Key points are not available for this paper at this time.
Artificial intelligence (AI) is rapidly reshaping cancer research, but high technical performance alone is not sufficient for cancer epidemiology, which requires population representativeness, measurement validity, causal reasoning, and demonstrable population-level benefit. This review evaluates AI applications across the cancer epidemiology continuum, from surveillance and data infrastructure through primary prevention, screening and early detection, prognosis, comparative effectiveness, and survivorship. We use a diagnostic matrix that crosses three epidemiologic pillars (study population, measurement, and inference) with six stages of the AI lifecycle, from problem definition to post-deployment monitoring. The current evidence base is promising but uneven. Natural language processing and large language models can improve cancer registration, yet cross-registry transportability and confidentiality safeguards remain incompletely documented. In etiologic research, AI-discovered associations often lack formal causal evaluation. In screening, randomized program-level evidence is strongest for mammography and colonoscopy, but endpoints remain largely intermediate, and no completed AI trial has shown cancer-specific mortality reduction. In prognostic modeling, most published models remain externally unvalidated beyond their development institutions. These limitations are compounded by an equity gap: Training data are concentrated in high-income and European-descent populations, while many regions with rapidly growing cancer burdens remain underrepresented. We propose minimum deployment standards for cancer AI, including multidimensional external validation, calibration assessment, decision curve analysis, equity-stratified reporting, privacy and consent governance, and algorithm vigilance systems for post-deployment monitoring. Realizing the potential of AI in cancer epidemiology will require not only accurate algorithms but also epidemiologic standards that make the population validity, clinical utility, and governance of such algorithms auditable.
Zhao et al. (Mon,) studied this question.