Protocol and databases — what the scoping review will do, what has already been run, and what is still to come.

Status, 19 Oct 2026: protocol in draft · pilot search run · formal search not yet run. The … rows on the Gap map come from an exploratory sweep over open bibliographic APIs (7 Oct 2026) and a pilot of the planned search on open databases (13 Oct 2026). Of these, … are primary studies that meet the draft eligibility criteria. The protocol below is the draft that will be registered on OSF before the formal search runs on the subscribed databases. Each section carries a PRISMA-ScR item number and a status tag.

Question and objectives items 3–4 drafted

What digital interventions for starting or sustaining physical activity have been evaluated in healthy, community-dwelling adults aged 50 and over, and how are they distributed across seven charting dimensions?

  1. RQ1 Which interventions have been evaluated in this population, and how do they sit on the seven dimensions?
  2. RQ2 How do they operationalise engagement (responding to the intervention) and adherence (doing the behaviour), and what is reported about the trajectory over time, in particular after the novelty window?
  3. RQ3 Which combinations of dimensions co-occur, and which are absent?
  4. RQ4 What do adjacent literatures (JITAI and micro-randomised trials in younger or clinical samples; LLM coaching in general adults; engagement mechanics from games) offer that has not been transferred to this population? Reported as a labelled "adjacent evidence" section, outside the primary search.

Protocol and registration item 5 to register

DESIGN
Scoping review, not systematic review

The question is "what exists and where are the holes", not "what is the pooled effect". The evidence is heterogeneous (HCI pilots, RCTs, control-systems papers, LLM evaluations), so there is no meta-analysis and no critical appraisal; design and sample size are reported for every study instead.

REPORTING
PRISMA-ScR, JBI method

Reported to the PRISMA-ScR checklist (Tricco et al., 2018, Annals of Internal Medicine, 10.7326/M18-0850) and conducted to the JBI scoping-review method. The checklist is pre-filled as the work proceeds.

REGISTRATION
OSF, before the formal search

Question, eligibility, databases, search strings and the charting form are fixed on OSF before any subscribed-database search runs. Not yet registered: the pilot below is what the registered strings are being tuned on.

TEAM
One reviewer, two supervisors

Screening and charting decisions are made by the candidate. Both supervisors review the charting on a sample and the paper. There is no independent second screener, so agreement statistics (κ) cannot be reported; this is stated as a limitation and is one of the questions below.

Eligibility criteria item 6 · PCC drafted

POPULATION

Adults aged 50 and over, community-dwelling, not recruited for a diagnosed condition. Mixed-age samples count when at least half the sample is 50 or over, or results are stratified. Patients are out: a diagnosis, a clinician in the loop and a prescribed target make a different motivational setting, so their results do not transfer upstream, where the prevention case is. Risk-factor samples (overweight, prediabetes): decision pending.

CONCEPT

A digital intervention intended to start or sustain physical activity: apps, wearables with feedback, message systems, conversational agents, social robots delivering activity support. It must deliver something, not only measure. Passive monitoring alone is out.

DURATION (CHANGED 19 OCT 2026)

Intervention or follow-up longer than 12 weeks. The time course is charted for every included study: intervention length, follow-up length, each timepoint at which engagement or an outcome was measured, and whether it held. Six months is then a number the review produces, not one chosen in advance; "24 weeks or more" is kept as a sub-code (it is what the map shows as D2 today).

CONTEXT · SOURCES · EXCLUDED

Any country; home or community setting; 2010 onward (smartphone era). Peer-reviewed primary studies of any design, including registered protocols; preprints included and flagged. Reviews are used for citation chasing and are not charted. Excluded: rehabilitation and disease-management populations, no digital component, abstracts without a paper.

Information sources and databases items 7–8 pilot run

Three searches so far or planned. The first two have run and their strings are published; the third is the one the protocol will register.

1 · Exploratory sweep — open APIs, 7 Oct 2026 run

Europe PMC, OpenAlex, Crossref and arXiv, seeded from the ten papers the research proposal cites, then outward by topic and citation chasing. No subscription database, no screening counts: it was run to pilot the seven-dimension framework, and it produced the first 138 charted papers. Exact strings are in the fold-out at the foot of this page and in search-log.md.

2 · Pilot of the planned search — open databases, 13 Oct 2026 run

Three concept blocks AND-ed (population 50+ or older · digital intervention · physical activity), and a four-block variant adding an outcome block (engagement, adherence, retention, attrition, "novelty effect", long-term, sustained, maintenance, abandonment, "continued use", dropout). Limits 2010–2026; English on Europe PMC. Run 13 Oct 2026, 08:02 NZDT.

DatabaseQueryHitsPulled
Europe PMC (REST, title/abstract fields)3 blocks2,7282,728
Europe PMC4 blocks (+ outcome)833833
OpenAlex (title and abstract search)3 blocks18,3383,000 most cited
OpenAlex4 blocks (+ outcome)6,0836,083
arXiv API(LLM OR chatbot OR conversational agent) AND (physical activity OR exercise OR walking OR fitness) AND (coach* OR intervention OR behaviour change OR older adults)119119
arXiv API(just-in-time adaptive OR JITAI OR micro-randomized OR contextual bandit) AND (physical activity OR mobile health OR mHealth OR steps)146146
ClinicalTrials.gov (API v2)physical activity × digital × older terms; minimum age ≥ 50 y; start ≥ 2010≥ 1,0001,000 (cap)
Records identified—12,909

Sensitivity check against the anchor papers: Bickmore 2013, Greysen 2024, King 2016, Patel 2019, Finkelstein 2016 and Hermsen 2017 were retrieved; Wong 2024 was not, because its abstract says "wearable monitoring device" and "healthy behaviors" with no physical-activity term. The behaviour block gains tracker-use and health-behaviour terms before the formal run.

3 · Formal search — subscribed databases planned

DATABASES

MEDLINE (Ovid) · Embase · APA PsycInfo · CINAHL · Scopus · Web of Science · IEEE Xplore · ACM Digital Library; arXiv for the LLM stream; ClinicalTrials.gov and ANZCTR for registered protocols. Run from the campus PC on a fixed date, each search logged (string, date, hits). Forward and backward citation chasing from every included study and the ten anchors.

CONCEPT BLOCKS (TO BE PEER-REVIEWED, PRESS)

Population: older adult* OR aged OR "middle-aged" OR senior* OR "50 years" … AND Intervention: mHealth OR "digital health" OR app OR wearable* OR "activity tracker" OR chatbot OR "conversational agent" OR "large language model" OR JITAI OR "just-in-time" OR "adaptive intervention" OR nudge* AND Behaviour/outcome: "physical activity" OR exercise OR walking OR steps OR sedentary; engagement OR adherence OR retention OR attrition OR "novelty effect", plus the tracker-use terms the pilot showed were missing. Strings are translated per database and checked by a University Library research adviser before running. Sensitivity rule: all anchor papers must be retrieved.

Selection of sources items 9, 14 pilot run

The pilot flow, as the formal flow diagram will be built. The machine pre-screen is a sorting aid: it orders the pile, it never includes or excludes a record.

  1. 12,909records identified across the seven searches
  2. 9,657unique after de-duplication (DOI, then normalised title)
  3. 27already charted from the exploratory sweep → 9,630 new
  4. 2,674tier A by machine pre-screen on title and abstract (1,455 strict); 208 tier B; 595 reviews set aside for citation chasing
  5. 81abstracts read by the reviewer: 55 strict-A records in the 50+ × (LLM | adaptive) cell, and 26 dimension-carrying records from a 275-record 50+ × ≥ 6-month batch
  6. 56charted into the table → 191 rows; … rows meet the draft eligibility criteria
FORMAL SCREENING (PLANNED)

Titles and abstracts, then full texts, in a screening tool (Rayyan or Covidence, to decide). Every full-text exclusion carries a reason. Counts at each step feed the PRISMA flow diagram. Decisions by one reviewer; supervisors check a sample.

PROVENANCE PER ROW (NOW)

Every row on the map records how much of the paper was read: full text, abstract, or record only. Current split: … full text · … abstract · … record only. A dimension not established in the text read is scored 0, never guessed.

Data charting items 10–11 piloted on 191

THE CHARTING FORM

Seven binary dimensions with written decision rules: population 50+, beyond the novelty window, state and context as input, behaviour theory, LLM delivery, closed loop with a fixed law, closed loop with a learned law. Definitions and rules are on the Dimensions page (revised 19 Oct 2026). Sub-codes: D2 time course (intervention length, follow-up, timepoints, held or not; ≥ 24 weeks); D3 input classes (wearable, environment, schedule, health state, user feedback) and what the input drives (display only, content, timing).

DESCRIPTIVE ITEMS

Year, country, design, n, age (floor and mean), delivery channel and device, theory named and techniques identifiable (BCTTv1), engagement metric used, attrition at last follow-up, primary outcome type, and delivery cost (human coaching hours, incentives paid, hardware supplied, cloud or on-device processing) as reported or not reported. Funding source.

RULES

Binary, conservative: 0 unless the text read establishes it. "Not stated" is recorded separately at extraction and collapsed to 0 in the matrix. Theory, framework and review papers are listed but not charted. Rules are piloted on the current rows and revised in writing before the formal charting; the change log is public (this page and the version history).

MACHINE-ASSISTED, HUMAN-DECIDED

First-pass codes and one-paragraph summaries are drafted by software from the text retrieved, then confirmed or recoded by the reviewer on the map itself (sign-in on the Gap map; each recode carries a reason and a name). The map shows published codes to visitors and the reviewer's recodes to the reviewer until they are accepted. Reviewed so far: … of ….

Synthesis items 13, 18 planned

Descriptive counts per dimension; co-occurrence of dimensions (the UpSet plot on the Gap map); an evidence-gap grid of dimension × design; a narrative per dimension; and a time-course table for duration (RQ2). No effect sizes are pooled. The study this thesis proposes scores 1 on D1 to D5 and D7, so the review's test is whether that combination exists. In the current eligible set it is … papers; of … eligible studies, … ran six months or longer, … adapt timing or dose with hand-set rules, … learn the rule from the person, and … use a large language model. These numbers are read live from the same data file as the map and move as papers are charted.

QUESTIONS FOR REVIEWERS
  1. One screener with supervisor checks on a sample: acceptable for a scoping review if declared as a limitation, or should a second screener code a random sample so agreement can be reported?
  2. Duration: is it sound to include trials longer than 12 weeks and let the charted time courses show where the novelty window closes, rather than filter at six months?
  3. The machine pre-screen orders the pile and a person reads everything that enters: should its rules be published in the protocol as part of the search, or described as a screening aid only?
  4. Registered protocols and preprints: chart them in the matrix (flagged), or report them separately?

Comments welcome in person; each change made from them will appear in the version history.

Two searches ran in parallel. A covered behavioural science, implementation science, population and cross-domain analogues; B covered LLM/AI, machine-learning state estimation, feedback and control, wearables, ethics and embodiment. Semantic Scholar was attempted and refused requests (HTTP 429). The complete log, including the exact-title checks, is in search-log.md. Duplicates were merged by DOI: 76 + 71 → 135, plus 3 JITAI-lineage papers added after verification → 138.

Search A · Europe PMC (REST, resultType=core)

  • TITLE:"novelty effect" AND (wearable OR "activity tracker" OR mHealth)
  • TITLE:"wearable activity trackers" AND (TITLE:"systematic review" OR TITLE:"meta-analysis") AND "older adults"
  • TITLE:"gamification" AND TITLE:"physical activity" AND TITLE:"meta-analysis"
  • TITLE:"gamification" AND TITLE:"older adults" AND TITLE:"physical activity" AND randomized
  • TITLE:"reinforcement learning" AND TITLE:"physical activity"
  • TITLE:"large language model" AND TITLE:"physical activity" AND (coach OR coaching)
  • TITLE:"chatbot" AND TITLE:"physical activity" AND (systematic review OR meta-analysis)
  • TITLE:"financial incentives" AND TITLE:"physical activity" AND TITLE:"meta-analysis"
  • TITLE:"Māori" AND (digital health OR mHealth OR telehealth)
  • TITLE:"dark patterns" AND (health OR app)
  • TITLE:"gamification" AND (ethics OR ethical) AND health
  • TITLE:"older adults" AND TITLE:"wearable" AND (abandon OR long-term use OR sustained use)
  • TITLE:"habit" AND TITLE:"physical activity" AND (meta-analysis OR systematic review)
  • TITLE:"exergames" AND TITLE:"older"
  • TITLE:"adaptive" AND TITLE:"goal" AND TITLE:"physical activity"
  • TITLE:"HeartSteps" AND micro-randomized

Search A · OpenAlex and Crossref (bibliographic look-ups)

Named-paper queries for the proposal's anchors and the foundational theory papers: PhysioLLM; Engage-PA; self-concordance in older adults; LumiHealth; Healthier SG; Ryan & Deci 2000; Behaviour Change Wheel; MINDSPACE; Lally 2009; Nahum-Shani JITAI; Yardley 2016; Perski 2016; Eysenbach 2005; RE-AIM; CFIR; Fogg 2009; Sheldon & Elliot 1999; Hunicke 2005; Hamari 2014; Deterding 2011; Sunstein 2020; Prochaska & Velicer 1997; Mathur 2019; Schüll 2012; Koivisto & Hamari 2019; Gray 2018; Jörke 2025.

Search B · OpenAlex (about 75 free-text queries, examples)

  • Just-in-time adaptive interventions in mobile health key components and design principles
  • HeartSteps micro-randomized trial physical activity suggestions Klasnja
  • control systems engineering behavioral interventions Rivera
  • adaptive step goals randomized trial physical activity Adams
  • large language model fine-tuned transtheoretical model physical activity coaching My Heart Counts
  • PhysioLLM wearable data large language model personalized health insights
  • receptivity just-in-time interventions mobile health prediction
  • conversational agents physical activity diet behaviour change systematic review meta-analysis
  • LLM generated motivational messages physical activity evaluation
  • Apple Watch heart rate variability validation · Apple Watch VO2max estimation validity
  • wearable activity trackers older adults use and abandonment
  • weather and physical activity accelerometer older adults
  • novelty effect wearable fitness tracker engagement decline
  • dynamic difficulty adjustment exergame older adults
  • digital nudging autonomy manipulation ethics
  • socially assistive robots older adults loneliness systematic review meta-analysis
  • large language model conversational agent older adults

Search B · Europe PMC free text

  • conversational agent older adults lifestyle
  • contextual bandit physical activity messages personalization trial
  • sedentary interruption JITAI older adults wearable
  • ElliQ older adults living alone loneliness
  • Maori perspectives on privacy

Search B · arXiv (by id)

  • 1802.07740 · 1909.03539 · 2506.07275 · 2312.06674

Limits: OpenAlex rate-limited mid-session (HTTP 429) and the remaining queries moved to Crossref and Europe PMC; Semantic Scholar refused all requests. Several intended sources could not be verified and were left out; they are named in the log.

See the … papers as they stand →

Framework demonstration — coding is machine-assisted and under human review (… of …). The formal search, the PRISMA flow diagram and the screening counts will appear here when the protocol is registered and the search has run.