Current regulations will not protect patient privacy in the age of machine learning

Abstract

Machine learning (ML) has shown great promise in advancing health outcomes by parsing ever more effectively through massive clinical and genomic datasets. These advances are tempered by fears that they come at the cost of privacy. Since data relating to health are particularly sensitive because of immutability and comprehensiveness, these privacy concerns must be seriously addressed. We consider examples (the Golden State Killer, the Personal Genome Project, and the rise of wearable fitness trackers) where the tension between technological progress and lost privacy is already apparent. We discuss, in light of ML capabilities, the current state of privacy regulation in healthcare. We note the Constitutional right to privacy does not yet in general protect voluntary disclosures of data; HIPAA, the current law regulating healthcare data in the US, does not apply to the burgeoning field of healthcare-adjacent companies and organizations collecting health data; and access controls remain subject to re-identification attacks. We then discuss the active research in algorithmic paradigms for privacy, highlighting their promise but also their limitations. In order to encourage technological progress, reframing privacy for the age of ML might involve extending the Constitutional right to privacy, extending the applicability of HIPAA, and/or enforcing transparent privacy policies.

Post-identifiability in changing sociotechnological genomic data environments

“…Such genome sequences of different organisms, including the human genome itself, have been the product of international consortia working to produce a common resource for the research community, producing data according to FAIR data principles (Wilkinson et al 2016 ), long before the term FAIR emerged. At a time big data and the use in research as well as applications, such as artificial intelligence are increasingly relevant for science and society and are attracting attention of scholars (Kitchin 2014 ; Borgman 2015 ; Crawford 2021 ; D'Ignazio and Klein 2020 ), identifiability through genomic data is becoming more salient along with developments in data sharing platforms as well as cloud and next-generation sequencing technologies (Martinez-Martin and Magnus 2019 ; Bonomi et al 2020 ; Narayan 2020 ; Carter 2019 ).…”