Artificial Intelligence in Primary Care for Disease Surveillance, Early Diagnosis, Preventive Care and Population Health Outcomes: A Critical Narrative Review
TL;DR
Progress will depend less on further gains in discrimination than on pragmatic evaluation in representative primary care populations, on prespecified patient-relevant and population-level endpoints, and on governance that treats deployed models as interventions requiring continued surveillance rather than as fixed devices.
Abstract
Primary care generates the longitudinal, population-wide records on which much contemporary health intelligence depends, and it is the setting in which most disease is first suspected and most preventive activity is delivered. Artificial intelligence has accordingly been proposed as a means of strengthening disease surveillance, accelerating early diagnosis, extending preventive care and improving population health outcomes. This critical narrative review examines whether the accumulated evidence supports those expectations, and where it does not. Peer-reviewed literature was identified through Europe PMC, which indexes MEDLINE and PubMed Central records, and through Crossref metadata records, supplemented by citation-informed retrieval and by targeted searching of an intergovernmental publication repository. Evidence was appraised for design adequacy, validation status, setting representativeness and the nature of the outcomes measured. The synthesis identifies a consistent and analytically important asymmetry across the four domains. Diagnostic performance evidence is comparatively mature for image-based and physiological-signal tasks, several of which have been evaluated prospectively or in randomised designs conducted wholly or partly in primary care. Surveillance and preventive applications remain dominated by retrospective model development, with external validation uncommon and prospective impact evaluation rare. Population health outcomes are almost never measured directly; the literature substitutes discrimination statistics, detection yield and process indicators, and the causal chain linking improved prediction to improved population health remains largely untested. Recurrent methodological weaknesses, including case-control sampling, optimistic internal validation, incomplete reporting of population characteristics and the use of convenient proxy outcomes, plausibly explain part of the gap between reported accuracy and demonstrated benefit. Evidence on differential performance across population subgroups is sparse relative to the strength of equity claims made for these technologies. Progress will depend less on further gains in discrimination than on pragmatic evaluation in representative primary care populations, on prespecified patient-relevant and population-level endpoints, and on governance that treats deployed models as interventions requiring continued surveillance rather than as fixed devices.