Pith. sign in

REVIEW 2 major objections 5 minor 26 references

Adaptive timing of digital health prompts consistently covers more brushing events than fixed, user-reported schedules.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 02:25 UTC pith:XLDH7C4C

load-bearing objection Solid, first-of-its-kind empirical case that online adaptive timing beats fixed user schedules for anticipatory habitual prompts; coverage gains are real, but the paper never links them to behavior or clinical endpoints. the 2 major comments →

arxiv 2607.09518 v1 pith:XLDH7C4C submitted 2026-07-10 cs.HC

Learning When to Intervene on Habitual Behaviors: A Case Study in Oral Health Care

Bhanu Teja Gullapalli , Vivek Shetty , Anna L. Trella , Asim H. Gazi , Susan A. Murphy This is my paper
classification cs.HC
keywords digital health interventionsadaptive intervention timingonline learningjust-in-time adaptive interventionsoral healthhabitual behaviorcoverage metric
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Digital health prompts for habits like tooth brushing only help if they arrive shortly before the person acts. Fixed times set from a user's report at enrollment often drift out of sync as routines change, so many prompts arrive after brushing or hours too early. This paper shows that an online, person-specific predictor of brushing time, updated as new data arrive, can continuously reschedule the candidate prompt time and thereby raise the fraction of windows in which a prompt falls inside a useful lead window. The claim is tested three ways: retrospective coverage on a prior micro-randomized trial, simulations in which timing can itself shift later brushing, and preliminary data from an ongoing randomized trial that already uses the adaptive rule. Across all three, adaptive schedules beat the fixed baseline, with the simplest Bayesian linear model performing best under sparse data.

Core claim

When intervention times for habitual behaviors must be chosen in advance, continuously updating each person's expected behavior time (and optionally its uncertainty) yields higher coverage of subsequent events than any fixed schedule derived from user-reported usual times, including fixed-offset versions of those times.

What carries the argument

Coverage: an intervention time I covers a brushing time B when 0 < B − I ≤ C hours; the online adaptive policies set I = μ̂ − Δ or I = μ̂ − α σ̂, where μ̂ and σ̂ are updated person-specific mean and uncertainty of brushing time.

Load-bearing premise

The paper treats higher coverage as a sufficient stand-in for usefulness, without showing that better-timed prompts actually improve brushing quality, duration, or oral-health outcomes.

What would settle it

In the ongoing RCT, if the adaptive arm does not produce higher coverage than the fixed user-time arm for C ≥ 3 hours once the full sample is enrolled, or if higher coverage fails to improve any clinical or behavioral endpoint, the central claim is refuted.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper identifies timing misalignment as a problem for anticipatory digital health interventions on habitual behaviors (e.g., tooth brushing), where fixed user-reported schedules drift from actual event times. It proposes an online framework that maintains individual-specific predictive models of behavior time (online BLR, Hoeffding tree, NN with MC dropout) and sets candidate intervention times via either a fixed offset from the predictive mean or an uncertainty-scaled offset (Mean + Unc). Using a prior oral-health MRT (N=69), the authors evaluate coverage (0 < B − I ≤ C) offline against observed brushing timestamps, in three simulation variants that allow interventions to shift subsequent brushing times, and with preliminary data from an ongoing RCT (N=20) that deploys Mean + Unc + BLR. Across settings, adaptive rules improve mean coverage over fixed user-input baselines for C ≥ 2–3 h (Tables 5, 7, 8), with BLR performing best.

Significance. The work fills a genuine gap: most JITAI literature either uses static schedules or reactive state detection, leaving anticipatory habitual timing under-addressed. The multi-pronged evaluation (retrospective coverage, causal-effect simulations, and live RCT deployment with ClinicalTrials.gov IDs) is a clear strength and raises the bar for methods papers in this area. If the coverage gains hold and later translate to behavioral endpoints, the framework offers a practical, low-data online method that can be dropped into existing sequential decision pipelines. Explicit credit is due for LOSO hyper-parameter and prior construction, open trial registration, and the honest reporting that high-variance participants (Fig. 2d) gain little.

major comments (2)
  1. [§4.2, Tables 5/7/8, Discussion] §4.2 defines coverage solely as 0 < B_i,t − I_i,t ≤ C for candidate times and Tables 5, 7 and 8 report only this metric. The paper never shows that windows covered under the adaptive rule produce higher subsequent brushing probability, longer duration, better quality, or any clinical endpoint relative to uncovered windows or the fixed-time arm. While the alignment claim itself is well-supported, the operational assertion that adaptive timing is preferable for digital-health interventions therefore rests on an unvalidated proxy; the abstract, discussion and conclusion should more explicitly bound the claim to coverage/alignment and flag the missing translation step as a limitation of the present evidence.
  2. [§6.1–6.2] In the offline evaluation (§6.1) brushing times are treated as fixed exogenous observations, so the reported coverage gains largely reflect improved prediction of historical times rather than closed-loop adaptation. The simulation variants (§6.2) restore a causal pathway, yet the moderate- and strong-effect modifications (p_target = 0.5/1) are arbitrary and no sensitivity analysis is supplied. A short ablation that varies p_target (or reports coverage under the pure data-driven generative model alone) would strengthen the claim that the ranking of strategies is robust to the strength of the timing–behavior link.
minor comments (5)
  1. [Fig. 2] Figure 2 caption contains the typo “Mean + Uncstartergy”; correct to “Mean + Unc strategy”.
  2. [Table 4] Table 4 reports the User+Offset improvement at C=4 as “(7.81)” without the “+” sign used elsewhere; make the sign convention uniform.
  3. [§4.2–5.2] The notation for candidate intervention time switches between I_t, I_i,t and 𝐼𝑡; pick one and use it consistently from §4.2 onward.
  4. [§6.3] In §6.3 the warm-start period is described as “7 days imes imes imes adaptive timing decisions are not evaluated”; the double “×” appears to be a typesetting artifact.
  5. [Table 10] Appendix Table 10 lists prior variances with a trailing “2” (e.g., “1.20 2”); clarify whether this is a formatting remnant or an intentional scaling factor.

Circularity Check

0 steps flagged

No significant circularity: coverage gains are empirical comparisons on held-out/prospective brushing timestamps, not forced by construction or load-bearing self-citation.

full rationale

The paper's central claim is an empirical finding that adaptive timing (Mean+Offset or Mean+Unc using online BLR/DT/NN) yields higher coverage than fixed user-provided times. Coverage is defined independently (0 < B_i,t - I_i,t ≤ C) on observed or simulated brushing events that are not used to set the candidate I_t under evaluation. Hyperparameters Δ/α and BLR priors are chosen by LOSO on other participants (or the prior trial for the ongoing RCT), then applied online; this is standard CV, not a fit-then-predict of the same quantity. Simulation generative models are fit to historical data and then used to produce new trajectories under three effect variants; adaptive policies are re-run online on those trajectories. The sole self-citation of note is [8] (overlapping authors) for the functional form I_t = μ̂_t - α σ̂_t and an alternative coverage definition; the paper still evaluates both adaptive strategies head-to-head against baseline on its own metric and data, so the citation is not load-bearing for the reported gains in Tables 5/7/8. No equation equates a reported coverage number to a fitted input by construction, no uniqueness theorem is imported, and no known result is merely renamed. The derivation chain is therefore self-contained empirical evaluation.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central claim rests on a small set of free timing hyper-parameters, standard online-learning and survival-model assumptions, and the invented but operationally defined coverage metric. No new physical entities are postulated; the work is empirical and algorithmic.

free parameters (5)
  • offset Δ
    Fixed temporal offset used by User+Offset and Mean+Offset strategies; selected per participant via LOSO to maximize coverage.
  • uncertainty scale α
    Multiplier on predictive standard deviation in the Mean+Unc rule I_t = μ̂_t − α σ̂_t; chosen by LOSO or from prior-trial data.
  • coverage tolerance C
    Hours defining the success window 0 < B − I ≤ C; results are reported for C ∈ {1,2,3,4,5}.
  • BLR prior (μ0, Σ0, α0, β0)
    Population-level Gaussian and inverse-gamma priors for the online Bayesian linear regression, estimated by GEE / LOSO on the earlier trial.
  • p_target in simulation variants
    Target conditional brushing probability inside the C-hour post-intervention window (0.5 moderate, 1.0 strong); hand-chosen to create effect-size scenarios.
axioms (4)
  • domain assumption Brushing time conditional on occurrence within a fixed 12-hour window can be modeled by online regression (BLR / DT / NN) that ignores the joint occurrence process.
    Stated in Section 5.3; right-censored windows are simply dropped for training.
  • domain assumption Candidate intervention times that fall inside a C-hour pre-brushing window are more useful than those that do not.
    Defines the coverage metric used for all claims (Section 4.2).
  • standard math Online Bayesian updates, Hoeffding-tree leaf statistics, and MC-dropout variance are valid uncertainty estimates under sparse sequential data.
    Standard online-learning and approximate-Bayesian machinery invoked in Section 5.3.
  • ad hoc to paper Intervention delivery probability can be fixed at 0.5 independently of the timing decision without invalidating the coverage comparison.
    Used in both simulation and live RCT (Section 6.2–6.3) to match the trial design.
invented entities (2)
  • coverage metric no independent evidence
    purpose: Scalar success criterion that counts whether a candidate intervention time precedes the next brushing event by at most C hours.
    Defined in Section 4.2; all quantitative claims are expressed in terms of this metric rather than clinical endpoints.
  • Mean + Unc timing rule no independent evidence
    purpose: Operational policy I_t = μ̂_t − α σ̂_t that converts online mean and uncertainty into a scheduled prompt time.
    Introduced in Section 5.2 and selected for live deployment; no external validation outside the authors’ trials.

pith-pipeline@v1.1.0-grok45 · 26212 in / 3080 out tokens · 39032 ms · 2026-07-13T02:25:48.896894+00:00 · methodology

0 comments
read the original abstract

A central challenge for digital health interventions aimed at improving habitual behaviors is deciding when to deliver an intervention prompt. For many daily habits, such as tooth brushing or eating, individuals tend to act around a usual time of day, but this timing is not fixed and can shift as routines evolve. When intervention timing is selected in advance and held constant throughout a study, it can gradually become misaligned with behavior, causing interventions to potentially arrive after the behavior has already occurred or too early to be effective. In this work, we address this habitual timing misalignment in digital health interventions by proposing an online decision-making framework that continuously adapts intervention timing as individual behavior patterns change. Rather than treating intervention timing as a static design choice, our framework adapts it over time and integrates it into a sequential process that determines both when and whether to deliver an intervention. Using data from a deployed oral health intervention trial as a case study, we evaluate our approach using both observed data and simulated settings to assess how well different intervention timing strategies align with the timing of brushing events. Across these evaluations, we measure performance using a coverage-based metric that captures whether an intervention is delivered sufficiently close to a subsequent brushing event. We find that adaptive intervention timing consistently improves coverage compared to fixed intervention times based on user-provided input. The proposed framework is currently deployed in an ongoing randomized controlled trial of a digital oral health intervention, with preliminary results that are consistent with and further support our prior evaluations.

Figures

Figures reproduced from arXiv: 2607.09518 by Anna L. Trella, Asim H. Gazi, Bhanu Teja Gullapalli, Susan A. Murphy, Vivek Shetty.

Figure 1
Figure 1. Figure 1: Illustrative examples of participant-level brushing-time patterns for the morning time windows over the 70-day study period, [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustrative examples of participant-level brushing-time patterns and corresponding adaptive intervention times for the morning [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 1 linked inside Pith

  1. [1]

    Samuel L Battalio, David E Conroy, Walter Dempsey, Peng Liao, Marianne Menictas, Susan Murphy, Inbal Nahum-Shani, Tianchen Qian, Santosh Kumar, and Bonnie Spring. 2021. Sense2Stop: a micro-randomized trial using wearable sensors to optimize a just-in-time-adaptive stress management intervention for smoking relapse prevention.Contemporary Clinical Trials10...

    2021

  2. [2]

    Dror Ben-Zeev, Ayesha Chander, Justin Tauscher, Benjamin Buck, Subigya Nepal, Andrew Campbell, and Guy Doron. 2021. A smartphone intervention for people with serious mental illness: fully remote randomized controlled trial of CORE.Journal of medical Internet research23, 11 (2021), e29201

    2021

  3. [3]

    Babin Bhandari, JianChao Lu, Xi Zheng, Sutharshan Rajasegarar, and Chandan Karmakar. 2017. Non-invasive sensor based automated smoking activity detection. In2017 39th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, 845–848

    2017

  4. [4]

    Michael S Businelle, Ping Ma, Darla E Kendzor, Summer G Frank, Damon J Vidrine, and David W Wetter. 2016. An ecological momentary intervention for smoking cessation: evaluation of feasibility and effectiveness.Journal of medical Internet research18, 12 (2016), e321

    2016

  5. [5]

    Yang Chen, Federico JA Perez-Cueto, Agnès Giboreau, Ioannis Mavridis, and Heather Hartwell. 2020. The promotion of eating behaviour change through digital interventions.International journal of environmental research and public health17, 20 (2020), 7488

    2020

  6. [6]

    Hugh HK Fullagar and Jonathan D Bartlett. 2016. Time to wake up: individualising the approach to sleep promotion interventions. 143–144 pages

    2016

  7. [7]

    Yarin Gal and Zoubin Ghahramani. 2016. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. Ininternational conference on machine learning. PMLR, 1050–1059

    2016

  8. [8]

    Asim H Gazi, Bhanu Teja Gullapalli, Daiqi Gao, Benjamin M Marlin, Vivek Shetty, and Susan A Murphy. 2025. SigmaScheduling: Uncertainty- Informed Scheduling of Decision Points for Intelligent Mobile Health Interventions. In2025 IEEE 21st International Conference on Body Sensor Networks (BSN). IEEE, 1–4

    2025

  9. [9]

    Stephanie P Goldstein, J Graham Thomas, Gary D Foster, Gabrielle Turner-McGrievy, Meghan L Butryn, James D Herbert, Gerald J Martin, and Evan M Forman. 2020. Refining an algorithm-powered just-in-time adaptive weight control intervention: a randomized controlled trial evaluating model performance and behavioral outcomes.Health informatics journal26, 4 (20...

    2020

  10. [10]

    Bhanu Teja Gullapalli, Stephanie Carreiro, Brittany P Chapman, Deepak Ganesan, Jan Sjoquist, and Tauhidur Rahman. 2021. Opitrack: a wearable- based clinical opioid use tracker with temporal convolutional attention networks.Proceedings of the ACM on interactive, mobile, wearable and ubiquitous technologies5, 3 (2021), 1–29

    2021

  11. [11]

    Hua Huang and Shan Lin. 2025. iBrush: Toothbrushing Monitoring using Smartwatch.ACM Transactions on Computing for Healthcare(2025)

    2025

  12. [12]

    Predrag Klasnja, Eric B Hekler, Saul Shiffman, Audrey Boruvka, Daniel Almirall, Ambuj Tewari, and Susan A Murphy. 2015. Microrandomized trials: An experimental design for developing just-in-time adaptive interventions.Health Psychology34, S (2015), 1220

    2015

  13. [13]

    Anna M Langener and Nicholas C Jacobson. 2026. Just in Time or Just a Guess? Addressing Challenges in Validating Prediction Models Based on Longitudinal Data.Advances in Methods and Practices in Psychological Science9, 2 (2026), 25152459261418960

    2026

  14. [14]

    Peng Liao, Kristjan Greenewald, Predrag Klasnja, and Susan Murphy. 2020. Personalized heartsteps: A reinforcement learning algorithm for optimizing physical activity.Proceedings of the ACM on interactive, mobile, wearable and ubiquitous technologies4, 1 (2020), 1–22

    2020

  15. [15]

    Inbal Nahum-Shani, Shawna N Smith, Bonnie J Spring, Linda M Collins, Katie Witkiewitz, Ambuj Tewari, and Susan A Murphy. 2016. Just-in-time adaptive interventions (JITAIs) in mobile health: key components and design principles for ongoing health behavior support.Annals of behavioral medicine(2016), 1–17

    2016

  16. [16]

    just-in-time

    Felix Naughton. 2017. Delivering “just-in-time” smoking cessation support via mobile phones: current knowledge and future directions.Nicotine & Tobacco Research19, 3 (2017), 379–383

    2017

  17. [17]

    Ju Lynn Ong, Stijn AA Massar, TeYang Lau, Ben KL Ng, Lit Fai Chan, Daphne Koek, Karen Cheong, and Michael WL Chee. 2023. A randomized- controlled trial of a digital, small incentive-based intervention for working adults with short sleep.Sleep46, 5 (2023), zsac315

    2023

  18. [18]

    Laura Michelle Pape, Annemieke van Straten, Sascha Yuri Struijs, Julian David Karch, Philip Spinhoven, Niki Antypa, and Caring Universities Consortium. 2025. Effects of a guided digital intervention on sleep and mental health outcomes in university students–a randomized controlled trial.SLEEPJ(2025), zsaf357

    2025

  19. [19]

    Abhinav Parate, Meng-Chieh Chiu, Chaniel Chadowitz, Deepak Ganesan, and Evangelos Kalogerakis. 2014. Risq: Recognizing smoking gestures with inertial sensors on a wristband. InProceedings of the 12th annual international conference on Mobile systems, applications, and services. 149–161

    2014

  20. [20]

    WhiteTeeth

    Janneke FM Scheerman, Berno van Meijel, Pepijn van Empelen, Gijsbert HW Verrips, Cor van Loveren, Jos WR Twisk, Amir H Pakpour, Matheus CT van den Braak, and Gem JC Kramer. 2020. The effect of using a mobile application (“WhiteTeeth”) on improving oral hygiene: A randomized controlled trial.International journal of dental hygiene18, 1 (2020), 73–83

    2020

  21. [21]

    Jay Thakkar, Rahul Kurup, Tracey-Lea Laba, Karla Santo, Aravinda Thiagalingam, Anthony Rodgers, Mark Woodward, Julie Redfern, and Clara K Chow. 2016. Mobile telephone text messaging for medication adherence in chronic disease: a meta-analysis.JAMA internal medicine176, 3 (2016), 340–349

    2016

  22. [22]

    Edison Thomaz, Irfan Essa, and Gregory D Abowd. 2015. A practical approach for recognizing eating moments with wrist-mounted inertial sensing. InProceedings of the 2015 ACM international joint conference on pervasive and ubiquitous computing. 1029–1040

    2015

  23. [23]

    Anna L Trella, Kelly W Zhang, Stephanie M Carpenter, David Elashoff, Zara M Greer, Inbal Nahum-Shani, Dennis Ruenger, Vivek Shetty, and Susan A Murphy. 2024. Oralytics reinforcement learning algorithm.arXiv preprint arXiv:2406.13127(2024). Learning When to Intervene on Habitual Behaviors: A Case Study in Oral Health Care 27

    Pith/arXiv arXiv 2024

  24. [24]

    Marcia Vervloet, Annemiek J Linn, Julia CM van Weert, Dinny H De Bakker, Marcel L Bouvy, and Liset Van Dijk. 2012. The effectiveness of interventions using electronic reminders to improve adherence to chronic medication: a systematic review of the literature.Journal of the American Medical Informatics Association19, 5 (2012), 696–704

    2012

  25. [25]

    Ashley Walton, Inbal Nahum-Shani, Lori Crosby, Predrag Klasnja, and Susan Murphy. 2018. Optimizing digital integrated care via micro-randomized trials.Clinical Pharmacology & Therapeutics104, 1 (2018), 53–58

    2018

  26. [26]

    Hang Yuan, Shing Chan, Andrew P Creagh, Catherine Tong, Aidan Acquah, David A Clifton, and Aiden Doherty. 2024. Self-supervised learning for human activity recognition using 700,000 person-days of wearable data.NPJ digital medicine7, 1 (2024), 91

    2024