Data Issues in AI Applications for Public Health — An Application of the C-E-A Framework: Compliance, Efficacy, and Autonomy

Abstract

Data concepts relevant to AI applications in public health — such as availability, relevance, and representativeness — are usually explained in their general, basic sense for a broad readership of engineers and public health professionals alike, without being organized by what kind of practical failure each one represents or connected to the specific compliance, efficacy, and autonomy concerns a public health professional must act on. This paper specifies a repeatable procedure for classifying such data concepts under the C-E-A framework (Compliance, Efficacy, Autonomy) and applies it to sixteen concepts named in the author’s public health AI textbook, with Compliance further split into its legal/regulatory, practical/adoptive, and multi-party-permission tiers, and independently regroups the same sixteen concepts by the grain at which each is actually checked — design-time, runtime, or both. The classification produces a sixteen-concept map (Table 1) and surfaces two asymmetries in this vocabulary — only one concept is classified as primarily an Autonomy concern, and only one as primarily a multi-party-permission concern — alongside a grain regrouping (Table 2) and three cross-pillar tensions discussed without resolution. The concept map and its accompanying tensions are offered as one useful way to view these data concepts, not the only one, extending the C-E-A framework’s fine/coarse grain distinction to the data lifecycle for the first time.