Stacey Fisher, Laura C. Rosella
6 min
As public health organizations seek to leverage artificial intelligence (AI) to improve population health, they face unique challenges distinct from those in clinical or individual-level healthcare. This literature review identifies six critical priorities for successfully integrating AI into public health functions, such as surveillance, health protection, and population health assessment. The authors argue that without a strategic, coordinated approach, public health organizations risk failing to realize the potential of AI or, worse, exacerbating existing health inequities.
The authors synthesize findings from various organizational data strategies and AI guidance reports to propose a framework for implementation. The six priorities are:
AI offers the potential for "precision public health," allowing for better-targeted interventions and real-time surveillance. However, the transition to an AI-enabled organization is not merely a technical challenge; it is a structural and cultural one. By focusing on these six priorities, public health agencies can move beyond pilot projects toward sustainable, equitable, and effective AI integration that keeps pace with the rapid growth of health-related data.
Artificial intelligence (AI) has the potential to improve public health's ability to promote the health of all people in all communities. To successfully realize this potential and use AI for public health functions it is important for public health organizations to thoughtfully develop strategies for AI implementation. Six key priorities for successful use of AI technologies by public health organizations are discussed: 1) Contemporary data governance; 2) Investment in modernized data and analytic infrastructure and procedures; 3) Addressing the skills gap in the workforce; 4) Development of strategic collaborative partnerships; 5) Use of good AI practices for transparency and reproducibility, and; 6) Explicit consideration of equity and bias.
Alex: Where does bias fit into this? Because that's often treated as a model-level problem — something you fix in the algorithm.
Sam: The paper pushes back on that framing directly. Their argument is that algorithmic bias is primarily an institutional failure, not a technical one. If the training data systematically underrepresents certain populations — because those populations have less contact with the formal health system, or because historical data encodes past inequities — then no amount of post-hoc debiasing fully corrects for it. The fix has to be upstream: in data collection practices, in how linkage decisions are made, in who is represented in the governance structures that oversee these systems.
Alex: So the "equity-first" framing they use isn't just ethical positioning — it has methodological content.
Sam: That's a fair reading. If you don't embed equity considerations into the pipeline from the start — into what data gets collected, how it gets labeled, what outcomes get predicted — you're not just risking harm, you're building a model that will perform worse on the populations that most need accurate prediction. The equity failure and the performance failure are the same failure.
Alex: And I'd imagine this connects to the point about traditional statistical methods sometimes outperforming AI in these settings?
Sam: Yes, and it's worth being precise about why. When you're working with sparse data, or data that's missing non-randomly across subgroups, the marginal gain from a complex model over a well-specified regression is often negligible — and the interpretability cost is real. The authors aren't arguing against machine learning, but they're noting that the conditions under which ML delivers meaningful lift over classical methods are exactly the conditions that good data infrastructure creates. Fix the infrastructure, and the case for more sophisticated models strengthens. Skip the infrastructure, and you're paying complexity costs for minimal gain.
Alex: So the six-priority framework they propose — is it primarily sequencing guidance? Telling agencies where to start?
Sam: It functions as both a diagnostic and a roadmap. The priorities run from foundational data infrastructure and governance, through workforce development and interoperability standards, to ethical oversight mechanisms. The sequencing matters because the later priorities depend on the earlier ones being in place. You can't do meaningful bias auditing if you don't have the data linkage to know which populations your model is underperforming on.
Alex: That's a useful reframe. The framework isn't "here are six things to do in parallel" — it's "here's the dependency structure."
Sam: Precisely. And the paper's contribution is less about novel empirical findings — this is a review and framework paper, not a primary study — and more about synthesizing the implementation literature into a coherent argument about sequencing and institutional preconditions. The honest limitation is that the framework is prescriptive without being deeply evaluative. We don't have strong evidence yet on which of these investments yields the highest return, or how long the infrastructure build takes before AI deployment becomes viable in a given agency context.
Alex: That's the gap a follow-up empirical literature would need to fill.
Sam: Right. The paper is making a structural argument — that the field has been optimizing the wrong layer — and that argument is well-supported by the implementation failures it cites. But the specific roadmap is still more logic model than evidence base. Which is appropriate for a review at this stage, but worth flagging for anyone who wants to use it as a policy template.
Alex: That's a useful distinction to end on. The diagnosis looks solid; the prescription is a reasonable inference from it, but hasn't been prospectively tested. Thanks for walking through it.
Sam: Thanks for having me. And thanks to everyone listening — if this is your area, the Fisher and Rosella paper is worth reading for the synthesis alone.
Alex: Thanks for listening to ResearchPod.