JMIR Publications

ABSTRACT

Digital health interventions can change behaviour and improve outcomes, but most deployed systems remain largely static: they deliver pre-authored content on fixed schedules rather than adapting to a person's moment-to-moment state. Just-in-time adaptive interventions (JITAIs) promise support precisely when a person is both receptive and at risk, yet the field lacks a coherent account of which machine-learning (ML) paradigm should answer which JITAI design question. We propose a decision-level framework that maps unsupervised, supervised, causal, and reinforcement-learning roles onto the core JITAI design decisions, composed within a single closed-loop architecture. We formalise six recurring design questions (whom, when, what, how, how much, and whether to intervene) and specify the inferential machinery that answers each. We embed this mapping in a sense-understand-predict-decide-intervene-learn loop (treated as functional roles rather than a mandatory serial pipeline), with an explicit constrained-safety layer in which the learned policy proposes candidate actions and auditable rules dispose of unsafe ones. First, we offer an empirically testable paradigm-to-decision mapping (each assignment names setting-specific evidence that would count against it) rather than another taxonomy of components. Second, we describe a constrained-action safety architecture in which a learned policy proposes and auditable rules dispose, making adaptive systems governable and every delivered action attributable, together with explicit limits on what a rule gate alone can guarantee. Third, we specify a paradigm-matched evaluation protocol spanning micro-randomised trials, off-policy evaluation, and calibration analysis, illustrated across smoking cessation, physical activity, and mood interventions. The framework gives designers a principled way to decide where learning belongs, where rules must dominate, and how to evaluate each component before deployment. It reframes "adding ML to a health app" as a set of scoped, testable decision problems, and it clarifies the boundary between what should be learned and what must remain clinically governed. This is explicitly a framework paper, not an empirical study: its contribution is the mapping, the architecture, and the protocol, all of which remain to be instantiated and tested.