Modern medical research is increasingly data intensive with every clinical study, laboratory experiment, diagnostic test, imaging procedure and patient interaction generating more information. Yet whilst individual patient data is unique, it is how it contributes to a larger population study that creates a corpus of knowledge that contributes to scientific discovery and medical innovation.
Why AI Is Increasing the Strategic Value of Medical Data
Collating relevant information from a number of different medical studies to form a larger data set was historically, and often still is when performed manually, a time-intensive task. Today, digitalised health records are transforming how data sets are created allowing curated data sets to be generated far faster than ever before to satisfy the thirst for medical data. Patient data is sensitive and may need processing to ensure a suitable degree of anonymisation of the patients’ personal information before it can be digitally shared.
In addition, as different techniques may generate the same type of information, understanding how this may affect the results of different studies, and processing the data to ensure normalisation across different studies is also usually needed. Whilst deterministic artificial intelligence (“AI”) systems have been used for many years to analyse medical data, more and more generative AI systems are being developed for a growing range of medical uses and applications, and their image and text capabilities are dramatically reducing the time-spans over which medical insights can be obtained.
Two examples of medical AI innovation using data-sets are the Delph-2M model which was trained on multiple diverse health records to understand patients’ past and current health states. The Delph-2M is a very small model of just 2.2 million parameters and training was primarily on the UK biobank dataset. Another LLM, Geneformer was trained on 30 million single-cell transcriptomes gathered from a large collection of publicly available single-cell RNA sequencing profiles to learn patterns of gene regulatory networks and cellular states. Geneformer used multiple diverse sources of healthcare data contributed by many different research groups world-wide, which exposed it to a huge range of cell states, tissues and disease contexts.
All this means is that it may be quite important to think of future uses when generating data as part of a medical research project and a big strategic question for many medical research projects is whether they should try to monetise their medical data/results for some uses or not at some point.
In the private sector, medical data sets and the activities and services that can be built on them can be incredibly valuable and academics who successfully licence their medical data (either with or without any other types of licensing) could generate significant income to fund future research activities. Other areas to consider for IP asset generation include patents, trade secrets, and copyright.
Whilst patents may provide useful rights protecting the techniques used to capture medical data and sometimes the techniques used to gain insights from raw medical data, it is becoming increasingly important to protect the raw data itself and also the related information that can be derived from raw data. Modern medical research generates extraordinary volumes of information from many different sources, including:
Historically, much of this information was simply stored and analysed using conventional statistical methods, but as AI technology has developed and is becoming ever more reliable, the analysis it can perform is fundamentally changing that landscape. Machine learning models can identify patterns that would be impossible for human researchers to detect, uncover previously unknown biomarkers, improve diagnostic accuracy, predict disease progression, optimise clinical trial design and accelerate drug discovery.
Private (still confidential) studies historically performed may be monetised as shared secrets and, as AI capabilities continue to develop, the underlying datasets used in historical studies may become significantly more valuable than when they were originally created. This all means a rounded IP strategy should accordingly address both near and long term data value and cover far more than just patenting an algorithm, process, or system.
Why Keep a Human in the Loop?
For research teams using AI in their research, the role of AI may also create issues when it comes to IP as AI systems cannot be named as inventors, and so careful consideration may be necessary to ensure that a human can be identified as having contributed the inventive concept being claimed in a patent. What this means in practice may differ case by case. For example, a human should ideally:
This leave the AI to perform detection of patterns, literature synthesis. However, where the AI suggests a candidate variable or specific compound, some additional care must be taken. A good practice which is strongly recommended is to maintain an invention log which shows at least: what input/prompt was given to the AI system, what was the raw output the AI produced, how the AI output was then subsequently processed/used, what criteria were applied by any human who reviewed the output and why they accepted/rejected/changed the output, and how the human conceived the final idea/invention from the raw AI output.
Modern medical research projects are often fertile IP landscapes and areas of innovation relating to using AI tools which could be generating IP assets such as data, trade secret and patent rights, might, to name just a few examples, include:
collecting medical data
cleaning and validating datasets
integrating heterogeneous data sources
AI-assisted searching and retrieval techniques
model training
explainable AI techniques
privacy-preserving computation
federated learning
By seeking to capture various different forms of IP a more complex moat can be created helping to defend the value of the IP being created around the core research concepts. In this context, medical data rights often require careful consideration of ownership/licensing potential as different legal rights may exist in relation to:
Ownership of these different types of data assets often does not rest with the same organisation and personally identifiable data and anonymisation of data add further layers of complexity. It is often possible for clinical collaborators, hospitals, universities, commercial partners, software providers and data processors to each own or control different aspects of the overall data ecosystem. Careful contractual drafting of any agreements from the outset may therefore be essential to ensure that any IP and data created by a research project has clear ownership for licensing downstream.
Building an integrated IP strategy for Medical AI technology
To recap, medical AI continues to evolve rapidly in terms of its raw analytical capabilities and as the technical capabilities of the AI analysis systems increase, so too does the value of the data that underpins it. Protecting that value requires a holistic approach spanning multiple forms of IP rights.
Timing is Important
The organisations that derive the greatest long-term value from medical innovation are often those that develop an integrated intellectual property strategy at an early stage. For modern research projects, any IP strategy should recognise data not simply as an operational resource, but as a strategic asset capable of generating competitive advantage and long-term commercial value. If you are actively developing or planning to develop technology related to AI-assisted diagnostics, digital health platforms, medical devices, drug discovery technologies or clinical decision support systems, a well-designed IP strategy may make all the different when it comes to giving you the control you want – whether that is sharing your data freely or putting in place a foundation for downstream licensing.
If you are interested in having a conversation around AI in this area, please reach out to Dr Coreena Brinck, or your usual Dehns contact in the first instance.