The misunderstood classic: Äquivalenter Aortendruck – ARC Seibersdorf (T 161/18)

Decision T 161/18 Äquivalenter Aortendruck – ARC Seibersdorf of 12 May 2020 relates to application EP 1955228 A2 (priority date 08.11.2005).

The application claimed a method of determining cardiac output of a patient by measuring a peripheral blood pressure, transforming it into an aortic pressure, and inferring the cardiac output from that. In particular, it was claimed to transform an arterial blood pressure curve measured at the peripheral region into the equivalent aortic pressure using an ANN. The description suggested using a fully connected deep feedforward neural network (i.e. an fc-MLP), trained by backpropagation on input (peripheral pressures) and corresponding output data (aortic pressures) from a variety of patients. Figure 3 shows the structure of the neural network:

Claim 1 reads as follows:

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, characterized in that

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.

The novelty over the closest prior art (D1, WO 92/12669 A1) was the use of the ANN. A further document (D4, US 5339818 A) disclosed using a neural network for calculations of blood pressures, but not for determining an aortic pressure. The Examining Division refused the application nonetheless as obvious over the combination of D1 and D4.

The applicant appealed. In a communication ahead of the scheduled oral proceedings, the Board of Appeal introduced a further prior art document (D8, Qasem et al.: A neural network for estimation of aortic pressure from the radial artery pressure pulse. 2001 Conf. Proc. 23rd Annual Int. Conf. IEEE Eng. Med. Biol. Soc., Istanbul, vol. 1 p. 237.) that had not been cited in first instance. D8 explicitly related to using an ANN for the claimed purpose. The Board also raised a lack of sufficiency objection, which had not been raised in first instance. In response, the applicant withdraw the request for oral proceedings and did not file any further arguments, thereby practically abandoning the case. The Board then dismissed the appeal both on the grounds detailed in the communication.

Comment

Interestingly, the Board could simply have – with some justification – dismissed the appeal based on the conclusion that the invention is obvious over a combination of D1 and D8. However, it added some more comments about lack of sufficiency and lack of credibility that do not appear necessary for deciding the case. Moreover, none of these issues were actually discussed with the applicant's representative, which could have changed the Board's mind.

Concerning sufficiency, the Board started from the premise that the application should disclose (No 2.1) either

• which input data are suitable for training the ANN, or
• a suitable dataset for solving the problem (leaving open if this should be a training dataset or the weights).

In general, this is a reasonable premise because the skilled person needs this knowledge to train the ANN, which is in turn a prerequisite for applying it in the claimed method. Of course, only the first option is practicable since filing entire datasets with the application is neither feasible nor helpful for the EPO.

It is arguable that the Board was overly strict on the application. The reader may look himself on the first sentence on p. 6 of the application publication (WO 2007/053868 A2), which literally answers the question raised by the Board: The training dataset comprises pairs of peripheral and aortic blood pressure signals. Concerning the more specific question on which data formats are to be used, it should be noted that the application described at length what an fc-MLP is and how it works. At the priority date (late 2005), MLPs were long known (Rumelhart et al., Nature 323, 533, published in 1986) and subject to ongoing research (Bengio et al., J. ML Res. 3, 1137, published in 2003). Therefore, the level of detail in the description would have enabled the skilled person to put it into practice and provide the data in a format that can be processed by such a network.

Concerning inventive step, the Board considered the technical effect not to be credibly achieved because the application did not specify how the neural network takes into account the narrow-band character and the resonance phenomena in the low-frequency part of the transmission path between periphery and aorta (No. 3.5). Here, the Board seems to have defined an overly high standard on the description. One of the specifics of neural networks is that their inner workings are not obvious and not every intermediate value has a clear semantic meaning. Latent representations are generally not human-readable. Therefore, the Board should have either not raised the objection at all, or at least invited the applicant to prove the effect by post-filed evidence. This should be possible nowadays, see G 2/21, since the description stated that an accurate prediction is possible and therefore provided enough instruction to the skilled person to achieve the effect already at the priority date.

Although some observers warned that the decision could be the beginning of a much stricter approach in AI sufficiency (e.g. F. Hagel, EPI Information 4/2020, p. 22), this did not seem to have materialized.

In summary, T 161/18 is a typical T decision of the EPO Boards of Appeal: Aimed at deciding the case at hand, its reasoning becomes clear in the context of the underlying application and the filed requests and arguments.

One might note that parallel US proceedings lead to grant of US8920327B2. The US examiner had cited the same prior art documents D1 and D4 as the EPO’s Examining Division, but not document D8 of the EP appeal proceedings. Moreover, the USPTO did not see any enablement problems.