JMIR Publications

ABSTRACT

Background:

NHS Wales routinely collects patient-reported outcome measures, and these together with other clinical data offer an opportunity to design machine learning (ML) technologies that could advance the implementation of prudent healthcare principles (a healthcare strategy encouraged by the Welsh Government). However, the wide adoption of such technologies is not only dependent on the development of technically well performing ML algorithms, but also on end-user barriers and facilitators.

Objective:

This study aimed to identify potential barriers and facilitators to the use of ML in healthcare decision-making in Wales. The study focused on the end-users of such potential technologies: members of the public (as potential patients) and healthcare professionals involved in therapeutic or treatment decision-making.

Methods:

An online survey using Microsoft Forms was conducted. It was open to anyone who was at least 16 years old and lived in Wales (member of the public criterion), or were registered healthcare professionals working in Wales and participating in treatment or therapy decision-making (healthcare professional criterion). The anonymous survey was open from the 4th of December 2024 to the 4th of March 2025. The survey utilised single choice, ranking, and free-text questions, which were phrased differently for both eligibility groups. Data analysis was based on respondent selected eligibility criterion and self-declared general attitude towards healthcare artificial intelligence (generally supportive, opposed or uncertain), using descriptive statistics and summary of free-text responses.

Results:

309 respondents filled out the survey, 179 selecting the member of the public criterion, and 130 selecting the healthcare professional criterion. 209 self-identified as having a generally supportive attitude towards healthcare AI, 31 as generally being opposed towards healthcare AI and 69 as being uncertain. Overall, respondents placed a large emphasis on the presence of evidence for the technologies effectiveness and humans being in control of the healthcare process, even if this meant that care processes were not as fast as they could be with a higher degree of automation. Those with a negative attitude towards AI placed more emphasis on human autonomy than other respondent groups.

Conclusions:

Those developing and implementing healthcare AI technologies should develop an unbiased evidence base for the effectiveness of their technologies, using transparent methodologies, and continue their evaluation when the technology is in place. Moreover, implementation should not decrease patient-clinician contact, but automate specific tasks only and maintain a human-in-the-loop.