PatentNext Summary: In order to prepare patent applications for filing in multiple jurisdictions, practitioners should be cognizant of claiming styles of the various jurisdictions that they expect to file AI-related patent applications in, and draft claims accordingly. For example, different jurisdictions, such as the U.S. and EPO, have different legal tests that can result in different styles for claiming artificial intelligence(AI)-related inventions.
In this article, we will compare two applications, one in the U.S. and the other in the EPO, that have the same or similar claims. Both applications claim priority to the same PCT Application (PCT/AT2006/000457) (the “’427 PCT Application”), which is published as PCT Pub. No. WO/2007/053868.
As we shall see, despite the application having the same or similar claims, prosecution of the applications in the two jurisdictions nonetheless resulted in different outcomes, with the U.S. application prosecuted to allowance and the EPO application ending in rejection.
Artificial Intelligence (AI) Overview
Pertinent to our discussion is an overview of AI. A brief description of AI follows before analysis of the AI-related claims at issue.
Artificial Intelligence (AI) is fundamentally a data-driven technology that takes unique datasets as input to train AI computer models. Once trained, an AI computer model may take new data as input to predict, classify, or otherwise output results for use in a variety of applications.
Machine learning, arguably the most widely used AI technique, may be described as a process that uses data and algorithms to train (or teach) computer models, which usually involves the training of weights of the model. Training typically involves calculating and updating mathematical weights (i.e., numeral values) of a model based on input that can comprise hundreds, thousands, millions, etc. sets of data. The trained model allows the computer to make decisions without the need for explicit or rule-based programming.
In particular, machine learning algorithms build a model on training data to identify and extract patterns from the data and therefore acquire (or learn) unique knowledge that can be applied to new data sets.
For more information, see Artificial Intelligence & the Intellectual Property Landscape
Sufficiency of Disclosure in the U.S.
AI inventions are fundamentally software-related inventions. In the U.S., as a practical rule, software-related patents should disclose an algorithm by which the software-related invention is achieved. An algorithm provides support for a software-related patent pursuant to 35 U.S.C. § 112(a) including (1) by providing sufficiency of disclosure for the patent’s “written description” and (2) by “enabling” one of ordinary skill in the art (e.g., a computer engineer or computer programmer) to make or use the related software-related invention without “undue experimentation.” Without such support, a patent claim can be held invalid. For more information regarding general aspects of the sufficiency of disclosure in the U.S. for software-related inventions, see Why including an “Algorithm” is Important for Software Patents (Part 2)
1. The ’457 PCT Application in the U.S.
U.S. Patent 8,920,327 (the “’327 Patent”) issued from the ’457 PCT Application. The ‘’327 Patent is an example of an AI patent that did not experience sufficiency issues in the U.S. The below provides an overview of the ’327 Patent.
The ’327 Patent is titled “Method for Determining Cardiac Output” and includes a single independent claim regarding a method for cardiac output from an arterial blood pressure curve. The method is implemented via a cardiac device, as illustrated in Figure 1 (copied below):
Fig. 1 illustrates device 1 for implementing the invention of the ‘327 patent, where measuring device 2 measures the peripheral blood pressure curve, and where related measurement data is fed into device 1 via line 3, and stored and evaluated there. The device further comprises optical display device 4, input panel 5, and keys 6 for inputting and displaying information.
The claimed method includes an AI aspect, i.e., namely the use of “an artificial neural network having weighting values that are determined by learning.”
Claim 1 is copied below (with the AI aspect bolded):
1. A method for determining cardiac output from an arterial blood pressure curve measured at a peripheral region, comprising the steps of:
measuring the arterial blood pressure curve at the peripheral region; arithmetically transforming the measured arterial blood pressure curve to an equivalent aortic pressure; and
calculating the cardiac output from the equivalent aortic pressure,
wherein the arithmetic transformation of the arterial blood pressure curve measured at the peripheral region into the equivalent aortic pressure is performed by the aid of an artificial neural network having weighting values that are determined by learning.
Figure 3 of the ‘327 patent (copied below) is a schematic illustration of the artificial neural network, as recited in claim 1.
The specification of the ’327 patent describes that “FIG. 3 illustrates the structure of the neural network…, and it is apparent that the neural network … is comprised of three layers 14, 15, 16.” The specification discloses that a supervised learning algorithm is used to train the weights of the model, e.g., “[t]he weights and the bias for the latter two layers 15 and 16 are determined by supervised learning.”
The input data fed to the supervised learning algorithm to train the AI model includes “associated blood pressure curve pairs actually determined by measurements in the periphery or in the aorta, respectively, are used.” The measurements used for the input data may come “from patients of different ages, sexes, constitutional types, health conditions and the like.”
No issues with respect to sufficiency were raised during the prosecution of the application in the U.S. that was issued as the ’327 patent.
More generally, issues of sufficiency in the U.S. typically arise in litigation, and result in expert testimony, i.e., “a battle of the experts,” where expert witnesses (e.g., typically university professors or industry consultants) from opposing sides opine on the knowledge of a person of ordinary skill in the art and sufficiency of disclosure in view of that person. For an example of how this played out in a U.S. case, see: Sufficiency of Disclosure for Artificial Intelligence Patents – U.S. Case Example
Sufficiency of Disclosure in the European Patent Office (EPO)
The EPO has developed its own, yet similar, stance on AI-related invention when compared with the U.S. Nonetheless, outcomes of prosecution can be different. The below provides a cursory overview of developments in the EPO with respect to AI-related inventions and analyzes the treatment of an EPO application as filed based on the PCT Application ’457 (which is the same PCT Application as for the ’327 patent discussed above).
1. Artificial Intelligence Inventions can be patented pursuant to EPO law
Generally, artificial intelligence inventions may be patented in the European Patent Office (EPO). For example, in its Guidelines for Examination, the EPO defines AI and machine learning as “based on computational models and algorithms for classification, clustering, regression and dimensionality reduction, such as neural networks, genetic algorithms, support vector machines, k-means, kernel regression and discriminant analysis.” Section 3.3.1 (Artificial intelligence and machine learning).
As such, the EPO dubs AI and machine learning as “per se of an abstract mathematical nature,” irrespective of whether such models may be trained with training data. Id. Thus, simply claiming a machine learning model (e.g., such as a “neural network”) does not, alone, necessarily imply the use of a “technical means” in accordance with EPO law.
Nonetheless, the Guidelines for Examination at the EPO recognize that the use of an AI model, when claimed as a whole with the additional subject matter, may demonstrate a sufficient technical character. Id. As an example, the Guidelines for Examination at the EPO states that “the use of a neural network in a heart-monitoring apparatus for the purpose of identifying irregular heartbeats makes a technical contribution.” Id. As a further example, the EPO Guidelines for Examination further states that “[t]he classification of digital images, videos, audio or speech signals based on low-level features (e.g. edges or pixel attributes for images) are further typical technical applications of classification algorithms.” Id.
2. Sufficiency of Disclosure for an Artificial Invention at the EPO
In a decision in 2020, the EPO Board of Appeals rejected a machine learning-based patent application that claimed an “artificial neural network” because the patent specification failed to sufficiently disclose how the artificial neural network was trained. See T0161/18 (Equivalent aortic pressure / ARC SEIBERSDORF). The application in question claimed priority to the PCT Application ’457, which is the same parent application as the ’327 patent, as discussed above.
The claims were the same or similar as to those in the U.S., where the claims-at-issue directed to determining cardiac output from an arterial blood pressure curve measured at a periphery, and recited, in part (with respect to AI), that the “blood pressure curve measured on the periphery is converted into the equivalent aortic pressure with the help of an artificial neural network, the weighting values of which are determined by learning.”
Claim 1 is reproduced below (in English based on a machine translation of the original opinion German):
1. A method for determining the cardiac output from an arterial blood pressure curve measured at the periphery, in which the blood pressure curve measured at the periphery is mathematically transformed to the equivalent aortic pressure and the cardiac output is calculated from the equivalent aortic pressure, characterized in that the transformation of the blood pressure curve measured on the periphery is converted into the equivalent aortic pressure with the help of an artificial neural network, the weighting values of which are determined by learning.
A. Sufficiency of Disclosure
The Board analyzed the claim in view of the specification pursuant to Article 83 EP (Sufficient disclosure). As described by the Board, Article 83 EPC requires that the invention be disclosed in the European patent application so clearly and completely that a person skilled in the art can carry it out. For this, the disclosure of the invention in the application must enable the person skilled in the art to reproduce the technical teaching inherent in the claimed invention on the basis of his general specialist knowledge.
The Board then turned to the specification to determine whether it disclosed enough support to meet these requirements in view of the claimed “artificial neural network.” However, the specification was found lacking because it failed to “disclose which input data are suitable for training the artificial neural network according to the invention, or at least one data set suitable for solving the technical problem at hand.”
Instead, the Board found that the specification “merely reveals that the input data should cover a broad spectrum of patients of different ages, genders, constitution types, health status and the like.”
Therefore, the Board found that the training of the artificial neural network could therefore not be reworked by the person skilled in the art, and the person skilled in the art can therefore not carry out the invention.
Because of these deficiencies, the Board found that the specification failed to provide sufficient disclosure pursuant to Article 83 EPC.
B. Inventive Step
For similar reasons, the Board further found that the claimed subject matter lacked an “inventive step” pursuant to Article 56 EPC. Specifically, the Board found that the claimed “artificial neural network” was not adapted for the specific, claimed application because the specification failed to disclose how the artificial neural network was trained, and specifically failed to disclose weight values that resulted from such training. For this reason, the claimed “artificial neural network” could not be distinguished from the cited prior art, which resulted in failure to demonstrate the requisite inventive step.
As the Board described:
In the present case, the claimed neural network is therefore not adapted for the specific, claimed application. In the opinion of the Chamber, there is therefore only an unspecified adaptation of the weight values, which is in the nature of every artificial neural network. The board is therefore not convinced that the claimed effect will be achieved in the claimed method over the entire range claimed. This effect cannot, therefore, be taken into account in the assessment of inventive step in the sense of an improvement over the prior art.
Accordingly, at least with respect to patent applications filed in the EPO, and where an AI or machine learning model is to be distinguished from the prior art, then a patent applicant may want to include an example training data set, example trained weights, or at least sufficiently describe the input used to train the model on a specific, claimed application or end-use. For example, at least one example of data can be provided (or claimed) to show the inputs used to train specific weights, which may allow for the claim to have sufficient disclosure, and, at the same time allow the claim to cover a spectrum of AI models trained with a particular set of data.
For the time being, such disclosure for an EPO case could be considered as additional when compared with the sufficiency of disclosure in the U.S. However, it is to be understood that the U.S. Patent office has also indicated the importance of including training data or specific species of data used to train a model in its example guidance. See How to Patent an Artificial Intelligence (AI) Invention: Guidance from the U.S. Patent Office (USPTO). In any event, while there have been few court cases on AI-related inventions in the U.S. (see How the Courts treat Artificial Intelligence (AI) Patent Inventions: Through the Years since Alice), future cases may indicate whether the U.S. will trend towards the EPO’s decision in T0161/18 with respect to the sufficiency of disclosure.
Subscribe to get updates to this post or to receive future posts from PatentNext. Start a discussion or reach out to the author, Ryan Phelan, at firstname.lastname@example.org or 312-474-6607. Connect with or follow Ryan on LinkedIn.