JMIR Publications

ABSTRACT

Artificial intelligence (AI) is increasingly embedded in digital public health systems, including web-based dashboards, mobile outreach platforms, electronic health record-linked registries, automated risk communication, and chatbot-based triage. Yet many evaluations still emphasize model performance, system deployment, dashboard use, chatbot availability, message delivery, or the number of alerts generated. These measures may be necessary for technical evaluation, but they do not establish whether AI has changed a public health action. This Viewpoint proposes public health action endpoints as proximal, auditable, AI-output-specific implementation endpoints positioned between technical performance and population-health outcomes. A public health action endpoint should specify the eligible denominator, AI output or threshold, action opportunity, accountable actor, expected action or predefined justified nonaction, observation window, numerator and denominator calculation, and audit log, together with paired monitoring for harm, workload, privacy risk, model or threshold drift, and inequity. The framework is distinct from model reporting, AI trial reporting, trustworthy-AI guidance, digital health implementation reporting, and broader implementation science frameworks because it forces the AI-output-to-action chain to be specified before implementation value or population-health benefit is claimed. Examples from outbreak surveillance, vaccination outreach, chronic disease registries, heat-health warnings, and chatbot triage illustrate how weak endpoints can be translated into accountable action endpoints. Action endpoints are not validated surrogates for morbidity, mortality, transmission, hospitalization, or equity improvement. They provide claim discipline by distinguishing AI availability from accountable action change, and accountable action change from population-health benefit.