Methodological Pitfalls in AI Driven Epidemiology: How to Maintain Representativeness

Objectives: Artificial intelligence (AI) and machine learning are increasingly used in epidemiology for predictive modeling and risk stratification. However, non-representative datasets can lead to biased predictions that exacerbate health disparities. This review identifies methodological pitfalls threatening representativeness in AI-driven epidemiology and proposes mitigation strategies across the AI lifecycle, emphasizing epidemiological rigor, fairness-aware design, and applications in low- and middle-income countries (LMICs).



Methods: We synthesized evidence on bias sources in public health AI, including selection, proxy, measurement, and optimization biases. Examples are drawn from LMIC contexts, such as fragmented records in Nigeria and pulse oximetry disparities. A lifecycle framework is proposed, integrating epidemiological principles, fairness-aware machine learning techniques, and continuous monitoring.



Results: Key pitfalls include data-level issues like urban over-representation and model-level issues such as spurious correlations and optimization paradoxes. The proposed lifecycle framework outlines stage-specific mitigations, including personalized federated learning, robust aggregation, and transfer learning for fragmented systems. A case study applying fairness-aware constraints to sub-Saharan African COVID-19 survey data demonstrates the measurable impact of these techniques on improving subgroup equity.



Conclusions: Maintaining representativeness requires intentional, lifecycle-aligned interventions. By prioritizing diverse data, rigorous validation, and equity-focused design—supported by hybrid machine learning architectures and pragmatic guidance on accuracy-equity trade-offs—AI systems can advance equitable public health outcomes.