A multimodal, risk-stratified framework for AI-driven early risk prediction and personalised prevention in obesity

Obesity is a multifactorial chronic disease whose worldwide prevalence in adults has more than doubled since 1990, demanding a shift from reactive treatment towards early, personalised prevention. Artificial intelligence (AI) provides a methodological pathway for this shift by integrating heterogeneous, longitudinal evidence-genomic, metabolomic, electronic health record (EHR), wearable Internet-of-Things (IoT), behavioural, and social-environmental-and by translating that evidence into individualised, time-varying risk estimates. Yet the field is fragmented: most existing tools are unimodal, validated on narrow cohorts, opaque to clinicians, and disconnected from the workflows that would render their predictions actionable. In this Perspective we propose an explicit multimodal, risk-stratified framework that links five data layers to a continuous dynamic risk score R(t), defined as a weighted, time-varying aggregation of clinical, anthropometric, behavioural, psychosocial and pharmacological domains. R(t) drives an A/B/C tiering policy that allocates monitoring intensity and intervention modality proportional to risk, and feeds a metabolic-behavioural digital-twin loop in which counterfactual interventions are tested in silico before deployment. We argue that three technical commitments are non-negotiable for translation: (i) cross-modal fusion architectures that respect informative missingness, (ii) explainable, equity-audited risk scoring, and (iii) a five-stage validation pipeline anchored in TRIPOD-AI, decision-curve analysis and post-market drift surveillance. We discuss how this framework reframes long-standing concerns-black-box opacity, demographic bias, real-world fragility-as design constraints rather than afterthoughts, and outline an actionable research agenda for clinically deployable, equitable AI in obesity prevention.