Medical Device AI
Patenting AI Medical Devices After Alice, Recentive, and Desjardins
Your engineers built something remarkable. A neural network that spots early-stage tumors human radiologists miss. A model that predicts sepsis six hours before symptoms appear. You want a patent.
Then your attorney says the word "Alice," and the mood in the room changes.
Alice Corp. v. CLS Bank did not ban software patents. It did not ban AI patents either. But it created a two-step test that has killed thousands of applications, and machine learning claims sit directly in its blast radius. In 2025, two decisions reshaped exactly how that test applies to machine learning, and they pull in different directions. In April, the Federal Circuit's Recentive Analytics, Inc. v. Fox Corp. issued the first precedential ruling squarely addressing AI patent eligibility — and the answer was not encouraging for anyone hoping that "we used AI" would carry a claim on its own. Then, in September, the USPTO's Ex parte Desjardins went the other way, holding that a claim describing a genuine improvement to how a machine learning model itself operates can be patent-eligible, even where the underlying idea is mathematical. Desjardins was later made precedential and folded directly into the agency's examination guidance. If you are building an AI-driven medical device, you now need to understand what both decisions require — and where they still don't agree with each other.
The Alice Problem, Stated Plainly
Step one asks: is your claim directed to an abstract idea? Courts treat "analyzing data and reaching a conclusion" as abstract, almost by definition. A machine learning model does exactly that. It ingests data. It applies weights. It outputs a prediction. Described that way, every AI diagnostic tool sounds like pure mental process, the kind of thing a doctor could once do with a chart and a hunch.
Step two asks: does the claim add an "inventive concept" that transforms the abstract idea into something patent-eligible? This is where most AI medical device applications live or die. Generic computer implementation does not count. "A processor configured to run the algorithm" adds nothing. The Federal Circuit has said this again and again — and in Recentive, it said it again about machine learning specifically.
So the question becomes narrow and urgent: what, specifically, makes your AI system a technical improvement rather than an automated judgment call?
What Recentive Actually Decided
Recentive Analytics held four patents covering the use of machine learning to generate optimized event schedules and "network maps" — essentially, which programs a broadcaster airs, where, and when. Recentive sued Fox for infringement. Fox moved to dismiss on Section 101 grounds, and both the district court and the Federal Circuit agreed the patents claimed ineligible subject matter.
The Federal Circuit framed the case as one of first impression: whether claims that simply apply existing machine learning methods to a new kind of data are eligible for patent protection at all. The answer was no.
At Alice step one, the court found the patents directed to the abstract idea of using an ordinary machine learning technique within a particular setting. At step two, it agreed with the district court that the claimed techniques were generic and only broadly described, adding nothing beyond that abstract idea.
Two arguments Recentive made are worth noting, because they are exactly the arguments medical device applicants tend to reach for instinctively. First, Recentive argued its claims were eligible because they applied machine learning to a genuinely novel dataset — scheduling data no one had trained a model on before. The court rejected that outright: pointing an existing technology at a new database does not, by itself, create eligibility. Second, Recentive argued its system was valuable because it did the scheduling work faster and better than a human team could. The court rejected that too, finding that outperforming human speed does not make an otherwise abstract claim eligible.
Importantly, the court did not close the door on AI patents generally. It acknowledged that machine learning is a fast-growing field capable of producing genuinely patent-eligible advances. But it drew a sharp line: patents that do nothing more than apply generic machine learning to a new data environment, without disclosing any improvement to the underlying model itself, are ineligible. The new environment is not the invention. The new dataset is not the invention. If the model architecture, training process, or computational mechanism is unchanged from off-the-shelf machine learning, pointing it at a novel problem does not make it patentable.
For medical devices, this lands with particular force. "We applied a convolutional neural network to chest X-rays for the first time" is precisely the fact pattern Recentive rejects. The clinical domain being novel does not help you if the underlying machine learning is generic.
Desjardins: The Other Side of the Coin
Five months after Recentive, the USPTO handed down a decision that reads almost like a rebuttal. Ex parte Desjardins involved a Google patent application for a machine learning training approach. The Patent Trial and Appeal Board had rejected the claims as directed to an abstract mathematical calculation. On rehearing, the USPTO's Appeals Review Panel — under Director John Squires — reversed.
The panel didn't disagree that the claims involved a mathematical concept at Alice step one. Instead, it focused on step two: whether the abstract idea was integrated into a practical application. The panel found that it was, because the claims described an improvement to how the machine learning model itself operates, not merely the mathematical calculation the model performs. The application disclosed concrete technical benefits — reduced memory use, reduced system complexity, improved multi-task learning, and resistance to "catastrophic forgetting" during training. The panel leaned on Enfish v. Microsoft, the Federal Circuit case holding that software can make non-abstract improvements to computer technology just as hardware can,and treated it as the correct lens for evaluating machine learning claims.
The decision was later made precedential, and in December 2025 the USPTO went further, issuing formal guidance folding Desjardins into the Manual of Patent Examining Procedure. Examiners are now instructed to look for concrete, claimed benefits — improved training efficiency, reduced complexity, better data structures, and similar technical advances — rather than issuing reflexive "abstract idea" rejections against AI claims.
For medical device applicants, Desjardins is the affirmative version of the argument this article has been making all along: claim the mechanism, not the domain, and not the output. A training pipeline that measurably reduces memory footprint or prevents a model from forgetting earlier-learned patterns while training on new clinical data is exactly the kind of disclosure Desjardins protects.
Keep in mind, however, the two decisions are not reconciled. Recentive is a Federal Circuit ruling; Desjardins is a USPTO Appeals Review Panel decision, and the agency does not bind the courts. Several practitioners who have written on both cases have flagged the tension directly: Desjardins takes a more permissive view of what counts as a technical improvement than the Federal Circuit did in Recentive, and it remains an open question whether the Federal Circuit will accept the USPTO's reasoning if a Desjardins-style claim is ever litigated. Until that happens, the honest state of the law is a split posture — the agency that examines your application may allow a claim that a court, if the patent is later challenged, might still view with more skepticism. That gap doesn't cancel out the practical guidance either decision offers; if anything, it raises the stakes for drafting claims that would survive under both standards, not just the more forgiving one. A claim built to satisfy Recentive's stricter bar — genuine change to the model or training process itself, not just a new dataset or faster output — will also tend to satisfy Desjardins. The reverse is not guaranteed.
Three Failure Modes to Avoid
Most rejected AI medical device claims fail in one of three ways.
Failure mode one: claiming the diagnosis, not the machine. A claim that recites "receiving patient data, applying a trained model, and outputting a risk score" describes an outcome. It reads on what a skilled physician does mentally, just faster. Examiners will map this straight onto Mayo v. Prometheus, where correlating lab results with drug dosage was deemed an unpatentable law of nature. The USPTO treats diagnostic correlations as inherently suspect, and AI does not escape that suspicion just because a model computed the correlation instead of a person.
Failure mode two: claiming generic infrastructure. A claim that adds "wherein the method is performed by a computer" or "using a neural network" does not fix the abstraction problem. Courts have made clear that invoking a known technology by name is not the same as improving it. If your model could be any model, and your computer could be any computer, you have not claimed a technical improvement. You have claimed a wish.
Failure mode three: claiming a new application, not a new model. This is the Recentive failure mode, and it is subtly different from failure mode two. Here, the applicant does describe a real, specific machine learning pipeline — but the only thing novel about it is the domain it was pointed at. "A neural network trained to detect diabetic retinopathy from retinal scans" can fail for the same reason Recentive's scheduling patents failed, if the network architecture, training approach, and inference process are all standard. A new clinical use case, by itself, is a field-of-use limitation. Recentive confirms that field-of-use limitations do not rescue an otherwise abstract claim, no matter how valuable or novel the medical application is.
What Actually Survives
Eligible AI patent claims share a common trait. They improve how the machine works, not merely what conclusion it reaches, and not merely what domain it operates in. The Federal Circuit's guidance, the USPTO's subject matter eligibility guidance, Recentive, and now Desjardins all point to the same categories of genuine technical improvement, even if they disagree about exactly how much is enough. For medical devices, four show up most often.
1. Architecture-level innovation.
Did your team change how the model itself is built? A novel layer structure that reduces false positives in low-signal ECG data is not "a neural network." It is a specific technical solution to a specific technical problem: signal noise in a physiological waveform. Claim the architecture. Claim what data flows into which layer and why. This is the direct counterpoint to Recentive: the model itself changed, not just what it was pointed at.
2. Sensor-model integration.
Many AI medical devices pair a novel sensor configuration with a model trained to interpret its output. This is fertile ground for eligibility. If your device uses an unconventional arrangement of optical sensors, and your model was specifically trained to extract signal from that arrangement, the model and the hardware are not separable. That inseparability is evidence of a technical solution, not an abstract idea wearing a lab coat. Courts respond well to claims where removing the software breaks the hardware's function, and removing the hardware breaks the software's function.
3. Training methodology that solves a technical problem.
This one gets overlooked constantly, and both Recentive and Desjardins make it more important than ever. Claims drafted around the trained model's output often ignore the training process. But if your team developed a new way to train on sparse, heterogeneous clinical data, or a technique that corrects for sensor drift during training, that is a technical process. It solves an engineering problem: how do you get a reliable model out of unreliable data? Post-Recentive commentary has converged on exactly this point: applications fare better when they include real disclosure of how the claimed invention solves a technical problem — through a novel training process or a novel data preprocessing technique, for example, rather than a generic pipeline pointed at a new dataset. Desjardins is the clearest validation of this yet: the claims that survived there described exactly this kind of thing — a training approach that reduced memory use and prevented a model from losing previously learned patterns as it trained on new data. That is not "organizing human activity." That is signal processing with extra steps, and it is the opposite of the generic training both decisions found wanting when it's absent.
4. Real-time computational improvement.
Speed and resource efficiency count as technical improvements when courts can see the mechanism — but note the limits Recentive puts on this. A claim that says "the model runs faster than a human scheduler" is the exact argument Recentive rejected; the Federal Circuit was clear that outperforming human speed does not confer eligibility. What survives is a claim describing the mechanism of the improvement: "the system reduces memory footprint by pruning redundant convolutional filters identified through gradient magnitude thresholding" describes how the computer's own operation changed, not just that the output arrived faster than a person could produce it.
The FDA Connection Nobody Talks About
Here is something patent attorneys underuse: regulatory evidence. If your device requires FDA clearance as Software as a Medical Device, that regulatory pathway generates documentation describing your device's technical function in granular detail. Validation studies. Failure mode analyses. Algorithm change protocols.
That documentation is a goldmine for patent drafting. It forces precision about what the algorithm actually does, mechanically, inside the device. Draft your claims using that same technical vocabulary. If your 510(k) submission describes a specific preprocessing pipeline that corrects for motion artifact before the model ever sees the data, that pipeline belongs in your independent claim. It is concrete. It is mechanical. It is exactly the kind of disclosure Recentive says was missing from Recentive Analytics' patents, and exactly the kind of "specific means or method" language the Federal Circuit rewards.
Regulatory rigor and patent eligibility want the same thing from you: precision about mechanism. Use one process to feed the other.
A Drafting Framework
When you sit down to draft claims for an AI medical device, ask these questions in order.
First, strip away the AI. What is left? If nothing technical remains, an abstract idea is doing all the work, and no amount of "processor" language will save it.
Second, strip away the clinical domain. What is left of the machine learning itself? If the answer is "an off-the-shelf model architecture, trained the standard way, pointed at a new dataset," you have Recentive's fact pattern. A novel disease, novel patient population, or novel data source is not, by itself, a technical improvement to the model. If instead you can point to something concrete that changed about how the model trains, stores information, or avoids losing previously learned patterns, you are closer to Desjardins territory — but draft it to survive Recentive's stricter reading too, not just the USPTO's more permissive one, since the two are not yet reconciled.
Third, identify the specific technical bottleneck your team solved. Not "diagnosis is hard." Something narrower: "conventional models overfit on imbalanced arrhythmia datasets because rare arrhythmia classes are underrepresented." Now you have a technical problem with a technical shape.
Fourth, describe your solution as a mechanism, not an outcome. Not "the model classifies arrhythmias more accurately" and not "the model classifies arrhythmias faster than a cardiologist." Instead: "the system applies a weighted loss function that penalizes misclassification of underrepresented arrhythmia classes proportional to their inverse frequency in the training set." That sentence could survive both Alice step two and Recentive. The other two could not.
Fifth, tie the claim to hardware or data structure wherever honestly possible. A claim anchored to a specific sensor array, a specific waveform preprocessing step, or a specific memory architecture is harder to dismiss as abstract than a claim floating in pure algorithmic space.
Sixth, write your specification like an engineer, not a marketer. Eligibility fights are often won or lost in the specification, not the claims. If your spec explains mechanism in detail — including how the model or training process itself differs from generic machine learning — your claims inherit that credibility. If your spec reads like a press release about a new clinical use case, so will your claims, and examiners notice.
The Bottom Line
Alice did not close the door on AI medical device patents, and neither did Recentive or Desjardins. All three changed what counts as an invention. A diagnosis is not an invention. A prediction is not an invention. Applying a standard model to a new disease is not an invention, no matter how urgently that disease needed better tools. But the specific technical architecture, training methodology, or computational mechanism that makes an accurate, reliable, clinically deployable prediction possible — that can be an invention, and a patentable one.
What's changed is that you now have both a warning and a blueprint from 2025, and they don't fully agree with each other. Recentive tells you what a court is likely to reject: generic machine learning, however novel the application. Desjardins tells you what the patent office is now willing to allow: a genuine, disclosed improvement to how the model itself trains, stores, or generalizes. The prudent path is to draft for the stricter of the two, since a court has the final word and hasn't yet weighed in on whether it will follow the USPTO's more permissive reasoning.
The lawyers who win these cases are the ones who resist the urge to claim the magic trick, and resist the urge to claim the domain. They claim the mechanism behind it. Find the engineering problem your team actually solved inside the model itself, describe it with precision, and let the eligibility argument write itself.
If your AI medical device claim can be summarized as "a computer figures out if the patient is sick" or "we pointed a standard model at a new kind of scan," start over. If it can be summarized as "a system trained on a novel data pipeline corrects a specific, previously unsolved technical limitation in the model itself," you are in good shape — under Recentive, under Desjardins, and under whatever court eventually has to decide where the two meet. That distinction is not academic. It is the entire ballgame.