Software as a medical device is one of the harder areas to patent well, because it combines two independent difficulties. Section 101 eligibility is contested for diagnostic and algorithmic inventions, and the regulatory pathway generates public disclosures on a schedule you do not fully control.
Both are manageable. Neither is manageable by accident.
The Eligibility Problem
Diagnostic software sits close to two judicial exceptions at once.
Abstract ideas capture claims that amount to collecting data, analyzing it, and displaying a result. That describes a great deal of clinical software at a high level.
Laws of nature capture claims that amount to observing a natural correlation. Mayo v. Prometheus invalidated claims to correlating drug metabolite levels with dosing needs, on the ground that the correlation itself was a natural law and the remaining steps were conventional. That holding sits directly across the path of any invention whose value is "we discovered that X predicts Y."
The distinction that decides these cases is whether you claim the correlation or a specific technical means of detecting and acting on it.
A claim to "determining a patient's sepsis risk by applying a machine learning model to vital sign data" is a losing claim. It recites an abstract result achieved with generic computing.
A claim reciting a specific architecture, for example ingesting asynchronous vital sign streams at differing sample rates, imputing gaps using a learned temporal model conditioned on the missingness pattern, and triggering an alert when the calibrated risk score crosses a threshold that adapts to unit-level baseline acuity, is claiming a technical mechanism that solves technical problems. Irregular sampling, missing data, and alarm fatigue are engineering problems, and solutions to them are patentable subject matter.
The invention is often identical in both versions. The difference is whether the application describes what the software concludes or how it computes.
What Strengthens a SaMD Application
Claim the pipeline, not the insight. Preprocessing, signal conditioning, feature extraction, model architecture, calibration, threshold adaptation, and the interface to the clinical workflow are all technical subject matter.
Name the technical problem in technical terms. Not "improves patient outcomes." Rather: reduces false alarm rate on noisy telemetry, operates within the memory constraints of an embedded monitor, handles sensor dropout without degrading sensitivity, maintains calibration across scanner vendors.
Tie it to something physical where you can. SaMD claims that interact with a sensor, a device, or a physical measurement process are meaningfully easier to sustain than pure data-processing claims. If your software runs on or controls a device, claim that relationship.
Include the validation data. Sensitivity, specificity, and comparative performance belong in the specification. They support non-obviousness through unexpected results and they demonstrate the technical effect that eligibility analysis looks for.
Describe the training methodology in detail even if you intend to keep parts of it secret, because you cannot add it later. Then decide deliberately what to claim.
The Trade Secret Question Is Sharper Here
For SaMD, the patent-versus-trade-secret analysis has an unusually clean answer in many cases, and it turns on detectability.
Model weights and training data curation are typically invisible in a deployed product. A competitor cannot extract your training pipeline from your API. Patenting it publishes the recipe in exchange for a right that is nearly impossible to detect infringement of. These are often better as trade secrets.
Architecture, workflow integration, and device interaction are frequently observable in operation, sometimes described in your own regulatory submission. Secrecy will not hold, so patent them.
Many digital health companies should run both tracks simultaneously: patents on the observable system architecture, trade secrets on the training methodology, with documented reasonable measures around the latter.
The Regulatory Interaction
Submissions publish. FDA posts 510(k) summaries and De Novo decision summaries. Those describe your software's intended use and often its operating principles. Once published, that is prior art against you.
Predetermined Change Control Plans complicate timing. FDA's framework for pre-authorizing certain model updates means your device is explicitly designed to change after clearance. Patent claims drafted only to the cleared version may not cover the version running in eighteen months. Continuations matter more here than almost anywhere else.
Clinical validation studies register publicly. Same disclosure problem as any device trial.
The pathway itself signals the field. A De Novo grant creates a new classification regulation that competitors can use as a predicate.
The sequencing rule is the same as for hardware: file before anything publishes, and keep the family open.
Inventorship With AI in the Loop
If AI tools contributed to developing your algorithm, inventorship still requires a natural person. The Federal Circuit held in Thaler v. Vidal that an inventor must be human, and USPTO guidance issued in 2024 confirms that AI-assisted inventions are patentable so long as a natural person made a significant contribution to conception. Document who contributed what.
The Short Version
SaMD is patentable, and the applications that fail usually failed at drafting, by claiming a clinical conclusion instead of a computational mechanism. Claim the pipeline. Name the technical problem. Tie it to something physical where you can. File before FDA publishes. Keep continuations pending because the software will change. And decide deliberately which parts to patent and which to keep secret, because for SaMD those are genuinely different answers for different components of the same product.
To discuss protection for a digital health or SaMD product, request a free consultation.
This article is general educational information, not legal advice, and reading it does not create an attorney-client relationship. Patent law is fact specific and deadlines are unforgiving. For advice on your situation, schedule a consultation.