We have just published a preprint of a comprehensive survey of technical challenges and solutions to achieve responsible and regulatory machine learning for medical applications.
Machine learning is set to affect the medical community significantly, but many challenges remain to be solved. To ensure beneficence and prevent harm, medical ML applications must be developed responsibly and in accordance with existing and future regulations. This raises a number of important technical challenges concerning the safety, security, privacy, transparency, and fairness of medical ML systems, many of which are the subject of current research and should not be considered solved.
In our survey, we first provide an overview of the regulatory landscape concerning medical ML systems, discussing both current regulations and requirements, as well as likely future developments. We then proceed to discuss the technical challenges to overcome in depth, as well as potential solutions to these challenges. Our survey is mainly aimed at medical ML developers, but should also be of interest to manufacturers, regulators, and ethicists.
In conceptualizing the paper, we were driven by the desire to close the principles-to-practice gap. To the best of our knowledge, there is a lack of literature that thoroughly addresses ethical and regulatory demands while at the same time going the extra mile to indicate ways in which technologists can contribute. In this article, we have mostly focused on technical approaches to solutions. However, in doing the complexity of responsible machine learning justice, we also highlight the importance of non-technological or hybrid approaches.
The work that led to our paper is part of the KI-SIGS project and was a collaboration with Yannik Potdevin and Dirk Nowotka of CAU Kiel, Esfandiar Mohammadi of the Institute for IT Security (Lübeck), Stephan Zidowitz of Fraunhofer MEVIS (Bremen), Sabrina Breyer and Christian Herzog of the Ethical Innovation Hub (Lübeck), Ludwig Pechmann of UniTransferKlinik (Lübeck), and Martin Leucker of the Institute for Software Engineering and Programming Languages (Lübeck), besides Eike Petersen, Sandra Henn, and Philipp Rostalski of the IME.
For any feedback, comments, or discussions, please contact Eike Petersen.
