple offerings that apply AI algorithms to raw MRI data in real time to produce sharper, high-resolution images at a drastically faster speed.
Beyond triage
These applications build on AI’s affinity for image analysis. As we wrote in 2023, AI-powered vision can carry out significantly speedier image analysis, often with accuracy that matches or exceeds that of human radiologists. Companies have now built upon that: the FDA list of approved medical technology devices shows that radiology is by far the most common medical specialty toward which such devices are geared, leaning on AI-powered computer vision to analyze images of patient internals. This has gone beyond emergency/triage situations and into more specialized areas of care.
Siemens Healthineers SMMNY has developed a multi-organ, multi-modal AI-powered scanning solution that applies AI to chest CT images of lungs and coronary arteries; brain MRIs to assess the progression of Alzheimer’s and other neurodegenerative diseases; and prostate MRIs that identify suspicious regions to prioritize when planning biopsies. The company’s major technology also focuses on applying AI to speeding the acquisition of medical images at high resolution/quality.
Similarly, Philips PHG 0.97% has developed a platform for AI-enhanced MRIs that enable faster, higher-resolution scans – the company claims an 80% improvement in image sharpness at a third of the usual time, as well as cardiac imaging (ultrasound and MRI) solutions that automate complex 3D cardiac measurements, isolating specific areas (heart chambers, for instance) for closer scrutiny.
On a specialized note:
- Pathologists are using AI to carry out high-throughput slide analyses, particularly in biopsies to detect prostate, breast, and cervical cancers. Paige Prostate (owned by Tempus AI TEM -6.38% ), along with privately held Ibex Medical Analytics and Aiforia. are among the companies offering a solution in this regard.
- HeartFlow’s HTFL -19.13% Plaque Analysis tool helps cardiologists use AI-powered image analysis to quantify and segment coronary artery plaque levels.
- Non-opthalmologist healthcare providers can perform diabetic retinopathy screening with the LumineticsCore AI diagnostic platform offered by privately held Digital Diagnostics.
AI is also finding its way into surgical applications. While we are not quite at a point where physical AI can be trusted to perform autonomous surgeries, Medtronic MDT 1.85% ’s Touch Surgery AI and Hugo Robotic-Assisted Surgery platform can now help monitor laparoscopic and robotic procedures in real time. On a more specialized note, Zimmer Biomet (ZBH -1.28% ) has an integrated knee replacement platform that combines robotic surgery with a smart knee implant embedded with accelerometers and sensors to monitor surgical recovery, joint performance, and potential joint failure.
Intuitive Surgical ISRG 0.82% does not have an AI-powered version of its da Vinci robotic surgical platform yet, but data from the procedures performed on da Vinci (20 million procedures and counting) is being used to train a model with the aim of achieving supervised autonomy.
Generalized clinical care
The applications above have the potential to improve patient outcomes, but there’s a case to be made that it is just as important to make the lives of physicians, nurses, and healthcare providers easier.
Such efforts tend to focus on three areas. The first is clinical reference.
Even those of us who aren’t medical professionals understand just how much raw knowledge needs to be acquired and internalized to become a healthcare professional. However, the nature of human memory means that for most, much of that knowledge is forgotten. Generally, only the knowledge that is used constantly ends up being retained.
A morbidly amusing semi-anecdotal meme arose during the days of the COVID pandemic that illustrates the point. The pandemic forced the postponement of many eye surgeries because they are often non-urgent in nature. At the same time, the pandemic led to an urgent need for additional doctors in emergency rooms and intensive care units. Consequently, ophthalmologists, most of whom hadn’t done anything that didn’t involve the eyes for years if not decades, were deployed to hospital wards where they found themselves needing to defer to battle-tested nursing staff in nearly all clinical matters and instead relegating themselves to supportive (though still important) work like communicating with families, handling paperwork, and overseeing proning (flipping ventilated patients). (Similar anecdotes exist for doctors in other specialties such as dermatology.)
The point is that medical practitioners absorb an enormous volume of knowledge during medical school and internship, but over time, only the most frequently used information is readily retained. When a doctor needs to recall the rest of it, clinical AI tools now make it easier and more effective. LLM-based evidence tools are increasingly being used so physicians can recall rarely accessed knowledge or update their understanding. Generalized LLM medical references include
- UpToDate Expert AI, provided by Wolters Kluwer (WTKWY), a well regarded publisher of clinical reference databases. UpToDate’s generative LLM is trained on more than 13,000 expert-vetted topics.
- Dyna AI, trained on privately held EBSCO Industries’ clinical summaries and pharmacology databases. Dyna AI is particularly useful at complex analyses of potential drug-to-drug interactions.
- OpenEvidence, a privately held platform trained on millions of PubMed articles and designed for point-of-care/bedside use. The AI provides chain-of-thought reasoning in response to clinical queries, with the ability to ground answers in peer-reviewed PubMed literature – with citations.
A notable specialized LLM reference is ClinicalKey AI, built by Elsevier RELX 1.17% and trained on the company’s range of surgical textbooks, procedural guides and high-resolution anatomical atlases. ClinicalKey enables surgeons and other specialists to quickly verify complex procedures and anatomical relationships as they prepare for surgeries.
Another platform, provides clinical reference with workflow tools, providing a bridge to our next area of discussion. Doximity Ask DOCS 0.42% provides AI-powered query functionality, enabling physicians to reference PubMed literature, medical society guidelines, and drug monographs to generate answers with citations. More importantly, it helps clinicians deal with something that nobody likes: paperwork. Doximity Ask serves as a Health Insurance Portability and Accountability Act (HIPAA)-compliant clinical assistant that can help draft patient notes and write authorization requests to insurers.
Ask physicians what the worst parts of their jobs are, and odds are patient notes and dealing with insurance companies will come up pretty quickly. In fact, a recent Medscape survey suggests that such tasks are not only a major expense for private health practices, but also the common cause of physician burnout – more so than just generalized overwork.
Yet writing up notes about patients is essential: after every patient visit, physicians need to make a record of their contemporaneous observations, the recommendations made, and the course of actions taken. This is essential because it helps to ensure continuity of care over long periods of time and across providers. Yet the traditional method of writing up patient notes and records often added hours to a provider’s workday.
Speech-to-text transcription tools have been around for quite some time now, and a number of AI-powered versions have become increasingly accurate and effective. Yet specialized transcription tools for physicians are still required for several important reasons.
- Privacy. General speech-to-text tools, whether Siri, the dictation feature in Microsoft Word, or a newer AI-powered service like Whisper, are not offered under the business associate agreements that HIPAA requires for handling patient information, at least in their standard consumer form.
- Specialized vocabulary. General transcription services are not trained on the specialized, complex vocabulary used by healthcare professionals and can often make rather alarming mistakes (inadvertently transcribing “hypotension” as “hypertension” for example.)
- Formatting. Specialized medical transcription is built around knowledge of the common formats used to write patient notes and reports, so spoken remarks are automatically understood and formatted correctly.
- Context. Medical transcription platforms go beyond dictation. A modern transcription tool can listen in on ambient doctor-patient conversations, identifying the voices of doctor, nurse, patient, and patient family members, accessing patient records to clarify context and filtering out social banter to draft documents that are ready for review before the patient even makes it out of the examination room.
One of the most widely adopted transcription platforms is Dragon Copilot. Developed by the Nuance Communications division of Microsoft MSFT -0.38% , Dragon makes use of a specially developed medical speech engine and has become embedded in major electronic health record (EHR) platforms like Epic, Oracle Health’s Cerner, and Meditech. As such, it has become widely adopted by large health systems. Dragon Copilot listens ambiently to doctor-patient conversations, drafting patient notes and clinical summaries in real time.
Two of these platforms now have their own AI transcription tools. Epic recently launched its own AI Charting tool, though it leverages Dragon’s transcription and note-generation technologies. Also, Oracle ORCL -2.06% , which owns Cerner, recently launched Oracle Health Clinical AI Agent, offering ambient transcription, realtime chart queries, and the automatic staging of orders.
These three have one other major competitor: Abridge AI’s Abridge. Abridge has been integrated into nationally known healthcare systems like the Mayo Clinic, Kaiser Permanente, and Duke Health. Most AI medical transcriptions still require human editing, but some studies suggest that Abridge transcriptions require significantly less editing and offer greater accuracy across medical specialties.
Other options include Suki, which doubles as an assistant that can look up clinical and patient data in real time during a visit; Commure Scribe, which features integration into revenue-cycle management and billing systems; and Freed, a lightweight option for smaller practices with quick setup and a monthly-fee model.
Transcription isn’t just important due to its role in ensuring continuity of care. Patient records are an important part of the documentation required by healthcare providers to collect insurance reimbursement for services rendered, and when AI transcription platforms are used, they assist in identifying patient complaints, diagnoses, assessments, services, and treatment recommendations and linking them to the appropriate billing codes to submit to insurers. AI tools can also ensure that providers collect the maximum to which they are entitled, when before, a tired physician might have rushed through their paperwork and missed submitting codes and claims for work that they did.
This might seem like a money grab for physicians. Indeed, insurers have complained that the growing adoption of AI scribes has increased submitted coding intensity, resulting in higher healthcare costs across the U.S. – and lower margins for insurance providers.
There is a counter argument, however: insurance companies in recent years have increasingly used AI tools to quickly identify reasons for denying or reducing claims submitted by providers. In recent years, insurers like UnitedHealth, Cigna, and Humana have begun using in-house AI tools to deny claims en masse, all without adequate human review. (Several pending class-action lawsuits allege that the use of such systems has led to routine rejection of medically necessary care.)
Monitoring
One of the longest-standing challenges in health care is due to the fact that meaningful health events do not tend to take place precisely when a healthcare professional is present and watching – during a doctor’s exam, for instance. For hospitalized patients, existing monitoring methods have tended to be suboptimal – sometimes intermittent rather than constant, often intrusive, and frequently not sufficiently predictive.
Companies and providers hope that a range of new AI-powered devices will make it easier for healthcare professionals to monitor patients while they are hospitalized, providing detailed data that can better inform ongoing treatment while also more quickly identifying emerging developments that require immediate human attention. Devices have been developed that help nurses better monitor patients for an escalating risk of serious conditions such as:
- Sepsis. Bayesian Health, a privately held spinoff from Johns Hopkins University, has developed an early-warning system for sepsis, a fast-moving and often-fatal condition in which the body overreacts to infection, destroying its own organs in a chain reaction. Its FDA-cleared “Targeted Real-Time Early Warning System” (TREWS) module tracks tiny shifts in temperature, heart rate, and other metrics and claims to be able to spot signs of sepsis an average of around six hours sooner than traditional methods. If true, this would be a game changer when traditionally, sepsis can result in death as few as 12 hours after detection.
- Respiratory distress/failure. Patients on opioid pain medications or recovering from surgical anesthesia sometimes quietly stop breathing deeply, leading to a severe shortfall in oxygen intake and carbon dioxide removal. Masimo, a division of Danaher DHR 0.37% , has developed Masimo Radius VSM, a wireless, wearable sensor that constantly monitors blood oxygen, pulse, and breathing patterns to provide earlier warnings before respiratory arrest actually occurs.
- Cardiac arrest/life-threatening arrhythmias. Becton, Dickinson’s BDX 1.08% Advanced Patient Monitoring product incorporates the ability to detect micro-level changes within a patient’s heartbeat waveform to spot changes in how blood vessels are expanding/contracting. The company claims that this can provide up to 15 additional minutes of warning before a dangerous cardiac event.
Monitoring patients who are managing chronic conditions or ailments can provide similar benefits, even if they are not judged to be in immediate danger or acute risk. However, outside of a hospital setting, this has traditionally been logistically challenging.
That could be changing. Consumer electronics companies like Apple, Garmin, and Samsung are increasingly seeking to incorporate a broader range of health monitoring capabilities, such as electrocardiogram (ECG) readings for atrial fibrillation, sleep apnea detection and notifications, and blood oxygen tracking.
However, serious, at-home, medical-grade monitoring is being developed for conditions including:
- Congestive heart failure. Zoll, a subsidiary of Asahi Kasei Corp. AHKSY, has developed an adhesive patch worn on the chest that uses low-powered radar waves to monitor lung fluid levels, heart rate, and posture. Alternatively, Abbott Laboratories ABT 1.18% has launched CardioMEMS HF System, which tracks internal pulmonary artery pressure through a small wireless sensor implanted in that blood vessel in order to provide diagnostic data to doctors.
- Diabetes. Dexcom’s DXCM 2.38% AI-powered G7 Glucose monitoring system replaces the traditional fingerstick (pin prick) blood tests with a small sensor worn on the back of the arm that continuously measures subcutaneous glucose levels, with predictive algorithms that tracks the speed and direction of blood sugar changes. This can be used to automate insulin-pump triggering, notify patients when they should eat something, and ultimately help them make more informed lifestyle choices. Abbott Laboratories ABT 1.18% has a competing FreeStyle Libre 3 Plus system that uses a small arm sensor to stream minute-by-minute glucose readings to a smartphone and can integrate with automated insulin pumps and meal-prediction tools.
- Cardiac arrhythmias. Certain types of irregular heart rhythms, notably atrial fibrillation, can serve as an important warning sign of increased stroke risk, but their occurrences rarely coincide with a patient visit to their cardiologist. IRhythm Technologies’ IRTC -1.76% Zio ECG Monitor System can provide cardiologists with up to 14 days of continuous ECG data, along with FDA-cleared AI waveform analysis and remote alerts, to hopefully serve as a more effective and timely warning system.
Conclusion
It’s important to note two major obstacles for companies seeking to develop solutions that integrate AI into the practice of medicine. The first is that, as with all sorts of AI applications, this is a nascent field, and sometimes it shows.
Somewhat infamously, one study showed that one of the earlier AI-powered methods to monitor patients for signs of sepsis, developed by Epic Systems, was noticeably less accurate and effective than the existing traditional methods, missing two-thirds of sepsis cases yet also generating numerous false alarms. (Epic later released a version that significantly improved, but did not fully resolve, these issues.) In January, the FDA disclosed that Abbott had reported 860 serious injuries and seven deaths associated with some FreeStyle Libre3 sensors, apparently due to inaccurately low glucose readings.
The second major obstacle is the U.S. insurance industry. The FDA serves as a major gatekeeper for what gets deployed in the U.S. healthcare system, but insurers are arguably just as important. For hospitals and other healthcare providers to buy into AI-empowered technology, it’s not enough to show that it will improve patient care or make providers’ jobs easier. Companies must also help convince insurers and healthcare providers that the tool in question will either enable/facilitate new billable services, reduce costs (possibly by reducing hospital stays or adverse health events), or make clinicians more efficient. In other words, companies in this space need to show that their products deliver positive ROI.
Despite these obstacles, however, our view is that AI in health care has moved past a pipe dream and proof of concept stage. The trend toward widespread adoption seems unlikely to us to be a steep, straight line upward. Nevertheless, a range of companies, both long-established and new, are making material advances in their efforts to put AI to use in improving health care, and we thus view this as supportive of the broader AI investment thesis in general.
As always, Signal From Noise should not be used as a source of investment recommendations but rather ideas for further investigation. We encourage you to explore our full Signal From Noise library, which includes a look at physical AI, cybersecurity stocks, companies making “moonshot” bets, the investment trends driven by Gen Xers, and the recent memory chip gold rush. You can also find our take on the future of malls, drone warfare, space-exploration investments, and more.
