Inventive step at the EPO for AI inventions – T 2803/18

Some evidence from the EPO’s Boards of Appeal suggests that establishing that an AI method contributes to the technical character of a claimed invention may be trickier than for other, more ‘deterministic’ or ‘predictable’ mathematical methods.

 

T 2803/18 is a Board of Appeal decision that illustrates this difficulty. The application in this case related to monitoring moisture in absorbent articles such as diapers, incontinence garments, dressings and pads resulting from ‘wetness events’ caused by, for instance, urinary incontinence. In particular, the application related to monitoring sensor signals representing such wetness events in absorbent articles, and processing the sensor signals to determine a characteristic, e.g. volume of wetness, of a wetness event on an absorbent article. Specifically, the application discloses performing a learning phase to determine a relationship between received sensor signal data and observation data indicative of an observed cumulative characteristic of wetness events. The learning phase included determining an optimal mathematical model for describing the relationship between the received sensor signal data and the observation data.

 

Claim 1 of the sole request under appeal recited a method for processing sensor signals representing a wetness event in an absorbent article. In particular, Claim 1 recited steps of the learning phase and further recited steps of an execution phase for automatically identifying a wetness event from sensor signals, comprising:

(i) generating a representative vector for an individual wetness event including values derived from a received sensor signal for that event;

(ii) normalising the elements of the representative vector with respect to a reference range;

(iii) allocating weightings to the normalised elements of the representative vector to generate a weighted representative vector for the wetness event; and,

(iv) comparing the weighted representative vector with clusters of weighted representative vectors to determine which of the clusters the weighted representative vector is most similar to, and allocating a wetness event characteristic of the determined cluster based on an identified optimal mathematical model obtained from the learning phase to the weighted representative vector for the wetness event, where the event characteristic indicates volume of exudate in the wetness event.

 

The Board of Appeal decided that the only novel feature of Claim 1 over the closest prior art document, D2, was feature (iv). In determining the objective technical problem, the Board noted that Claim 1 recited a mixture of technical and non-technical features, the non-technical features relating essentially to mathematical method steps and including feature (iv).

 

The Proprietor had argued that Claim 1 – including feature (iv) – allows for estimating the volume of collected exudate from a wetness event with increased accuracy. However, the Board was not convinced that the accuracy of the estimation would be increased. In particular, the Board stated: “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 claimed method are not necessarily more accurate than the results obtained the regression analysis and the threshold criteria applied in D2. The patent in suit also does not support such an alleged benefit by comparative data.”

 

As the alleged technical effect of increased estimation accuracy was not acknowledged by the Board, then this could not be used as the basis for arguing the presence of an inventive step over the closest prior art. The Board did find that the claimed invention had a technical effect of providing an estimation (which may be more or less accurate than the prior art method) of the volume of exudate collected in an absorbent product during an individual wetness event in a plurality of wetness events. However, the claimed invention was ultimately found to lack an inventive step when the objective technical problem was formulated using this acknowledged technical effect.

 

Returning to the increased estimation accuracy point, even if this had been acknowledged by the Board as being a technical effect of the claimed invention, it is unclear whether this would have been sufficient for the Board to conclude that the claimed invention was inventive over the cited prior art.

 

However, the takeaway is that, for certain types of mathematical methods, a greater level of detail may need to be provided in the specification for a technical effect to be acknowledged (such that an assessment of inventive step can be based thereon).

 

For more ‘deterministic’ or ‘predictable’ mathematical methods, merely stating the presence of an associated technical effect may be enough for it to be acknowledged, as it can more readily be envisaged across the entire scope of a claim. On the other hand, for more ‘probabilistic’ or ‘unpredictable’ mathematical methods – including AI-based methods – additional details of under which specific circumstances an alleged technical effect is achieved, and/or specific examples in which the technical effect is achieved, may need to be provided for the technical effect to be acknowledged. This could be specific sets of model parameter values and/or specific sets of training data used to train the model, or specific characteristics of a training set needed to train the model, for instance. As a technical effect or advantage associated with a claimed invention must be achieved across the entire scope of a claim, then evidencing exactly how to achieve the advantage and when it is achieved can assist in convincing the EPO to acknowledge that an alleged technical effect is indeed provided by a claimed invention.