Machine Learning–Based Analysis of Encrypted Medical Data in the Cloud: Qualitative Study of Expert Stakeholders’ Perspectives

Abstract

Background Third-party cloud-based data analysis applications are proliferating in electronic health (eHealth) because of the expertise offered and their monetary advantage. However, privacy and security are critical concerns when handling sensitive medical data in the cloud. Technical advances based on “crypto magic” in privacy-preserving machine learning (ML) enable data analysis in encrypted form for maintaining confidentiality. Such privacy-enhancing technologies (PETs) could be counterintuitive to relevant stakeholders in eHealth, which could in turn hinder adoption; thus, more attention is needed on human factors for establishing trust and transparency. Objective The aim of this study was to analyze eHealth expert stakeholders’ perspectives and the perceived tradeoffs in regard to data analysis on encrypted medical data in the cloud, and to derive user requirements for development of a privacy-preserving data analysis tool. Methods We used semistructured interviews and report on 14 interviews with individuals having medical, technical, or research expertise in eHealth. We used thematic analysis for analyzing interview data. In addition, we conducted a workshop for eliciting requirements. Results Our results show differences in the understanding of and in trusting the technology; caution is advised by technical experts, whereas patient safety assurances are required by medical experts. Themes were identified with general perspectives on data privacy and practices (eg, acceptance of using external services), as well as themes highlighting specific perspectives (eg, data protection drawbacks and concerns of the data analysis on encrypted data). The latter themes result in requiring assurances and conformance testing for trusting tools such as the proposed ML-based tool. Communicating privacy, and utility benefits and tradeoffs with stakeholders is essential for trust. Furthermore, stakeholders and their organizations share accountability of patient data. Finally, stakeholders stressed the importance of informing patients about the privacy of their data. Conclusions Understanding the benefits and risks of using eHealth PETs is crucial, and collaboration among diverse stakeholders is essential. Assurances of the tool’s privacy, accuracy, and patient safety should be in place for establishing trust of ML-based PETs, especially if used in the cloud.

Trust in and Acceptance of Artificial Intelligence Applications in Medicine: Mixed Methods Study

“…After full-text screening, 90 (73.8%) records were excluded primarily due to the absence of concepts of trust or acceptance and a lack of factors related to trust or acceptance in the main text or data-containing figures. As a result, 32 (26.2%) papers and reports [ 7 , 9 - 12 , 15 , 16 , 19 , 21 , 31 - 53 ] were included in the data analysis.…”

Barriers and Facilitators to Trustworthy and Ethical AI-enabled Medical Care From Patient’s and Healthcare Provider’s Perspectives: A Literature Review

“…To address these trust barriers, the literature discussed the importance of keeping AI systems updated by introducing new rules and cases along with routine performance assessments to enhance the accuracy of decisions made by AI-based medical devices. 73,78 Further regulations and legislation could also increase trust by ensuring the balance between innovation and patient safety and confirming that AI algorithms meet appropriate standards of clinical benefit. 79,80…”

Structural and functional explanations for informing lay and expert users: the case of functional encryption

“…This is especially the case if PETs, such as FE and homomorphic encryption, are novel for users and different from traditional (encryption) schemes that the user may have heard of or are familiar with. Furthermore, they could be perceived as counter-intuitive, as no existing real-world analogies exist [5]. Therefore, research on usable explanations of FE (and of PETs in general), which can contribute to usable privacy notices as part of consent forms, requires further work.…”