Seven dimensions for charting the included studies. The study this thesis proposes scores 1 on D1 to D5 and D7.
Charting dimensions, not search terms: only population, intervention type and duration can be searched for. D3 to D7 are coded from the full text after screening. Definitions revised 19 Oct 2026; tap a card for the decision rule.
D1
Healthy adults 50+
Community-dwelling adults aged 50 or over, not recruited for a diagnosed condition.
Rule: an inclusion age of 50 or over, or at least half the sample aged 50 or over (or a mean age of 50+ with the sample described as older adults), and not recruited for a diagnosed condition. Why patients are out: a diagnosis as the cue, a clinician in the loop and a prescribed target make a different motivational setting. Results from it do not transfer upstream, to people with nothing wrong yet, which is where the prevention and the economic case sit. Open: risk-factor samples (overweight, prediabetes) — decision pending.
D2
Beyond the novelty window
Longer than 12 weeks, with the time course charted, so the review shows where the window closes.
Rule: intervention or follow-up longer than 12 weeks. For every included study chart the intervention length, the follow-up length, each timepoint at which engagement or an outcome was measured, and whether it held at that point. Sub-code: 24 weeks or more (the earlier "≥ 6 months" rule, still shown on the map as D2). Why: most digital interventions report a novelty window of weeks to about three months. Charting the time course lets the trial length (six months, five, twenty weeks) come out of the review rather than be chosen in advance. Caution: most trials report an endpoint, not a trajectory, so the answer will rest on the studies with repeated measures.
D3
State and context as input
State and context are collected during the trial and used as an input to what the intervention does, rather than a goal and inputs fixed at baseline.
Input classes: W (wearable biometrics), E (environment: location, weather), S (schedule: calendar, work/holiday), H (health status: sick, injured, pain), F (explicit user feedback). Sub-code — what the input drives: display or feedback only · the content of the intervention · its timing. State-triggered "just-in-time" delivery by fixed rules lands here under timing. Rule: coded 1 only if a class is used as an input to a decision, not merely collected or displayed; report the class set and input breadth 0–5.
D4
Behaviour science
A named behaviour-change theory or framework shapes the intervention: COM-B and the Behaviour Change Wheel, Fogg's B=MAT, self-determination theory, the transtheoretical model, social cognitive theory.
Rule: the theory must drive design choices (which techniques, which messages, which moments), not only be cited in the introduction. Techniques are also identified with the BCT taxonomy (BCTTv1) where the text allows. Not D4: implementation-science frameworks (RE-AIM, CFIR) are a different literature; they are tagged separately.
D5
LLM delivers
A large language model generates or selects the participant-facing content at run time.
Rule: generative or selective output at run time. Rule-based chatbots score 0; an LLM used only to author a message bank offline scores 0.
D6
Closed loop, fixed law
The person's response to earlier interventions is an input to the next one, and the rule that uses it is set in advance.
Rule: a feedback path from the person's response (steps achieved, prompts answered, goals met) to the next intervention; the parameters of the rule do not change during the trial. Example in the map: Adams et al. 2013 — the next step goal is a fixed percentile of the person's own recent steps. Not exclusive with D7: one system can run a fixed law for one output and a learned law for another.
D7
Closed loop, learned law
The person's response updates the parameters of the decision rule: a bandit, reinforcement learning, regression, anything that learns from the person.
Rule: the parameters of the decision rule (when, whether, what or how much) are updated from the person's own responses during the trial. The method does not matter; "probabilistic" is not required. Example in the map: Liao et al. 2020 — the HeartSteps bandit learns, per person, when a walking suggestion is worth sending. The thesis: D7 with D1 to D5. The gap claimed is that combination, not "all seven".
D6 and D7 differ in one thing: whether the rule learns. Both take the person's response as an input, and one system can do both, so they are not coded as exclusive. A state-triggered open loop (the weather or the calendar decides when to prompt, by a fixed rule, with no feedback from the response) is D3 with the sub-code timing, not D6.
Coding status. These definitions are dated 19 Oct 2026. The Gap map still shows the 8 Oct coding, and its labels say what was coded: D2 as six months or more, D6 as adaptive timing, D7 as closed-loop calibration. Re-coding under the revised D2, D3, D6 and D7 is in progress; counts will move when it lands, and the change will be a new version in the history.