Reporting follow-up budgets for diagnostic and predictive AI.

Abstract

To argue that diagnostic and predictive AI should be evaluated by both classification performance and the downstream work their outputs create. Prior work has framed AI as an intervention and has quantified alert burden or decision utility under resource constraints. This perspective extends that logic beyond surveillance alerts by proposing follow-up budget reporting for diagnostic and predictive AI. Across lung nodule detection, sepsis alerting, and autonomous diabetic retinopathy screening, similar classification performance can imply very different demands on imaging, specialist access, nursing attention, and patient waiting. A follow-up budget statement should specify the comparator, scope and stopping rule, downstream action volume, resource types consumed, yield per clinically meaningful true-positive case, and the local pathway assumptions and stakeholders responsible for those estimates. Clinical AI evaluation should make follow-up capacity visible alongside discrimination, calibration, and decision-analytic measures.