Artificial intelligence (AI) has the potential to advance medical diagnosis and the overall provision of healthcare. As part of this evolution, the number of medical devices incorporating machine learning (ML) functionality continues to increase. The FDA has already reviewed and authorized over 690 AI/ML-enabled medical devices1. Similarly, the number of medical AI-related patents has seen a significant surge2. Thus, the emergent patent landscape for medical AI inventions remains an area of considerable interest and complexity. The underlying legal intricacies coupled with the rapid technological advancements necessitate a deeper exploration, particularly in areas that have previously been identified as open questions, such as the quality of these patents2,3. Legal regulation of this technology will play a big part in its future, including patent law which regulates economic rewards for innovation.

In our preceding study2, we undertook a comprehensive review of medical AI patenting trends. One salient finding was the recent increase in medical AI patenting activity, despite the doctrinal uncertainties surrounding AI patentability. Such growth raises questions about the robustness and quality of these patents, especially the sufficiency of their disclosures.

Ensuring that patent specifications sufficiently describe and support the claimed invention is a cornerstone of patent law4 (Box 1). While there are differences in the exact wording of this requirement, patent offices across the world generally require that the patent specification provides enough detail to teach a person skilled in the relevant art to make and use the claimed invention without undue experimentation5,6. Given the complex and more abstract nature of AI/ML technologies, ensuring adequate disclosure is particularly challenging. A pressing question arises: Do the patent specifications of medical AI inventions meet this standard?

Computer and software-related patents, including AI-enabled inventions, have been criticized in the literature for inadequate teaching, and overly broad claims which extend beyond their teaching7,8. Some scholars even suggest that disclosure of black-box machine learning is nearly impossible9. Conversely, other authors argue that disclosure of AI inventions is being effectively managed10. Diverging opinions persist, given the limited empirical data available. Better data would help assess whether AI/ML technologies uphold the quid pro quo underlying the patent system. Widespread disclosure shortcomings could lead to extensive patent invalidity, undermining the ‘private value’ of patents. It may also weaken the ‘public value’ of medical AI patents, with patent claims impractical to reproduce after the patent ends, or that extend beyond the inventor’s contribution to the state of the art. These issues will hinder technology diffusion and fair competition after patent protection ends.

In this follow-up study, we dive deeper into the sufficiency of disclosure for medical AI patent specifications. Our objective is to generate empirical data essential for a critical evaluation of the adequacy of disclosures in AI/ML technologies in the medical field. Are patent applicants revealing enough details about their AI/ML models, user-specified parameters, algorithms, and training details to enable replication by others in the field? Or are there gaps that might hinder subsequent application and innovation?

Research Questions Specifically, we seek to address the following research questions: (1) To what extent are medical AI patents disclosed? How does this level of disclosure compare to the standards required in peer-reviewed journals publishing medical AI articles? (2) How does it compare to the legal standards? (3) What policy implications follow from the observed disclosures with medical AI patents?; and (4) How can we improve the quality of patent disclosures for AI-related inventions to enhance their ‘public value’?