JMIR Publications

ABSTRACT

Healthcare artificial intelligence (AI) is advancing rapidly, yet improvements in computational capability do not reliably translate into corresponding improvements in healthcare. Existing implementation, sociotechnical, realist, clinical utility, and AI evaluation approaches provide substantial resources for understanding how AI is evaluated, adopted, embedded, and made consequential in healthcare systems. However, evidence about changes in computational capability and changes in realised healthcare value is often generated across different research traditions, stages of the technology lifecycle, and units of analysis, leaving their relationship insufficiently examined as an explicit object of explanation. We describe this relationship as capability–value coupling: the extent and manner in which variation in task-relevant computational capability contributes to variation in realised healthcare value through context-dependent mechanisms over time. We argue that studying this relationship requires investigators to specify the capability expected to matter, determine whether additional capability creates actionable utility, examine the processes through which it may influence healthcare, prespecify the relevant healthcare value, and distinguish co-occurring capability–value alignment from evidence supporting a coupling explanation. Longitudinal and comparative investigation can further identify mechanisms and boundary conditions under which additional capability becomes consequential, is amplified, attenuated, delayed, or fails to generate additional value. Capability–value coupling is proposed not as a new translational framework, but as an explanatory focus for connecting existing forms of evidence and developing more cumulative knowledge about when, how, and why advances in AI become advances in healthcare.