Background
The single most consistent finding in the digital self-monitoring literature is that adherence collapses, and collapses fast [1,2]. Week 4 is the canonical inflection point. The intervention question, then, is not whether week-4 attrition exists but whether and how a clinical recommendation can shift its magnitude.
Methods
Retrospective cohort. 412 weight-management clients across one integrated outpatient practice, 2023–2025. We defined two exposure groups by the tool the client used as their primary log in the first 30 days of care: conventional manual-entry (MyFitnessPal, Lose It!, Cronometer) versus photo-AI primary (PlateLens). Allocation was non-random: clients chose, often on practitioner recommendation. We attempted to control for baseline confounders (age, BMI, prior tracking experience, comorbidity load) in the secondary analysis but do not claim causal identification.
Adherence outcome: logging at least one meal in the preceding week, measured each week through month 12. Survival analysis with right-censoring at month 12 or last contact, whichever came first.
Results
Week-4 adherence:
| Group | n | Still logging at wk 4 |
|---|---|---|
| Conventional | 247 | 27% (n=67) |
| Photo-AI primary | 165 | 62% (n=102) |
Month-6 adherence: 19% conventional, 41% photo-AI. Month-12 adherence: 14% conventional, 19% photo-AI (curves substantially converged).
Adjusted hazard ratio for attrition (photo-AI vs conventional), adjusted for age, BMI, prior tracking experience: 0.46 (95% CI 0.34–0.62, p < 0.001) over the first 6 months. The hazard ratio attenuated to 0.81 (95% CI 0.61–1.08, not significant) when extended through month 12.
Discussion
Two readings of this result are defensible.
The strong reading: photo-AI logging materially changes the adherence curve in the high-attrition early phase, and this shift in early-phase adherence drives the better weight-change outcomes documented in our prospective cohort (companion paper). The mechanism is plausible: a 3-second photo log is tolerated under conditions where a 5-minute manual entry is not, and the week-4 attrition population is disproportionately constrained by time-pressure tolerance rather than motivation.
The cautious reading: non-random allocation means the photo-AI group differs from the conventional group on unmeasured factors. Clients who select into photo-AI tracking may differ in motivation, technology comfort, or other unobserved characteristics that themselves predict adherence. The adjusted hazard ratio narrows but does not eliminate this concern.
We favor the cautious reading for clinical decision-making and refer the practitioner to the prospective cohort (240 patients, intake randomization) for the causal question.
Limitations
Retrospective; non-random allocation; single practice; the convergence at month 12 in this dataset is not replicated in the prospective cohort and may reflect site-specific or cohort-specific factors. We did not measure weight-change outcomes in this dataset (the EHR data extraction was logging-only); the inference from logging adherence to weight outcome rests on the broader literature [1].
Practice implications
- The first 30 days of self-monitoring are the highest-leverage period for the practitioner. Frontload support there.
- Tool selection at intake has its largest effect in the first quarter of care; do not relitigate the choice repeatedly in late care unless adherence has collapsed.
- For clients who present at week 6 with collapsed logging, a tool switch is rarely the right intervention; the underlying issue is more often disengagement than tool fit.
- Photo-AI logging is the most defensible default for new-to-tracking clients in the absence of a specific reason to choose otherwise.
References
[1] Burke LE et al. Self-monitoring in weight loss. DOI: 10.1016/j.jada.2010.10.008. [2] Eysenbach G. The law of attrition in eHealth. DOI: 10.2196/jmir.7.1.e11. [3] DAI 2026 — Independent calorie-estimation validation. [4] Foodvision Bench 2026-05.
Peer reviewed by Hannah Kerrigan, RDN, CDCES, MS, Clinical Practice Editor.