Are applications of AI to healthcare patentable?

Once the pitfalls are avoided, this is a healthy area for obtaining patent protection. Our experts James Varley and Tom Leanse point the way:

Artificial intelligence (AI) is increasingly being used across society for a wide range of applications, from voice and image recognition to predicting our shopping habits. Healthcare has become an area of particular focus and activity, with AI being used to search for new treatments, and to assist in the diagnoses of patients. From the perspective of European patent law, the application of AI to healthcare is particularly interesting because, at least at first glance, it seems to relate to several different types of patent exclusions: the exclusions for mathematical methods, computer programs and methods of treatment and diagnoses performed on a human body.

Article 52(2) of the European Patent Convention (EPC) lists a number of subjects that are not regarded as inventions by the European Patent Office (EPO). Among them are programs for computers and mathematical methods. On the face of it, AI would fall into both of these categories. However, Article 52(3) of the EPC goes on to state that applications related to these subjects are only excluded “to the extent to which a European patent application or European patent relates to such subject-matter or activities as such”.

In practice, the EPO considers both mathematical methods and computer programs to be patentable where they “contribute to the technical character of an invention, i.e. contribute to producing a technical effect that serves a technical purpose”. The most recent version of the EPO Guidelines for Examination provides a non-exhaustive list of technical purposes that may be served by a mathematical method (and by extension, a computer program). Among these are “providing a medical diagnosis by an automated system processing physiological measurements”.

As an example, consider the use of AI in the paper “Artificial Intelligence Clinician learns optimal treatment strategies for sepsis in intensive care” (Nat Med. 2018 Nov; 24(11):1716-1720). Here, a reinforcement learning algorithm is applied to 48 different bits of patient data in order to identify sepsis treatments that would maximise the patient’s 90-day survival chances. The results showed that mortality rates were lowest among patients whose clinician has followed the recommendation of the algorithm.

Comparing this method to the technical purpose provided in the EPO guidelines suggests that a claim to such a method should not be excluded as a mathematical method/computer program. The reinforcement learning algorithm can be considered an “automated system”, as it automatically processes the input patient data, and the 48 bits of patient data used are said to include vital signs and lab test results, which can be considered “physiological measurements”. While identifying an appropriate treatment is not exactly the same as “providing a medical diagnosis”, it is almost completely analogous, so would likely be considered to meet this requirement.

Another example to consider is the AI system for identifying factures in bone scans described in “Deep neural network improves fracture detection by clinicians” (PNAS November 6, 2018 115 (45) 11591-11596). In this paper, a convolutional neural network (i.e. an automated system) was trained to identify factures (i.e. provide a medical diagnoses) in X-ray images (i.e. physiological measurements), and showed an accuracy comparable to that of a specialised orthopaedic surgeon. Under the current EPO guidelines, such a method would be considered technical. It also has the added advantage that it falls into another of the categories of mathematical methods/computer programs that the EPO has long considered to be technical – namely “digital audio, image or video enhancement or analysis”.

There is, however, a further complication. Being considered technical is not enough to avoid a patentability exclusion. Article 53 of the EPC also contains a list of subject matter that is excluded from patentability, whether or not is considered technical. Among this list of exclusions is “methods for treatment of the human or animal body by surgery or therapy and diagnostic methods practised on the human or animal body”. This exception would appear at first glance to be damaging to any prospect of patenting the use of AI to assist diagnosing patients.

However the EPO Guidelines for Examination once again provide some clarity on the extent of this exclusion. The Guidelines state that a claim is only considered to fall within this exclusion if it contains all four of the following elements: (i) an examination phase, involving the collecting of data; (ii) a comparison phase, comparing the data with standard values; (iii) the finding of a significant deviation during the comparison; and (iv) a decision phase, attributing of the deviation to a particular cause. Additionally, the technical parts of the method must be “practised on the human or animal body” in order to fall within this exclusion, by which it is meant that an interaction with a human or animal body takes place.

The fact that each of a sequence of four steps are required in order for subject matter to be considered a diagnostic method means that careful drafting can often avoid the exclusion of Article 53(c). For example, a claim directed to a method of data acquisition or data processing that could be used in a diagnostic method would not be excluded. However care must be taken when drafting to ensure that a non-excluded claim can be formulated without adding subject matter contrary to Article 123(2) EPC, and without removing the novel and inventive features of the invention.  

In conclusion, the prospect of achieving patent protection for the use of artificial intelligence in healthcare applications is not as bleak as it may first appear. By taking the above considerations into account when drafting an application to the use of AI in diagnostic methods, the exclusions of Article 52(2) and 53 of the EPC can be avoided.