Chapter excerpt from the ebook MedTech IP: Lessons and Strategies for Success - view all chapters here.
For the last four years Marks & Clerk has produced an annual AI Report, providing data and insights from our many experts in AI. The following graph, showing the number of AI patent applications in MedTech published by the EPO each year, is taken from our 2023 AI Report.
As can be seen from the graph, the number of publications of European patent applications for AI inventions in the MedTech sector has increased virtually exponentially over the last decade.
Let's look at a decision of the EPO relating to a MedTech AI invention, which has helped to drive the trend shown above...
Teuvo Kohonen was born in Finland in 1934. Growing up in Finland he spent time in the Scouts, where he would have learnt map reading, and maps were to form an important part of his career. He was fortunate to have a good teacher in physics, which led to him doing a PhD on the lifetimes of positrons, and starting life as a physicist.
Teuvo became a professor in physics in Finland but, as there were not enough university academics at the time, he also had to teach students about computers. He had never been taught this subject, and so a few weeks before each lecture he hastily read up about computers. In 1962 he came across an article on computer learning which interested him greatly, and so began his interest in artificial intelligence.
Teuvo Kohonen was to become one of the world's best-known neural network researchers. He made pivotal contributions in the field of artificial neural networks, and is best known for developing the Kohonen map, which brought Finnish artificial intelligence research onto the world stage in the early 1980s.
A Kohonen map, also known as a Self-Organizing Map (SOM), is a machine learning technique which uses a (usually) two-dimensional representation of a higher dimensional data set while preserving the topological structure of the data. Kohonen maps have been used in finance, trade, natural sciences, linguistics, speech recognition and robotics. The Kohonen map was considered by experts to be one of the most significant inventions in computational science, and has been the subject of more than 8,000 scientific papers.
On 3 September 1996 a UK company, Cardionics Limited, filed International Patent Application No. PCT/GB96/02169 for a heartbeat monitoring apparatus and method using a Kohonen neural network.
The invention related to the analysis of electrocardiograph signals obtained from a patient using a neural network to monitor changes in the functioning or performance of the heart of a patient.
Prior to the invention it was known to detect electrical signals of the heart by means of conductive pads attached to the patient's chest and directly wired to a suitable machine which provided a graphical trace of the waveform for analysis by a doctor.
The invention sought to provide a heart monitoring apparatus which could monitor changes in heart condition automatically.
Figure 1 of the patent application, shown below, illustrates a typical electrocardiograph trace wherein various features P, Q, R, S and T can be seen. The shape and size of each of the features is an important indication of the condition and operation of the heart.
Instead of using a fixed apparatus to display such traces, the invention envisaged that the patient would wear a portable heart monitor device, and that this would be connected wirelessly to a remote base station, thus allowing the patient's heart to be monitored whenever the patient wore the portable monitor device.
Analysis of the electrocardiograph signal could take place either in the portable monitor device itself, or in the base station or another computer.
Feature mapping in a Kohonen feature map is a process in which example training vectors are clustered in feature space.
In the invention, pre-processed values from an electrocardiograph signal were used to define a vector position in a multi-dimensional feature space. The dimensionality of the feature space is determined by the number of features measured. In some embodiments of the invention 64 values were provided, and the feature space was therefore a 64 dimensional space. In simple terms, in a Kohonen map these vectors can be represented in a 2D pattern space.
Figure 10 from the patent application is shown below, and illustrates a simple case of a two dimensional pattern space having eight reference vectors (shown as small circles), each with their own area of influence.
Figure 10 also illustrates two electrocardiograph values (shown using the letter "x") for this two dimensional hypothetical example: one representing a normal electrocardiograph signal which falls within the threshold of normality and one representing a "novel" (i.e. unknown) electrocardiograph signal which falls outside the threshold of normality.
According to the invention, this approach could be used to define different regions in the 2D pattern space representing different specific heart conditions. During the monitoring phase, when the patient is wearing the portable heart monitor, if the electrocardiograph signal values fall outside the threshold of normality and inside a region of abnormality, the apparatus can detect and indicate the specific heart condition which has arisen.
For example, a region of the 2D pattern space may represent a myocardial infarction, commonly known as a heart attack, caused by reduced blood flow to the muscular tissue (myocardium) of the heart. Myocardial infarctions can sometimes be mild, and can therefore sometimes go undetected in the absence of heartbeat monitoring.
The European patent application granted on 15 December 1999 as European Patent No. EP 0 850 016 B1. However, European patents can be opposed by third parties within 9 months of grant, and the patent was opposed on the last day of the 9 month period. Unfortunately for the patentee, the opposition division revoked the patent on the ground that the invention was obvious. The patentee appealed.
After two rounds of oral proceedings the appeal board finally gave its decision (T 0598/07) on 19 May 2010, some 10 years after the original opposition, and nearly 14 years after the application's filing date.
Fortunately, these days appeals are much quicker at the EPO, which has an objective to settle 90% of cases within 30 months.
The appeal board noted that the independent claims had now been limited to a Kohonen neural network, and more specifically recited that:
..... the n-dimensional vectors representative of the monitored ECG are compared with a first n dimensional volume representative of irregular heartbeats, which are spurious with regard to monitoring heart conditions, and subsequently with a second n dimensional volume representative of regular heartbeats.
The appeal board had to consider what prior art documents indicated to be already known at the priority date of the application. There were two documents which were particularly relevant.
Document 1 disclosed a heart monitoring apparatus with input means for receiving an electrocardiograph signal from a patient during a monitoring phase. Pre-processing means could carry out feature extraction in order to extract important features of the shape of a sequence of pulses of the electrocardiograph signal to obtain a plurality n of values representative of the shape of said sequence of pulses of the electrocardiograph signal.
However, the invention differed from Document 1 in that the plurality n of values obtained in the invention was representative of the shape of each pulse of the electrocardiogram, and not of a sequence of pulses as was the case in Document 1. A further difference was that the invention used Kohonen networks.
Document 2 also disclosed a heart monitoring apparatus, which could carry out feature extraction to extract important features of the shape of each pulse of the electrocardiograph signal, and which could store a plurality of four-dimensional reference vectors, each said four-dimensional reference vector comprising four
values representative of the shape of a reference pulse.
At this point you might be thinking that all was lost, as Document 2 is quite close to the invention. However, the board noted the following two important features of the invention which were not present in Document 2:
a) The n dimensional vector representative of each pulse was first compared with an irregular heartbeat n dimensional volume, to identify distinctive irregular beats, and subsequently with a regular heartbeat n dimensional volume; and
b) The data processing was carried out by a Kohonen neural network.
The board noted that the technical effect (advantage) obtained by feature (a) was to allow the second comparison with the regular heartbeat n dimensional volume to be carried out only for n dimensional vectors formed from a regular heartbeat which did not include a distinctive irregular heartbeat.
This allowed the invention to improve the signal to noise ratio and thereby reduce the number of false identifications of novel electrocardiograph signals,
The board next had to decide whether the invention was excluded from patentability by our old friend Article 53(c), which you will recall states that states that European patents shall not be granted in respect of:
(c) methods for treatment of the human or animal body by surgery or therapy and diagnostic methods practised on the human or animal body ...
One of the independent method claims included the step of "outputting an indication" (e.g. an alarm) if the heartbeat was abnormal. Although certain examples in the patent specification allowed the system to provide an audible or visual indication of specific heart conditions, such as the myocardial infarction mentioned above, this feature was not included in the claims.
Therefore, the claims did not include a step relating to diagnosis for curative purposes stricto sensu representing the deductive medical decision phase. The diagnosis would instead be carried out by a physician after the alarm had been sounded.
For these reasons the patent was allowed.
From this example it can be seen that a heartbeat monitoring method using a neural network to identify irregular heartbeats was considered patentable by the EPO, and was found to make a technical contribution.
