Individual risk prediction and cancer prevention for the digital health era

Cancer prevention is, in practice, a prediction problem. Decisions about risk reduction, screening, and survivorship depend on forecasts of incident cancer, progression, recurrence, and treatment-related harms, and on whether earlier or more intensive action improves the benefit-harm balance. For prevention-oriented use, prediction tools should provide calibrated absolute-risk estimates over a defined follow-up period, be anchored to local incidence and competing mortality when relevant, and be embedded in explicit clinical or public health pathways in which prespecified thresholds trigger actionable steps. These steps may include intensified lifestyle support, eligibility assessment for chemoprevention, risk-adapted screening decisions regarding start age, interval, or modality, and referral for confirmatory evaluation. Using breast cancer as the most mature archetype, we illustrate how individualized risk estimates can be linked to multiple prevention levers while making the trade-offs of threshold-based strategies explicit, including induced follow-up procedures and the potential for overdiagnosis. We then synthesize requirements for prevention-ready prediction, spanning longitudinal follow-up with registry-linked outcome ascertainment, absolute-risk estimation anchored to local incidence and competing mortality, evaluation of calibration overall and in relevant subgroups, transparent accounting of threshold-induced follow-up procedures and harms, and planned recalibration and updating as baseline risk and practice patterns change. As cohort-enabled illustrations in the Health Examinees (HEXA) infrastructure, we describe a breast cancer prediction benchmark, CancerFree, a questionnaire-first, registry-anchored multi-cancer framework for site-specific risk-to-action specification, alongside an AI-derived diet layer that converts food frequency questionnaire data into interpretable dietary-pattern features for scalable prevention-oriented profiling. Across emerging data layers, added model complexity should be justified not by discrimination alone, but by demonstrable gains in calibrated absolute-risk estimation and in the benefit–harm balance of the pathways they are intended to inform.