This guide focuses on systems that encompass physical components, software, and model operation. Standalone software as a medical device may require specific claim architecture, and the corresponding specification should disclose model implementation details and training-data characteristics to the extent needed to support and reproduce the claimed effect. Referring to a ‘neural network’ or ‘AI model’ does not always supply missing technical detail.
The inventive contribution of an AI-assisted medical application may lie in one component, an interaction among components, or a processing step obscured by the generic label “AI.” Drafting from the perspective of the model alone may omit support for claims to device integration, and adding generic AI terminology to a conventional device may leave the asserted scope inadequately supported.
Before selecting claim categories for claims to be filed in the application, map five parts of the system: physical acquisition; signal and data processing; AI operation; system response; and the human workflow. For each part, record the problem, mechanism, output, responsible actor, and implementation alternatives.
Consider a hypothetical wearable device that uses an optical sensor to collect a physiological signal and produces an alert for review by a clinician.
Use of a trained model does not alone identify the feature(s) above on which novelty or inventive step may depend. The patenting process should thus start with the technical problem: What failed, or could not be done reliably, before the invention? Which mechanism associated with the device addresses that problem? What observable result follows from that mechanism? Guided conversations with inventor teams or conducting prior art searches may be helpful and/or necessary to pin down the answers to these questions. In Solve Intelligence’s Drafting product, users could also upload invention disclosure forms and query the tool to pinpoint the inventive features at hand.
The above questions also help separate an invention from the context of its implementation: An application focused only on the device casing and sensors may not explain the processing that produces the asserted result. It is important to document the technical chain as it identifies claim options at filing and records the relationships that may matter during prosecution.
Depending on the invention and jurisdiction, the application may support several claim forms:
These are not interchangeable versions of the same claim, as each can address different infringing conduct and different parts of the commercial system. Multi-actor arrangements can also complicate enforcement, particularly under US law, so the allocation of functions should be deliberate for each claim type. For system claims, determine which entity makes, uses, or supplies the claimed combination. For method claims, determine whether one actor performs, directs, or controls all claimed steps.
Dependent claims should provide technically meaningful fallback positions. Useful limitations may concern a sensor type, sampling condition, filtering sequence, input representation, model feature, confidence threshold, deployment location, output action, or fail-safe behavior. Simply replacing a broad result with an arbitrary implementation detail in a dependent claim is not a substitute for a fallback strategy; the narrower feature should retain the feature-to-effect relationship on which patentability is expected to depend.
“A sensor communicatively coupled to a processor” rarely explains a medical-device invention in useful depth. When discussing interfaces, explain how a physical measurement is represented digitally and how preprocessing changes it. For example, if an operation depends on a timing relationship between two signals, a particular calibration procedure, or a defined resolution, describe it explicitly. This information can support claims directed to the underlying engineering contribution. Where relevant to the proposed scope, the application should describe details including but not limited to:
Alternatives should extend beyond a list of components. If several sensor types could perform the same role, explain what they measure, what preprocessing each requires, and whether they produce the same asserted result. If inference can occur on the device or in the cloud, describe the consequences for latency, bandwidth, power, privacy, and control behavior where those consequences matter to the invention.
The disclosure need not survey machine learning generally. It should contain enough implementation detail to satisfy, at a minimum, USPTO enablement and EPO sufficiency requirements across the claimed scope. Relevant disclosure may include:
The required level of detail depends on what is claimed. If the contribution concerns a new hardware arrangement and the model is conventional, extensive model architecture may add little. If the asserted improvement depends on a particular input representation, training population, loss function, or network structure, omitting that information may leave the asserted result unsupported.
When it comes to training data, the EPO Guidelines on artificial intelligence and machine learning expressly make the distinction between disclosure and scope. The useful question is not simply whether the dataset should be disclosed (in fact, the specific dataset generally need not be), rather which specific characteristics of the data that are needed to reproduce the claimed effect should be identified. If the technical effect depends on particular characteristics of the training dataset, the characteristics required to reproduce that effect must be disclosed unless the skilled person can determine them without undue burden. These characteristics may include the signal source, measurement conditions, population characteristics, labeling method, class balance, preprocessing, or coverage of known sources of variation.
Broad functional language may be difficult to sustain if the application discloses only one narrow route to the result. Under 35 U.S.C. §112(a), a US specification must enable the skilled person to make and use the full scope of the claimed invention without undue experimentation.
The USPTO's enablement guidance applies the familiar Wands factors, including claim breadth, predictability, direction in the specification, working examples, and the experimentation required. A medical-device application that claims many sensor types, input conditions, model families, and outputs should contain disclosure proportionate to that range.
Analogously, the EPO Guidelines on sufficiency require at least one way of carrying out the invention. Where a claim covers a broad field, the description generally needs examples, alternative embodiments, or other information that allows the skilled person to work across substantially the claimed area without undue burden.
A useful drafting process starts with a concrete embodiment and then tests each proposed generalization:
Actual and prophetic examples must also be distinguished. MPEP §608.01(p) permits prophetic or “paper” examples, but predicted work should not be written in the past tense or otherwise represented as an experiment that was performed. In a field where clinical and technical validation matter, ambiguity about the status of results creates avoidable legal, credibility, and even ethical risk.
The same technical disclosure can support a global filing strategy, but the claims and legal emphasis require jurisdiction-specific review. This can be performed either by legal counsel or automatically, such as by using Solve Intelligence’s review instructions.
US drafting should treat §101 eligibility and §112(a) written description and enablement as distinct questions. The Alice/Mayo framework reflected in the USPTO's AI eligibility examples should be applied to (i) determine whether claims directed to AI usage recite a judicial exception and, if so, (ii) whether additional elements integrate that exception into a practical application/the claims nevertheless recite additional elements amounting to significantly more than the exception. A technological improvement is one route to a practical application. Other routes identified in USPTO guidance include a particular treatment, use of a particular machine, a transformation, or another meaningful limitation. Hardware recitations matter only to the extent that their claimed role affects the analysis.
The Federal Circuit's July 2026 opinion in Dental Monitoring SAS v. Align Technology, Inc. provides a current US example. The challenged claims concerned dental-arch image analysis using a “deep learning device.” The court found the claims directed to abstract ideas and found no inventive concept because the generic trained deep-learning device did not supply a specific technological solution. Naming a deep-learning device and restricting its training to domain-specific images did not establish a technological improvement on the facts and claim language before the court.
Depending on the invention, a claimed mechanism might instead include:
An assertion of improved speed or accuracy does not identify the claimed mechanism that produces the improvement. State how the improvement is produced, what baseline is being compared, and, where possible, how the result is measured.
For §112(a), explain how the claimed function is achieved. MPEP §2161 notes that computer-implemented functional claims may require disclosure of the computer and algorithm in sufficient detail to show possession of the claimed subject matter. The specification should describe the inventor's implementation rather than rely on the proposition that a skilled programmer could later produce one.
See also Solve Intelligence's PTAB case studies of AI disclosure requirements.
Under EPO guidance, AI and machine-learning computational models and algorithms are per se abstract and mathematical in nature. The EPO Guidelines do, however, identify specific technical purposes, including deriving physical or physiological quantities from sensor measurements and providing a medical diagnosis through an automated system processing physiological measurements. Merely specifying physiological data is not sufficient; the claim must link the processing to a specific technical purpose and effect.
Article 53(c) excludes qualifying method claims; claims to devices are not excluded by that provision. Under the EPO Guidelines on diagnostic methods, a diagnostic method is excluded only if the claim includes all four phases:
The technical steps constitutive of the first three phases must also satisfy the “practised on the body” criterion. A method for obtaining or processing physiological data is not automatically an excluded diagnostic method. European counsel should assess the exact claim language rather than the product's general description as “diagnostic.”
Figures should do more than show the finished product. For a device combining hardware, software, and AI, useful figures may include:
Each figure should be supported by text explaining component relationships and data flow. A block labeled “AI model” does not explain which information enters it, how that information was produced, or what happens to its output.
The application should also distinguish training from inference. The entities, data, hardware, and steps involved in training may differ substantially from those used by the deployed medical device. Combining them into one high-level flow diagram can obscure which components and steps belong to training rather than the deployed system, complicating both claim support and infringement analysis.
Before filing, consider whether each material limitation could later be detected or proven from an accused product, public documentation, testing, logs, or observable behavior. A claim dependent on an inaccessible training parameter may be difficult to investigate or prove. Similarly, a claim limited to visible housing features may omit the processing that distinguishes the commercial implementation. Rather, the claim set should reflect the invention while addressing commercially relevant conduct from several perspectives.
Future amendments must remain within the disclosure as filed under the USPTO's written-description guidance and Article 123(2) EPC. Describe alternatives while the inventors and engineers can still explain them. Record why a threshold matters, how a calibration step works, what happens when data quality falls, and which component arrangements remain viable. These details may permit amendments that preserve the asserted mechanism rather than narrow the claim to an incidental feature.
Before filing, confirm that the application answers these ten questions:
Patent counsel should convert inventor, engineering, and clinical input into jurisdiction-appropriate claims, disclosure, and fallback positions. A structured technical record provides the material to do so.
Follow the system from physical input to technical or clinical output. Describe how the device acquires a signal, how that signal is processed, what the model does, and how the result affects device operation or supports a clinical task. Then consider claims directed to the relevant device, system, method, and software aspects, with disclosure and fallback positions for each.
If a technical effect depends on particular training-dataset characteristics, disclose the characteristics required to reproduce that effect unless the skilled person can determine them without undue burden using common general knowledge. Relevant characteristics may include the signal source, measurement conditions, population characteristics, or preprocessing. The EPO Guidelines state that the specific dataset generally need not be disclosed.
This depends on where the invention lies and which conduct matters commercially. A claim set may include device or system claims, data-acquisition and signal-processing methods, inference or control methods, and jurisdiction-appropriate software claims. The specification should support the relationships among those parts even when an independent claim addresses only one of them.
Potentially. Novelty or inventive step may depend on the sensor arrangement, calibration, preprocessing, input representation, constrained deployment, control loop, or another interaction in the system rather than on the model architecture itself. The application should identify the mechanism relied upon and the result it produces instead of treating “AI” as the inventive feature.
In the United States, §101 eligibility and §112(a) written description and enablement require separate analysis. At the EPO, identify which claimed features contribute to a technical solution and assess inventive step based on that technical contribution. The application must also satisfy Article 83 EPC.
No. Article 53(c) excludes qualifying diagnostic method claims, not claims to devices. Under the EPO Guidelines, the method claim must include all four diagnostic phases, and the relevant technical steps must satisfy the “practised on the body” criterion. Claims directed to data acquisition, intermediate processing, or a device may raise different issues.
