The data challenge: from biases to regulation

Abstract

Algorithms based on machine learning and deep learning unlock the technological possibility of using aggregated healthcare data to produce powerful models that facilitate and improve the accuracy of diagnosis, tailoring treatments and targeting resources with maximum effectiveness. However, there are key ethical and technical challenges that need to be evaluated before AI solutions can be widely implemented in the clinic – and often, the two are closely intertwined. Issues about ownership and control of confidential health data of patients, their privacy, public trust, accountability and responsibility about the application of an AI tool all require consideration; and ethical considerations naturally come to the fore. Two of the biggest technical barriers are data standardization and data sharing, both of which are essential to fight possible algorithmic biases. In turn, data sharing poses its own set of questions, including regulation, ownership of data and privacy protection, protection and cybersecurity, and accountability. We have found that, although there have been advances in all these areas, they remain insufficient. The rapid development of the field is making it difficult for regulatory bodies to keep up – and similarly with the attempts to standardize data. There are increasing numbers of data-sharing initiatives such as biobanks and international consortia in the European Union and abroad, and frameworks focusing on interoperability and clinical translation. However, current methodologies for anonymization and de-identification are often suboptimal; and while there is a critical need to provide high-quality diverse data to train unbiased algorithms, maintaining patient privacy must always be carefully maintained. These aspects are critical to the success of the whole field an-d will require close collaboration on a global scale.