JMIR Publications

ABSTRACT

As artificial intelligence (AI) systems approach and sometimes exceed human diagnostic performance on medical imaging tasks, a foundational question for clinical informatics is what sustains the safety value of human oversight once individual accuracy is no longer the differentiator. We argue that the operative parameter is not accuracy but decorrelation: the degree to which human and AI errors differ on the same cases. The classifier-combination literature establishes that pairwise error correlation (ρ) governs whether additional overseers meaningfully extend the vetting ceiling; when ρ approaches one, the effective number of independent overseers collapses to one regardless of how many systems are nominally deployed. We propose that ρ should be treated as a first-class quantity in the informatics of algorithmic safety. Drawing on published agreement data across 22 modality-task combinations in ophthalmology, radiology, and pathology, we offer illustrative cross-domain patterns suggesting a qualitative association between interpretive complexity (an author-proposed heuristic we call interpretive dimensionality) and concordance, and we observe that concordance appears to rise within fixed tasks over time as AI absorbs the relevant feature space. We call this erosion decorrelation wasting and contrast it with decorrelation injection, in which a novel substrate or task specification transiently lowers ρ. These patterns motivate a multi-lever Strategic Cognitive Reserve combining (1) maintaining clinicians in the diagnostic loop, (2) redesigning human-AI interaction, (3) exploiting the broader clinical context window, and (4) periodically introducing novel substrates. The framework is hypothesis-generating rather than prescriptive; we close with testable predictions requiring paired case-level error data. Multimedia Appendices 1 and 2 provide the task-level source data and the formal dynamics.