Addressing bias in AI healthcare models — could it be patentable?

A recent article in The Lancet highlights an important challenge facing developers of AI-powered healthcare solutions: so-called “algorithmic bias”. AI models trained on datasets that do not adequately reflect the subset of patients receiving care based on those models risk contributing to unequal health outcomes for the underrepresented demographics.

The most direct approach to alleviate such bias is to ensure that training of the AI models is performed based on balanced datasets including sample cases from a wide variety of ages and backgrounds.  

From a UK and European patent perspective, however, applications directed to selecting appropriate candidates for a health study (or, equivalently, for training a model), are often viewed as addressing an administrative problem rather than a technical problem. This can make surmounting the “inventive step” hurdle to patentability challenging, though other jurisdictions such as the US are generally more lenient in this field. In any case, a more balanced training dataset may not be readily available or practical to obtain.  

In such cases, a different approach would be to adjust some aspect of the underlying AI model, or the way it is trained, to offset some bias arising from skewed training data. Here, care should be taken not to overlook an innovation that could be patent-eligible. Patent applications to such methods can certainly stand a reasonable chance of achieving grant in Europe, especially when supported by the right data and framed optimally for success.

Talk to your usual Marks & Clerk attorney or feel free to get in touch if you think any of these issues could be relevant to your technology.