Spruijt-Metz et al (2022)
Advancing Behavioral Intervention and Theory Development for Mobile Health: The HeartSteps II Protocol
Spruijt-Metz, D. et al. · Int. J. Environ. Res. Public Health, 19(4):2267 · 2022
HeartSteps II is a year-long Micro-Randomized Trial deploying a JITAI for physical activity, designed not only to test intervention effectiveness but to build computational modeling frameworks capable of expressing dynamic, mathematically rigorous theories of health behavior change — a capability that static, box-and-arrow theories have never provided.
The Core Problem
Despite decades of intervention research, available behavior-change interventions remain only moderately effective in the short term and largely fail to sustain long-term maintenance. A key reason is a mismatch between the static, cross-sectional theories on which most interventions are based and the genuinely dynamic, moment-to-moment nature of real health behavior. Classic health-behavior theories — social-cognitive theory, the health belief model, theory of planned action — share a common form: box-and-arrow diagrams depicting time-invariant relationships. They cannot capture the co-interacting, continuously shifting influences that determine whether a person takes a walk or stays on the couch on any given afternoon.
The Two-Part Purpose of HeartSteps II
The study pursues two central goals simultaneously. First, it aims to develop modeling approaches for operationalizing dynamic, mathematically rigorous theories of health behavior — moving beyond static diagrams to computational models that evolve over time. Second, it serves as a testbed for the development of learning algorithms that JITAIs can use to individualize intervention provision in real time at multiple timescales (minute-to-minute, day-to-day, week-to-week).
Why Pure Data-Driven Approaches Are Not Enough
It might seem that reinforcement learning (RL) could bypass the need for behavioral theory entirely by learning purely from data. HeartSteps II rejects this shortcut explicitly: the high noise and sparsity in mHealth data mean that purely data-driven JITAIs are severely limited in what they can learn. Even RL-based systems benefit from theory-grounded models, at minimum as a basis for the initial intervention-provision policy before data accumulates. Theory and data must work together.
Building on HeartSteps I
HeartSteps II extends a series of earlier NIH-funded MRTs that optimized and tested the HeartSteps JITAI in patients with cardiovascular disease. The contextual bandit RL algorithm that drives walking-suggestion randomization in HeartSteps II was directly derived from the HeartSteps I data. HeartSteps II retains the same core JITAI structure but adds quarterly EMA burst weeks, richer theory-driven psychosocial measures, and the dual modeling agenda — making it simultaneously a clinical trial and a theory-development engine.
The Three Micro-Randomized Components
| Component | Frequency | Randomization Logic | Proximal Outcome |
|---|---|---|---|
| Motivational Messages | Once daily (6 a.m.) | Fixed probability: p = 4/7 message sent; four framings (promotion/prevention × active/sedentary), each p = 1/7 | Daily step count |
| Walking Suggestions | Up to 5× daily at participant-chosen times | Contextual bandit (Bayesian Thompson sampler) updated nightly; personalizes probability by context (location, time, weather) | Step count in the 30 min following the decision point |
| Anti-Sedentary Suggestions | Whenever sedentary ≥40 min; max avg 2.5/day budget | Budget-managed algorithm predicts daily sedentary patterns and distributes prompts across the day | Time to sedentary-behavior disruption (time until steps begin accruing) |
■ Core Concepts ■ Study Design & Methods ■ Behavioral Theory ■ Technology & Modeling
Core Concepts
Study Design & Methods
Behavioral Theory
Technology & Modeling
Study Design at a Glance
HeartSteps II is a single-arm, year-long Micro-Randomized Trial targeting sedentary, overweight adults aged 18–65 in Southern California. Following a 1-week passive baseline with the Fitbit only, participants use the HeartSteps app for 12 months while three intervention components are micro-randomized daily. The study was powered to detect a very small decreasing effect (d = 0.05, 90% power) for the least-frequent component (motivational messages), requiring 46 participants; 60 were targeted to allow for 20% attrition, and 95 were ultimately fully enrolled.
Participants & Recruitment
- Target: 60 independent participants at study end; 80 recruited to absorb ~25% attrition; 95 fully enrolled
- Inclusion: BMI 25–45 kg/m², ages 18–65, sedentary lifestyle (IPAQ-screened), Southern California residents, iOS or Android smartphone
- Exclusion: inability to give consent, psychiatric disorder limiting protocol adherence, orthopedic problems preventing walking, significant peripheral neuropathy, vigorous activity exceeding IPAQ moderate threshold
- Fully remote recruitment and onboarding due to COVID-19 (email threads, newsletters, community forums, local media)
- Baseline week: 7 days of Fitbit wear ≥8 h/day with HeartSteps app locked; failure to complete = withdrawal
HeartSteps Intervention Components
- Fitbit Versa Lite: continuous step count and heart rate; auto-detects activity bouts ≥10 min; custom clock face polls steps every 5 min for real-time sedentary detection
- Pull components (always accessible): Dashboard (self-monitoring, goal progress, activity plans), Planning tab (implementation intentions), Activity Log (all-time statistics)
- Push components (notifications): morning motivational messages (6 a.m.), walking suggestions (5 user-chosen times/day), anti-sedentary suggestions (sedentary ≥40 min, max avg 2.5/day), weekly reflection (Sunday 8 p.m.)
- Adaptive weekly activity goal: automatically set to 20 min MVPA above prior week’s achieved minutes, capped at 150 min; participants may manually adjust during weekly review
Randomization Logic
- Motivational messages: fixed daily probability p = 4/7 (any message); each of four framings = p 1/7; no message = p 3/7
- Walking suggestions: Bayesian Thompson sampler (contextual bandit); updated nightly per participant; personalizes to location, time, weather; availability excludes participants who recently received another intervention, walked >250 steps in the past hour, walked >2000 steps in the past 2 hours, or are offline
- Anti-sedentary suggestions: budget algorithm (max avg 2.5/day) predicts sedentary pattern and distributes prompts across the day; runs whenever fewer than 150 steps detected in preceding 40 min; same exclusion criteria as walking suggestions plus Fitbit/clock face non-wear
- All randomization probabilities saved per decision point to enable post-study MRT analyses
Measures — Active (Self-Report)
- Baseline/follow-up: demographics, TIPI personality, perceived stress (PSS), self-efficacy for physical activity, motivation for physical activity (BREQ-modified), neighborhood walkability, social isolation, app usability (MARS), COVID-19 activity disruption
- Daily EMA (morning survey): restedness, busyness, mood, commitment to activity; motivational orientation items (SDT-grounded)
- Weekly EMA (Sunday reflection): social support, enjoyment, loneliness, barriers to activity, confidence in weekly goal
- Activity EMA: post-activity enjoyment, fit into day, social context, motivation (sent with base probability 0.1; probability 1 during burst weeks)
- EMA bursts (quarterly, 7 days): activity questionnaires after every detected bout + walking-suggestion questionnaires (8 mood/affect items at each of 5 daily decision points)
Measures — Passive (Sensor)
- Fitbit: minute-level step counts; minute-level heart rate; auto-detected activity bouts with MVPA minutes; GPS location
- HeartSteps app logs: timestamps of every page view, notification open, and survey interaction
- Limitation: non-interactive notification views (e.g., auto-expanded iOS lock screen) cannot be logged
Analysis Plan
- Primary MRT analyses: centered and weighted least-squares (CWLS) method — a robust generalization of linear regression that handles nested MRT data (decision points within participants) and estimates causal treatment effects; primary aim is average effect + change in effect over time for each component
- Secondary analyses: time-varying moderation by dose, weather, current activity level, and other contextual moderators
- Modeling work: Dynamic Bayesian Networks (flexible, multi-timescale, handles missingness) and Dynamical Systems Models (idiographic, differential-equation-based, fluid analogy); Model-on-Demand (MoD) for nonlinear parameter estimation; DSPSA for efficient search over model features
Strengths
- Year-long duration enables study of long-term dynamics and habituation — rare in mHealth trials
- Dual-purpose design: generates both causal intervention evidence (MRT) and rich ILD for theory building simultaneously
- Adaptive randomization personalizes intervention exposure, avoiding one-size-fits-all dose schedules
- Theory-grounded EMA measures SDT constructs longitudinally, enabling computational modeling of motivation dynamics
- Fully remote design increases accessibility and ecological validity
Limitations
- Single-arm design precludes comparison against a no-intervention control; cannot estimate total HeartSteps efficacy
- Fitbit heart rate data in 24/7 mode can be unreliable, especially for moderate-intensity activities, compromising MVPA accuracy
- High sensor data missingness is expected due to non-wear and technical failures; adds analytic complexity
- English-only app and Southern California geographic restriction limit generalizability
- Social desirability and self-report biases in EMA and activity logging cannot be fully excluded
Five foundational references that anchor the HeartSteps II study’s conceptual and methodological framework. Click any card to expand the analysis.
Nahum-Shani et al. (2015) — Building health behavior models to guide JITAI development
Why this reference matters
This is one of the seminal papers establishing the JITAI concept and articulating why health behavior models are indispensable for JITAI design. It provided a pragmatic framework for identifying the four core components a JITAI must specify: decision points, tailoring variables, intervention options, and decision rules.
Key contribution to this paper’s argument
Spruijt-Metz et al. use this reference to justify the “promise of JITAIs” claim and to frame the gap the HeartSteps II study addresses: that realizing JITAI promise requires better models of behavior, not just better technology. It is also part of the author’s own citation lineage — Nahum-Shani, Hekler, and Spruijt-Metz have collaborated on JITAI theory.
Connection to the wider citation network
This paper sits at the center of a cluster that includes Klasnja et al. (2015) on MRTs and Hekler et al. (2016) on digital behavior change models. Together these three papers form the theoretical-methodological backbone that motivated the HeartSteps program of research.
Klasnja et al. (2015) — Microrandomized trials: an experimental design for developing JITAIs
Why this reference matters
This paper coined and formally defined the Micro-Randomized Trial design, providing the statistical and conceptual rationale for repeatedly randomizing intervention components within participants over time. Without this design framework, HeartSteps II would not exist in its current form.
Key contribution to this paper’s argument
Every claim Spruijt-Metz et al. make about MRT validity, proximal outcomes, and decision-point structure is grounded in this reference. It also justifies the single-arm design: MRT evidence comes from within-person randomization, not between-group comparison.
Connection to the wider citation network
Published simultaneously with the Nahum-Shani JITAI framework paper (ref 8) in a special issue of Health Psychology, this pair of papers jointly launched the modern MRT + JITAI research paradigm. HeartSteps II is one of the most ambitious applications of this paradigm to date.
Klasnja et al. (2019) — Efficacy of contextually tailored suggestions: HeartSteps MRT
Why this reference matters
This is the primary results paper for HeartSteps I — the first MRT in the HeartSteps program, conducted in patients with cardiovascular disease. It established that contextually tailored walking suggestions could causally increase short-term step counts and provided the training data for the RL bandit used in HeartSteps II.
Key contribution to this paper’s argument
HeartSteps II is positioned as a direct scientific successor to this work. The paper provides empirical proof-of-concept that the micro-randomization approach works and that the walking-suggestion component has a detectable proximal effect — justifying the year-long scale-up in HeartSteps II.
Connection to the wider citation network
Together with Liao et al. (2020, ref 13), which derived the personalized RL algorithm from HeartSteps I data, this reference forms a cumulative evidence chain: initial trial (ref 12) → algorithm development (ref 13) → year-long adaptive deployment (HeartSteps II). The progression illustrates the MRT-to-JITAI development pipeline the protocol advocates for.
Liao et al. (2020) — Personalized HeartSteps: a RL algorithm for optimizing physical activity
Why this reference matters
This paper describes the Bayesian Thompson-sampling contextual bandit that personalizes walking-suggestion probabilities in HeartSteps II. It is the technical specification for what makes the system “adaptive” beyond rule-based thresholds — the algorithm that learns each participant’s responsiveness patterns and adjusts future intervention provision accordingly.
Key contribution to this paper’s argument
HeartSteps II does not just deploy a static algorithm — it treats the RL system itself as an object of study and improvement. Liao et al. provides the foundation from which HeartSteps II extends: the algorithm was trained on HeartSteps I data and is now being re-evaluated in a year-long deployment, generating new data to refine it further.
Connection to the wider citation network
This reference bridges the computer science / ML cluster with the behavioral health cluster in HeartSteps II’s citation network. The first author (Liao) is a statistician; the collaborators include Murphy, Klasnja, and the mHealth engineering community — illustrating the genuinely cross-disciplinary coalition behind the paper.
Deci & Ryan (2012) — Self-Determination Theory
Why this reference matters
Self-Determination Theory (SDT) is the primary behavioral theory around which HeartSteps II’s modeling work is organized. SDT’s specification of autonomous vs. controlled motivation, and of the three basic psychological needs (autonomy, competence, relatedness), provides a theoretically grounded set of constructs for the EMA measures and the computational models the study will build.
Key contribution to this paper’s argument
HeartSteps II explicitly commits to modeling SDT constructs dynamically — not just measuring them cross-sectionally. This is the study’s translation challenge: taking a well-established static theory and operationalizing it as a system of differential equations or a DBN that evolves over time.
Connection to the wider citation network
SDT is the only “classic” behavioral theory in the HeartSteps II reference list — and it is cited not as a model to test in the traditional sense, but as the starting raw material for computational re-formulation. This positions HeartSteps II within the broader agenda of upgrading behavioral science to be compatible with the data richness that mHealth now enables.
Murphy, K.P. (2002) — Dynamic Bayesian Networks: Representation, Inference and Learning
Why this reference matters
This dissertation remains the definitive technical reference for Dynamic Bayesian Networks, establishing their formal probabilistic structure, exact and approximate inference algorithms, and learning procedures. It is the foundational text the HeartSteps II modeling team draws on when specifying how DBNs can represent temporal dependencies between health constructs.
Key contribution to this paper’s argument
Spruijt-Metz et al. cite this alongside Murphy’s 2012 textbook to ground the DBN modeling framework in rigorous probabilistic theory. The reference demonstrates that the team is not applying DBNs loosely — they are adopting a well-specified formalism with known inference and learning properties.
Connection to the wider citation network
Together with ref [39], this forms the pure-theory anchor of the engineering/modeling cluster. All the applied modeling work (refs 40–48) implicitly relies on the formal machinery laid out here.
Murphy, K.P. (2012) — Machine Learning: A Probabilistic Perspective
Why this reference matters
A comprehensive graduate textbook that covers probabilistic graphical models, Bayesian inference, and machine learning methods, including DBNs, Kalman filters, and hidden Markov models. It places the specific modeling choices in HeartSteps II within a broader ML landscape, providing the team and readers with a shared technical vocabulary.
Key contribution to this paper’s argument
Cited alongside ref [38] to lend further authority to the DBN approach. The 2012 textbook is more accessible than the 2002 dissertation and serves as the practical implementation reference for the team’s probabilistic modeling language development work.
Connection to the wider citation network
Forms a pair with ref [38] — together these two Murphy references are the theoretical backbone for the DBN thread of HeartSteps II’s modeling agenda, distinguishing it from the dynamical systems thread (refs 40–45) which draws on a completely different engineering tradition.
Riley et al. (2016) — Development of a dynamic computational model of social cognitive theory
Why this reference matters
This paper demonstrates — for the first time — that a major behavioral theory (social cognitive theory) can be translated into a dynamic computational model using control-systems engineering methods. It proved the concept that the fluid-analogy/differential-equation approach was viable for behavioral health, paving the way for HeartSteps II’s more ambitious multi-theory modeling agenda.
Key contribution to this paper’s argument
Spruijt-Metz et al. use this as their primary illustration that dynamic computational modeling of health theories is not merely aspirational — it has already been done with SCT, and HeartSteps II is extending the approach to SDT with richer longitudinal data. Rivera and Hekler, co-authors here, are also co-authors of HeartSteps II.
Connection to the wider citation network
The bridge paper between the behavioral science cluster and the engineering cluster in HeartSteps II’s citation network. It represents the moment the two traditions first collaborated productively, and HeartSteps II is the direct institutional successor to this proof-of-concept.
Rivera et al. (2018) — Intensively Adaptive Interventions Using Control Systems Engineering
Why this reference matters
This book chapter provides two detailed worked examples of applying control-systems engineering principles to adaptive behavioral interventions. It moves beyond the conceptual framing of earlier papers to show what the modeling workflow actually looks like in practice — making it a key how-to reference for the HeartSteps II team’s dynamical systems modeling work.
Key contribution to this paper’s argument
Cited to support the claim that dynamical systems modeling of adaptive interventions is practically achievable, not merely theoretically possible. The two illustrative examples demonstrate the tank-and-valve fluid analogy that HeartSteps II’s modeling section describes.
Connection to the wider citation network
Rivera is a co-author of both this chapter and HeartSteps II, and the chapter draws on the same Arizona State University control-systems lab tradition that underlies refs 42, 43, 44, and 45 — collectively forming the applied dynamical systems sub-cluster within the broader engineering/modeling cluster.
Martin et al. (2020) — A Control-Oriented Model of Social Cognitive Theory for Optimized mHealth Interventions
Why this reference matters
Published in an IEEE engineering journal — unusual for behavioral health research — this paper formalizes a control-oriented model of social cognitive theory as a system of differential equations amenable to closed-loop control design. It demonstrates that behavioral constructs can be expressed in the mathematical language of engineering control systems.
Key contribution to this paper’s argument
Supports the dynamical systems modeling thread in HeartSteps II by providing a concrete, peer-reviewed example of the differential-equation approach applied to a behavioral theory. Its IEEE venue signals that this work is credible from both a behavioral science and an engineering standpoint — exactly the cross-disciplinary credibility HeartSteps II needs.
Connection to the wider citation network
Together with ref [40] (Riley 2016), this is the second major instantiation of the SCT computational modeling program. HeartSteps II builds on both, shifting the target theory from SCT to SDT and the data source from simulated/short-term data to a year-long ILD dataset.
Hekler et al. (2018) — Tutorial for Using Control Systems Engineering to Optimize Adaptive mHealth Interventions
Why this reference matters
Published in JMIR — the field’s leading mHealth journal — this tutorial bridges the technical engineering literature and the broader behavioral health community by explaining control-systems concepts in accessible terms. It is the primary on-ramp for behavioral scientists who want to apply dynamical systems modeling to their own interventions.
Key contribution to this paper’s argument
Hekler is a co-author of HeartSteps II, and this tutorial represents the team’s commitment to making their methodological innovations available to the field — not just using them internally. It also demonstrates that the approach has been reviewed and accepted by the mHealth research community, not just engineering journals.
Connection to the wider citation network
The most widely accessible entry point into the engineering/modeling cluster, making it the reference most likely to be followed up by behavioral health readers of HeartSteps II who want to understand the dynamical systems approach without reading engineering textbooks or IEEE papers.
Korinek et al. (2017) — Adaptive step goals and rewards: a longitudinal growth model for a smartphone walking intervention
Why this reference matters
This paper provides empirical evidence that adaptive step goals — dynamically adjusted based on individual performance trajectories — can support sustained physical activity in a smartphone-based intervention. It directly validates one of HeartSteps II’s core design features: the adaptive weekly MVPA goal algorithm.
Key contribution to this paper’s argument
Supports the adaptive goal-setting component of HeartSteps II by demonstrating that the approach has been tested and shown to work in a real-world walking intervention. Rivera and Hekler are also co-authors, maintaining the cross-team lineage of this research program.
Connection to the wider citation network
Sits at the intersection of the dynamical systems modeling cluster and the HeartSteps intervention design literature — it is both a methods validation paper and an empirical evidence paper, making it one of the cluster’s most directly applied references.
Rivera, Pew & Collins (2007) — Using engineering control principles to inform the design of adaptive interventions
Why this reference matters
This is the earliest paper in the engineering/modeling cluster, representing the first published argument that control-systems engineering principles — developed for managing industrial processes — could be applied to designing adaptive behavioral interventions. It is the original conceptual bridge between the two disciplines.
Key contribution to this paper’s argument
Provides the intellectual genealogy of the dynamical systems approach, showing that it was not invented for HeartSteps II but has a 15-year development history. Its drug and alcohol dependence context also signals that the approach generalizes across behavioral health domains beyond physical activity.
Connection to the wider citation network
The earliest node in the dynamical systems sub-cluster; everything in refs 40–44 builds on the conceptual framework first articulated here. Rivera’s presence as a co-author across refs 41, 42, 43, 44, and 45 — and as a co-author of HeartSteps II itself — shows remarkable continuity of scientific vision across nearly two decades.
Stenman, A. (1999) — Model on Demand: Algorithms, Analysis and Applications
Why this reference matters
This dissertation introduced the Model on Demand (MoD) estimation framework — a local, nonlinear system-identification method that fits models to data from a temporally local neighborhood rather than imposing a single global model. It is the foundational reference for the nonlinear parameter estimation approach HeartSteps II plans to apply to its dynamical systems models.
Key contribution to this paper’s argument
MoD is cited as HeartSteps II’s solution to the nonlinearity problem: behavioral dynamics are unlikely to follow a single global linear model across a full year. By fitting locally adaptive parametric models, MoD allows the dynamical systems framework to capture the shifting, context-dependent nature of behavior without requiring an intractably complex global specification.
Connection to the wider citation network
Paired with ref [47] (Braun et al. 2001), which extended MoD to non-linear process systems. These two references form the numerical methods sub-cluster within the engineering/modeling cluster — distinct from both the DBN cluster (refs 38–39) and the control-theory cluster (refs 40–45).
Braun, Rivera & Stenman (2001) — A ‘Model-on-Demand’ identification methodology for non-linear process systems
Why this reference matters
This paper extended Stenman’s MoD framework to nonlinear process systems — demonstrating its practical value for engineering applications where system dynamics change over time and cannot be captured by a single fixed model. It validates MoD as a robust tool for exactly the kind of time-varying, nonlinear system HeartSteps II aims to model in human behavioral dynamics.
Key contribution to this paper’s argument
Rivera is co-author of both this paper and HeartSteps II, providing direct continuity between the engineering process-control literature and the behavioral mHealth application. The paper shows that MoD is not borrowed speculatively — it has been validated in applied engineering contexts and is now being transferred to health behavior modeling.
Connection to the wider citation network
Completes the MoD pair with ref [46]. Rivera’s co-authorship here and across the dynamical systems cluster (refs 41–45) makes him the central node bridging process-control engineering and behavioral health science within HeartSteps II’s citation network.
Wang & Spall (2011/2014) — Discrete Simultaneous Perturbation Stochastic Approximation for resource allocation in public health
Why this reference matters
This conference paper introduces Discrete Simultaneous Perturbation Stochastic Approximation (DSPSA) — a gradient-free optimization algorithm designed for discrete parameter spaces. It is the search procedure HeartSteps II plans to use for efficiently identifying optimal model features and MoD design parameters from the large, noisy ILD dataset without requiring a closed-form objective function.
Key contribution to this paper’s argument
DSPSA solves a practical bottleneck in the dynamical systems modeling workflow: determining which model features and MoD parameters to use requires searching a large discrete space, and standard gradient-based methods fail when the objective function cannot be expressed in closed form. DSPSA approximates gradients from simultaneous two-sided perturbations, making the search tractable.
Connection to the wider citation network
The most technically specialized reference in the engineering/modeling cluster — it addresses a narrow but critical computational problem in the modeling pipeline. Its public-health application context in the original paper is an unusual point of connection between the optimization algorithms literature and health intervention research, mirroring HeartSteps II’s own cross-disciplinary ambition.
HeartSteps II’s citation network divides into three distinct clusters, each representing a different discipline the paper synthesizes. The behavioral health / JITAI cluster (refs 6–10) establishes the theoretical motivation — why static theories fail and what JITAIs require. The MRT/statistics cluster (refs 11, 35–37) provides the experimental and analytic design scaffolding, anchored by the foundational Klasnja 2015 MRT paper. The engineering/modeling cluster (refs 38–48) is unusual for a behavioral health journal paper: it includes Murphy’s DBN textbooks, Rivera’s engineering-control-systems papers, and specialized numerical methods (MoD, DSPSA). The bridge between clusters is the HeartSteps program itself — refs 12 and 13 (HeartSteps I efficacy + RL algorithm) connect the behavioral trial evidence to the computational modeling agenda. Notably, many co-authors of HeartSteps II also appear in the reference list, indicating that the paper represents a large multi-disciplinary research coalition (USC, UMass, Northeastern, Arizona State, UCSD, Michigan) publishing their joint roadmap, not a single team’s retrospective report.
Click an answer to reveal feedback. Each question locks after answering.
The real frontier of mHealth is not whether interventions work, but how to build dynamic, computationally rigorous theories of behavior change that can drive JITAIs capable of adapting to each person across the full complexity of daily life — and HeartSteps II is the first study explicitly designed to advance both goals simultaneously.
-
Static Theories Cannot Drive Adaptive Interventions
The fundamental critique in HeartSteps II is directed at the entire classical health-behavior theory canon — social-cognitive theory, health belief model, theory of planned action — all of which share a box-and-arrow, time-invariant structure. These theories were built on cross-sectional data and cannot capture the moment-to-moment co-interactions that actually determine behavior. Effective JITAIs require theories expressed as dynamic computational models, not diagrams.
-
The MRT as Theory Engine, Not Just Efficacy Test
Most trials ask “does the intervention work?” HeartSteps II asks that question and simultaneously asks “what do the patterns of response tell us about how behavior works?” The ILD collected across 95 participants over a year — with thousands of micro-randomization decision points — is treated as a dataset for building and testing dynamic computational theories, not just for estimating an average treatment effect.
-
Burden Management Is an Algorithmic Problem
HeartSteps II treats intervention burden as a first-class design challenge. The anti-sedentary suggestion budget (maximum average 2.5 per day), the contextual bandit’s habituation response, and the availability windows that prevent “piling on” after recent activity all reflect the insight that an intervention people ignore or resent provides no benefit. Burden management is not a side constraint — it is baked into the randomization algorithms themselves.
-
Theory and Data Must Co-Evolve
HeartSteps II explicitly rejects both pure theory (untestable static models) and pure data (RL without behavioral structure). The modeling frameworks — DBNs and dynamical systems models — are designed to be expressive enough to encode theory-grounded dynamic hypotheses and then to be estimated and revised with actual ILD. The study’s contribution is as much the methodology for bridging theory and data as it is the specific HeartSteps results.
-
Timescale Is Everything
One of HeartSteps II’s most underappreciated contributions is the insistence that behavior is multi-timescale: mood fluctuates within minutes; sedentary bouts resolve within hours; weekly goals operate over days; motivation for exercise may shift over months. Any model of behavior change must specify which timescale each construct operates on and how constructs at different timescales interact. The EMA burst design was created precisely to capture this structure.
-
Cross-Disciplinary Collaboration Is Not Optional
HeartSteps II required behavioral scientists (USC, UCSD), computer scientists (UMass), engineers (Arizona State), and clinical informatics researchers (Michigan) working together from the study’s inception. The modeling approaches — borrowed from engineering control theory and Bayesian machine learning — could not have been brought to behavioral health without a team structured to bridge these worlds. For mHealth researchers, this signals that the next generation of JITAI development requires organizational change, not just new tools.
Why do the authors take aim at classical health-behavior theories without engaging deeply with what those theories actually say?
The Move the Paper Makes
The critique of classical health-behavior theories in HeartSteps II is structural rather than substantive. Spruijt-Metz et al. do not engage with what social-cognitive theory, the health belief model, or the theory of planned action actually say. They make a single move — “these theories share a common form: box-and-arrow diagrams depicting time-invariant relationships” — and treat that formal observation as sufficient grounds to dismiss the entire tradition. From that point forward, the theories serve as a foil for the modeling agenda, not as objects of inquiry in their own right.
Three Reasons the Paper Works This Way
The genre excuse. This is a protocol paper. Its purpose is to justify a study design, not to conduct a theoretical review. Deep engagement with the content of SCT or the health belief model would be beside the point of the article’s function. The theory critique is instrumental — it exists to motivate the modeling agenda, nothing more.
The critique is about form, not content. The authors are not asking whether SCT is right about self-efficacy, or whether the health belief model correctly identifies perceived barriers. They are asking a narrower engineering question: can this theory be expressed as a system of differential equations or a DBN? From a modeling standpoint, most classical theories cannot — they were not written to be computationally operationalizable. That is a legitimate formal critique that does not require reading the theories deeply.
Strategic efficiency. A protocol paper needs a clean narrative arc: problem → gap → study. Nuanced engagement with whether each theory does or does not contain implicit dynamics would destabilize that arc. The straw man is doing load-bearing work.
Where the Argument Overreaches
Social-cognitive theory in particular — Bandura’s work — does describe feedback loops, reciprocal determinism, and the iterative, context-sensitive nature of self-efficacy. It is not obviously as static as the paper implies. SDT itself, which HeartSteps II adopts as its theoretical anchor, distinguishes between autonomous and controlled motivation in ways that already imply dynamics: need satisfaction shifts over time, contexts shift motivational regulation. The claim that “all classical theories are static boxes and arrows” is a straw man, and the authors never confront that tension — including regarding their own chosen theory.
The Deeper Conflation
The paper conflates two distinct problems: the problem of measurement (we did not have dense enough data to test dynamic theories) and the problem of theory (theories were written statically). These are not the same thing. A behavioral theorist could reasonably respond: “SDT always implied dynamics — we simply could not test them. You have not discovered a new theory; you have finally built the instrument that the theory always needed.” The authors’ framing positions their engineering modeling approach as generating new theory, when it might be better understood as finally providing the empirical machinery to test theory that was already more dynamic than their caricature of the classical canon suggests.
The authors take aim at the models’ form — the box-and-arrow structure — because that is the specific thing their computational framework improves on, and deep engagement with the models’ actual content would complicate the clean narrative they need to justify their study design. Whether that is intellectually honest depends on how charitable you are willing to be about the protocol-paper genre.