Crafting a claim for an AI: Two case studies
As artificial intelligence creates a complex legal landscape for healthcare and medtech, Janine Swarbrick of HSF explores two case studies that shed light on how to claim an AI invention.
Artificial intelligence (AI) is playing an increasingly significant role in the rapidly evolving fields of healthcare, medical technology and biotechnology.
From identifying new drug molecules, diagnosing diseases, advancing bioscience, and predicting patient outcomes, AI is revolutionising the way medical professionals approach patient care and obtain insights into medical conditions, treatments and genetics. However, with this innovation comes a complex legal landscape, particularly in relation to patent protection for AI-focused inventions in the life sciences sector.
A key consideration in software patenting is to demonstrate there is a clear technical character to the claimed invention, rather than the claim being understood as a mathematical method, mental act or computer automation of a manual task. Another requirement for AI inventions is that there must be sufficient details of the new technique in the patent specification.
Simply reciting that there is a) some input, b) a black box running an AI algorithm, then c) useful output produced, is unlikely to meet the requirements (even if AI has not been used for that purpose before). These considerations of AI software claims also apply to AI-based healthcare, in which processing health and medical data using AI modelling to provide a medical indication or indication of treatment, for example, is a rapidly growing area of innovation.
Recently, I received an examination report for a European patent application drafted by another party, and the subject matter of the claims related to a machine learning model operating on medical data to produce a medical indication. Without giving the game away as to their overall opinion of patentability, the examiner helpfully noted two European Patent Office (EPO) Board of Appeal cases relating to using software methods in the biological / medical field as examples of when the claim does not include sufficient detail to be considered technical, and when the claim does provide sufficient detail to be found technical.
These cases, which are summarised below, are great examples of how, and how not, to craft a claim for an AI-based invention, and provide useful direction in how much detail is expected in software innovations when applied to healthcare/medtech.
While the focus of the data processing in these cases is in statistical analysis, AI methods are examined similarly to other computer implemented mathematical methods, so the considerations of technical character in these statistical processing examples also apply more broadly to other AI data processing methods.
Not technical: T 0784/06 Beckman Coulter —Automatic genotype determination
The claims related to a method and device for determining a genotype in genetic material. Originally, the sixth auxiliary request was allowed during opposition proceedings, but this decision was appealed by the opponent who succeeded in having the patent revoked for a lack of inventive step—in this case through a lack of technical character.
The method claim recited the following:
A method of determining the genotype at a locus within genetic material obtained from a biological sample, the method comprising:
A. reacting the material at the locus to produce a first reaction value indicative of the presence of a given allele at the locus;
B. forming a data set including the first reaction value;
C. establishing a distribution set of probability distributions, including at least one distribution, associating hypothetical reaction values with corresponding probabilities for each genotype of interest at the locus;
D. applying the first reaction value to each pertinent probability distribution to determine a measure of the conditional probability of each genotype of interest at the locus, and
E. determining the genotype based on the data obtained from step (D).
The applicant argued that the claimed method involved technical steps including in the data treatment phase set out in steps B to E. The arguments focussed on the claimed method providing a way of determining something real about physical, genetic material, and that the probabilistic approach particularly in step C allowed for automatic scoring of the obtained data to get an improved result compared with the deterministic approach taken in the prior art. The claimed steps were argued by the applicant not to be an excluded mental act but a tool for processing real world data representing a physical entity, per decision T208/84.
However, the board concluded that the data treatment steps B to E were a mental activity which did not interact with the technical features A of the method to produce a tangible technical effect, and so those data treatment steps were ignored in the assessment of inventive step.
Step A was found to clearly be a technical act involving carrying out a reaction step. However, steps B to E were considered to relate to data processing steps using the “reaction value” obtained from step A to determine the genotype. The board’s opinion was that these data processing steps, of forming a data set, determining probability distributions and then conditional probabilities to obtain the genotype, were not technical because they could be performed as a mental act (a mathematical method), and did not combine with the technical features of step A to be considered in the consideration of the presence of an inventive step.
That is, steps B to E were understood by the board to recite a statistical method to obtain a probability, which is a purely mathematical process (which was also a known process, even if it was not known to perform the method on the data obtained in step A) and therefore could not be considered in assessing inventive step.
This case illustrates how it is important to be able to convincingly argue that an invention which uses mathematical analysis of input data is technical (rather than a mental act). Only those steps which are technical, or which are non-technical (eg, mathematical processing) but which clearly combine with technical steps, can be considered in determining whether a claimed invention is inventive. A contributing factor in this case was a lack of specific detail about how the mathematical processing acted on the input data to obtain a meaningful and concrete technical output rather than a statistical indication which the board could not conclude was a meaningful technical output.
Often, it is helpful for arguments in support of technical character if the claim features provide a real world effect, which in the field of digital health may be a clinical indication or output providing some medical guidance, and in the life sciences may be an identified genotype obtained by the claimed steps, for example. Ensuring that there is sufficient technical detail present in the claims, supported by the description, to argue for the determination of a meaningful real world parameter is important to be able to present a convincing case for technical character.
This case illustrates that the EPO may not be convinced that determining a parameter (a genotype here) through mathematical analysis, even of data obtained through a technical process, is sufficiently technical alone to be considered in assessing patentability (inventive step) without specific implementation details of the processing taking place to support the determination of a real world and physically meaningful output.
Technical—T 2050/07 Perlin—Method and system for DNA mixture analysis
In this case, the claimed invention was initially refused for lack of novelty but was remitted back to the examining division for further prosecution by the board. After submission of auxiliary claim sets, the claims which were admitted into proceedings for further prosecution were considered by the board to be technical and therefore not excluded from patentability on the grounds of being non-technical. The application was granted in 2013.
The claims recite a method of DNA mixture analysis as follows:
A method of analysing a DNA sample that contains genetic material from at least two individuals to determine a probability distribution of genotype likelihood or weight in the sample, comprising the steps:
(a) amplifying the DNA sample to produce an amplification product comprising DNA fragments, wherein each allele at a locus is amplified to generate relative amounts of DNA fragments of the alleles that are proportional to the relative amounts of template DNA from the alleles in the DNA sample, and wherein the amplification product produces a signal comprising signal peaks from each allele the amounts of which are proportional to the relative amounts of the alleles;
(b) detecting signal peak amounts in the signal and quantifying the amounts using quantifying means that include a computing device with memory to produce DNA length and concentration estimates from the sample;
(c) resolving the estimates into one or more component genotypes using automated resolving means, said resolution into one or more genotypes including solving the coupled linear equations d = G.w+e for the relevant loci (i), individuals (j) and alleles (k), in which d is a column vector which describes the peak quantitation data of a DNA sample from the signal, G is a matrix that represents the genotypes in the DNA sample, with a column j giving the alleles for individual j, w is a weight column vector that represents relative proportions of template DNA in the sample and e is an error vector, wherein the solution includes calculation of data variance sigma**(2) from the linear model d = G.w+e together with the global minimal solution Pd = Gw0, where Pd is the perpendicular projection point which is the closest point to d in mixture space C(G) and w0 is the minimum weight vector, using linear regression methods, and calculating a probability distribution of the data assuming a normal distribution and that the error is unbiased, so that E(e) = 0, but has a dispersion D[e] = sigma2V in which V is the covariance matrix of the data; and
(d) determining, using the probability distribution of the data, a probability distribution of genotype likelihood or weight in the DNA sample.
The board questioned whether the claims related to a mathematical method, and as such whether they were excluded from patentability. Their decision was that the method related to analysing a DNA sample, including a step of DNA amplification in which the amplification product comprises relative amounts of DNA fragments of the alleles, proportional to the relative amounts of template DNA from the alleles in the sample, and these amplification products produced a signal comprising signal peaks from each allele proportional to the amount of each allele, detecting peak amounts, and quantifying those amounts using a computational method to produce an estimation of genotype. Steps (a) and (b) are carried out using laboratory equipment and are therefore technical.
Regarding the mathematical processing steps, the board considered that they contributed the feature of ascertaining the reliability of the method for analysing DNA samples for determining genotype. By providing the estimated error, estimates of the variance and standard deviation can be computed and used to estimate probabilities, giving a quantitative estimate of the solution quality. Thus the distinguishing features provide a method of improving the confidence of the genotype estimate, thereby contributing to the technical character of the invention.
The board referred back to the first decision discussed above, T784/23, and found that its disclosure was of too general a nature to provide the skilled person with the information needed to proceed and obtain the probabilistic basis of the genotype in step E to provide a tangible technical result, and so were ignored in analysing inventive step.
The description of the present case was considered to be sufficiently detailedin how to perform the method and how the data processing contributions interacted with the initial technical laboratory steps to provide a technical result of a genotype estimate with improved confidence compared to the prior art.
Comparison of the two cases
Overall it can be seen that what helped demonstrate technical character in T 2050/07, and what caused issues in T 0784/06, was the level of detail of the mathematical processing performed on the input data, which has a knock-on effect for the consideration of the technical character of the outputs obtained.
Without sufficient detail of the mathematical processing being included, in the claims and the description, it is challenging to demonstrate that the output provided is a concrete technical result which can clearly be likened to a real world parameter, and thus that the claimed invention provides a technical effect.
Vague reference to statistical methods, particularly known methods, even if used in a new context, are at risk of being considered to relate to mathematical processing of input data without providing a real world technical outcome, and thus to be seen as non-technical and not contribute to an inventive step in Europe.
Training machine learning models
Recent updates to the EPO guidelines for examination, published on April 1 2024, set out in section G-II-3.3.1, provide useful guidance on the level of detailed disclosure required in the patent specification in relation to training data to fulfil the legal requirement that sufficient detail must be provided to enable the claimed invention to be reproduced by a person skilled in the art.
The new guidelines state that the necessary features of the training data set must be disclosed—unless they are common general knowledge—to show how the claimed machine learning model is trained if those features of the training data contribute to the technical output provided by the model. This is so that someone following the teaching in the patent application can arrive at the same or comparable result.
Specific disclosure of the training dataset itself is not mandatory but “particular characteristics of the training dataset used” which contribute to the technical effect of the invention are referred to, to enable reproduction of the invention. For example, statements of the nature of the training data (such as the resolution of input images, number of training data sets, and/or types of demographics of personal data) may meet the EPO’s requirements.
Useful checks to make are, if you could not readily determine enough detail from the description to be able to implement the invention and get a similar outcome to that claimed, or you could not credibly argue for a technical effect based on the claim features, this indicates that more details may be needed to properly define the invention.
Conclusion
In conclusion, the EPO needs to be confident that there are sufficient technical details present in the claims and description—both in the claimed processing steps and any training data used to train a machine learning model—to demonstrate that an obtained output relates to a concrete real world output, and that the invention can be reproduced to obtain the same output by a person skilled in the art. These issues should be considered carefully at the patent drafting stage.
Janine Swarbrick is a patent director at HSF.
