A recent decision from the EPO’s Technical Boards of Appeal (T2803/18) highlights a pitfall for applicants wishing to patent systems and methods that utilise machine learning algorithms.
The claimed invention in this opposition appeal relates to the monitoring and processing of sensor signals to automatically determine the characteristics of wetness events in moisture absorbing articles such as diapers, incontinence garments, dressings, etc. The invention seeks to automatically provide information about the wetness event to avoid the need for manual checking.
This inventive concept was encapsulated within a claim that recited receiving sensor signals representing the wetness event in an absorbent article, and the processing of those sensor signals. The claim was characterised in the mathematical steps that define that processing in order to arrive at an indication of volume of exudate in that wetness event.
When the available prior art was considered, a single novel feature was identified, distinguishing the claim from the prior art disclosure. That single feature involved, firstly, a comparison of a weighted representative vector representing the wetness event with clusters of weighted representative vectors to determine which one or more of the clusters the weighted representative vector is most similar to, and then, secondly, using an optimal mathematical model derived in a learning step to allocate a volume of exudate to the weighted representative vector.
The Board of Appeal doubted that this novel feature provided for any technical effect going beyond the normal operation of a processor performing the processing of the sensor signals.
The applicant had put forward the argument that the novel feature would provide an increased accuracy of the estimation of the characteristic (the indication of volume of exudate in the wetness event). However, the Board considered that this could not be guaranteed, noting that:
“the accuracy would depend on many factors (size of training sets, number and type of elements/variables constituting the representative vectors, etc.), none of which are defined in claim 1, so that the results obtained by the claim method are not necessarily more accurate than the results obtained by the regression analysis, the resulting mathematical model and the threshold criteria applied in D2”
That is, although the use of the comparison of the weighted representative vector with the clusters of weighted representative vectors, along with the optimal mathematical model obtained during a learning phase, could plausibly have resulted in a technical effect over the prior art, the level of generality with which it was expressed in the claim failed to bring out the technical effect and so deprived the claim of an inventive step over the prior art.
The take-home message from this case is that, in some circumstances, it may be insufficient to merely distinguish a machine learning method or other mathematical method from the prior art in a generic sense in a claim. Such methods, while able to perform at greater accuracy (or achieve other advantages), might only achieve those advantages in a limited set of circumstances. In line with the general approach of the EPO to inventive step, an advantage must be achieved across the full breadth of the claim.
As such, it is important when drafting a patent specification to bring out the specifics of those machine learning methods or mathematical methods that achieve any improvements in performance.
Our AI group has extensive experience with machine learning inventions. We can provide advice on patentability and assist in obtaining protection in Europe and across the world. Please contact your usual Boult advisor for more details.