Physical activity (PA) promotes long-term health and wellness by providing physiological benefits such as weight management, blood glucose and blood pressure control, and reduction of the risk of chronic diseases such as heart disease and diabetes. In addition, it enhances cognitive functions (eg, improved alertness, attention, and memory) [1-3] and emotional well-being (eg, elevated mood and confidence and reduced anxiety and stress) [1,3,4]. Despite the many documented benefits of PA for health and wellness, increasing PA at a population level remains a challenge. Research shows that, in 2020, only 24.2% of American adults met the recommended levels of PA (ie, 150 minutes of moderate-intensity PA and 2 days of muscle-strengthening activity per week) [5]. In addition, PA levels decline significantly with age, particularly after midlife, contributing to the heightened prevalence of cardiometabolic risks—including hypertension and insulin resistance—among older adults [5]. With an aging population and extended life expectancy, promoting sustained PA in middle-aged and older adults represents an increasingly significant public health challenge.

With a high penetration of smartphones and wearables across a broad range of populations, there have been many interventions that have attempted to provide support for PA through mobile health (mHealth) apps [6]. Compared to the more traditional modes of intervention, such as in-person cognitive behavioral therapy, the key benefits of mHealth apps are high accessibility, low cost, support for user autonomy [7-9], trackability of progress over time with continued intervention support in daily settings, and the potential for adaptation and tailoring to individual needs [10-12]. However, despite these advantages, mHealth interventions continue to face persistent challenges in sustaining long-term engagement, particularly among middle-aged and older adults [13,14]. In this population, psychological barriers—such as low intrinsic motivation, reduced self-efficacy, fear of injury, negative age stereotypes, and previous negative experiences with PA [13-16]—are often compounded with age-related physical limitations (eg, chronic conditions and fatigue) and technological barriers (eg, lower digital literacy and poor usability for older adults) [17,18]. These overlapping challenges can create a self-reinforcing cycle of disengagement. To break this cycle, sustaining long-term engagement among older adults requires a fundamentally different design approach—one that is more resilient to age-related changes, fluctuating motivation, and psychological barriers.

Scholars in the mHealth space have relied on behavior change theories and models such as self-determination theory, the Health Belief Model, the theory of planned behavior, and social cognitive theory to understand the determinants of health behaviors and guide the design of intervention strategies. These theories provide valuable insights into the psychological factors that drive intention and motivation for behavior change, such as perceived benefits, self-efficacy, and autonomy. By identifying the why behind behavior, these frameworks have informed the development of theory-based strategies to create mHealth interventions tailored to psychological and behavioral needs.

Despite the theoretical rigor of traditional behavior change models, their heavy reliance on self-regulatory strategies such as goal setting and planning assumes that health behaviors are primarily driven by reflective cognitive processes. However, growing evidence highlights the significant role of subconscious and automatic processes, including implicit attitudes, emotions, and habits, in shaping behavior [19-21]. These implicit mechanisms operate below conscious awareness, shaping perceptions and actions through repeated experiences and associations [22]. In recognition of these limitations, health psychology has increasingly shifted toward dual-process theories, which integrate both reflective and automatic processes to better capture the complexity of behavior change. This shift is particularly relevant for PA adherence, where motivation type plays a critical role. Research shows that individuals driven by intrinsic motivation (eg, enjoyment) sustain activity longer than those driven by external motives (eg, appearance or peer approval) [23,24]. Affective states—feelings of pleasure or displeasure—are central to intrinsic motivation, shaping behavioral choices through subconscious affective processing [23,25-27]. Fostering positive affective experiences with PA has proven especially promising for encouraging sustained engagement, particularly among sedentary individuals [27].

The affective-reflective theory (ART) of physical inactivity and exercise is a dual-process framework for understanding how both automatic and reflective processes shape exercise behavior. Central to ART is the concept of affective valuations—momentary emotional responses to exercise behavior based on previous experiences—which arise automatically with minimal cognitive effort [20,28]. These fast, automatic responses interact with slower, reflective processes that involve conscious reasoning through working memory [20,28-30]. According to ART, repeated experiences with PA form affective valuations—instant reactions of attraction or aversion triggered by thinking about or preparing to engage in exercise. Over time, these valuations, reinforced through reflection on one’s experiences, develop into affective attitudes—more stable emotional evaluations of the behavior (eg, likes or dislikes) [22]. Together, affective valuations and attitudes—referred to collectively as affective associations—play a critical role in shaping one’s behavior, performance, and exercise decision-making [30-34]. Therefore, sustaining positive affective associations (eg, enjoying PA) is essential for strengthening intrinsic motivation and intention to engage in regular exercise [20,28,29].

However, despite the close relationship between affective associations and the motivation and intention to engage in PA, mHealth interventions largely overlook this critical determinant of PA. Recent mHealth interventions have attempted to track and monitor affective states, but most of them fall short of directly influencing the core affective association processes that underpin the intrinsic motivation for engaging in PA [30,31,35]. While a few recent studies have shown promising results in modifying implicit attitudes within controlled laboratory settings [36-38], this approach has not been widely tested in mHealth interventions that support users in the dynamic, real-world contexts of everyday life. One notable exception is HeartPhone [39], an mHealth intervention that presented emotionally positive images of people exercising on users’ smartphone lock screens over an 8-week period, resulting in improved affective judgments (ie, enjoyment, intrinsic motivation, and integrated regulation).

Although the research that specifically focuses on improving affective associations with PA is still relatively nascent, there is extensive literature on messaging that indicates that message framing can effectively impact motivation for healthy behaviors such as PA [40-43]. Affective content can play an important role in such messaging interventions, supporting the idea that messaging can engage automatic affective processing and elicit subconscious emotional responses. For instance, a study using affective message framings demonstrated promising results in influencing exercise behavior and shifting attitudes toward PA [44]. Previous research has also shown that emotional imagery enhances affective arousal, valence, and attention [45,46], suggesting that images might be a powerful way of actuating affective processes in messaging interventions.

How best to leverage these insights in mHealth interventions to effectively influence affective processes toward PA, particularly among middle-aged to older adults, remains an open research challenge. To begin to address this gap, this paper presents a proof-of-concept study of the WalkToJoy mHealth intervention. Grounded in ART, this message-based mHealth intervention was designed to improve middle-aged and older adults’ affective valuations of and affective attitude toward walking (ie, a moderate PA) to support their ability to sustain PA in the long term.

WalkToJoy was designed to be a low-cost, scalable intervention that can augment the standard reflective self-regulatory strategies, such as self-monitoring and goal setting, found in commercial activity trackers (Fitbit in this case) with SMS text messages aimed at improving middle-aged and older adults affective associations with PA. Theoretically, the WalkToJoy messages implemented 2 key strategies to improve affective associations: evaluative conditioning [37,38] and belief updating [47]. Evaluative conditioning [37,38] involves repeatedly pairing an inherently positive (for promotion) or negative (for cessation) stimulus with representations of the behavior one is trying to impact. WalkToJoy used evaluative conditioning by pairing a PA-related stimulus (ie, a walking suggestion) with smile-inducing animated images of cute animals in GIF images to subconsciously enhance users’ affective valuations of walking. Unlike HeartPhone [39], which facilitated evaluative conditioning of PA by showing attractive images of people being active on the lock screens of users’ phones, WalkToJoy implements evaluative conditioning by embedding positively valenced imagery within push notifications that prompt walking. This design embeds evaluative conditioning into a behavioral cue, enhancing its potential impact on walking behavior. We hypothesized that repeatedly pairing, via push messages, suggestions to go for a walk with an intrinsically positive emotional experience—smiling induced via a GIF image of a cute animal—would, over time, strengthen associations between walking and positive emotions, resulting in increased walking behavior as a distal outcome (Figure 1).

Belief updating is a strategy that aims to reshape individuals’ expectations and previous beliefs about a behavior [47]. According to ART, beliefs about PA—including affective attitudes—can be reshaped by prompting individuals to notice and reflect on the positive experience of being active [48-50]. Previous studies have emphasized the effectiveness of positively framed messages and action planning on aging adults’ PA [41,51-53]. WalkToJoy builds on this research by specifically targeting belief updating by making enjoyable aspects of walking more cognitively accessible and by encouraging PA exploration by planning more enjoyable walking experiences. Mechanistically, these components were intended to evoke reflection on the discrepancy between the positive aspects of walking, which planning and salience components supported, and one’s previous attitudes toward walking (Figure 2). The intended proximal outcomes were for participants to report greater enjoyment and anticipation of their daily walks, which, over time, would enhance affective attitudes toward walking and increase the likelihood of engaging in PA (distal outcome).

To assess the feasibility of this intervention and gather initial evidence for its mechanistic impacts, we conducted a 6-week proof-of-concept study of WalkToJoy. This study was conducted within the Multiphase Optimization Strategy framework [54], and it used a hybrid experimental design that combined a factorial experiment with a microrandomized trial (MRT) to investigate the potential of each of WalkToJoy’s 3 experimental components to influence affective associations and walking behavior. In addition, we interviewed a subset of participants to understand their experiences with the intervention and how they incorporated walking into their daily lives.

A proof-of-concept study of the WalkToJoy intervention was conducted as a 6-week full factorial experiment with 3 factors (ie, for a total of 8 conditions) and an embedded MRT [11,13,41,55]. The primary aim of this study was to assess the feasibility of the WalkToJoy intervention package as a whole for improving affective associations with walking. The secondary aim was to assess how each of the 3 components contributed to an improvement in affective associations with walking and the walking behavior itself. In addition, this study investigated the impact of the microrandomized components on their corresponding proximal outcomes to inform the decision rules (eg, timing and context) for the delivery of those components. This study was retrospectively registered with the Open Science Framework [56] after data collection and analysis were completed.

The target study population consisted of adults aged ≥40 years residing in the United States with access to a smartphone and a Fitbit (Google) activity tracker. Recruiting was conducted through MHealthy and UMHealthResearch, 2 web-based platforms at our university focused on wellness and health research, respectively. Recruitment emails containing a detailed study description, contact information, and a link to the Qualtrics eligibility screener (Qualtrics International Inc) were sent to potential participants who expressed interest in the study.

The eligibility screening involved a conditional 2-item self-report measure assessing participants’ satisfaction with their current activity levels. The first question asked the following—I want to be more physically active—with responses on a 5-point Likert scale ranging from strongly disagree (1) to strongly agree (5). Participants who scored ≥3 were then shown the second question—To be more active, I want to increase how much I walk—also rated on the same 5-point scale. Participants who scored ≥3 on both items (ie, indicating low satisfaction with their current activity level and strong motivation to increase PA through walking) were invited to participate.

This proof-of-concept study intentionally included participants across a wide range of baseline activity levels rather than restricting recruitment to sedentary individuals to assess the feasibility of WalkToJoy broadly. In addition, due to feasibility constraints (ie, limited funding), we recruited individuals who already owned a Fitbit and expressed interest in becoming more active but struggled with maintaining enjoyment and consistency. Eligible participants were subsequently asked to confirm their intent to take part and complete an informed consent form. Upon completion, they provided their email address and phone number to receive further instructions for study enrollment. The inclusion criteria are summarized in Textbox 1.

This study was approved by the Health Sciences and Behavioral Sciences Institutional Review Board at the University of Michigan (HUM00217566). All participants provided informed consent electronically via a secure Qualtrics survey, which included study details, contact information, and information about voluntary participation and withdrawal rights. Participants could withdraw at any time and request deletion of their data, which was honored immediately. To ensure confidentiality, each participant was assigned a unique ID, and all survey and Fitbit data were linked only to these IDs. Personally identifiable information (eg, names, emails, and phone numbers) was stored separately in secure, access-restricted databases. A subset of participants was invited to optional Zoom interviews, which were recorded, transcribed, deidentified, and securely stored. Participants were compensated up to US $52 based on their level of participation.

Participation in the study involved several steps: onboarding through the WalkToJoy web application, a 1-week baseline assessment, a 6-week intervention period, a postintervention assessment, and an optional semistructured interview. The overall study flow is illustrated in Figure 3 and described in detail in the following sections.

Due to COVID-19 restrictions, the study was conducted entirely remotely. Upon providing consent, eligible participants were randomly assigned to 1 of 8 study conditions and received a participant ID, account token, and detailed instructions via email. Using these credentials, participants logged into the WalkToJoy web application, accessed through a mobile web browser on their smartphones, to complete the onboarding tasks and the baseline survey.

During onboarding, participants personalized their experience by entering their preferred name and typical waking and sleeping hours, connecting their Fitbit account, and configuring Fitbit and phone settings. Figure 4 provides an overview of this onboarding workflow.

As part of the baseline survey, participants provided demographic and psychosocial information, including gender; motivation for PA; and baseline measures of affective attitudes toward and affective valuations of walking, exercise self-efficacy, and body awareness. Fitbit data were collected for a minimum of 7 days before the start of the intervention to establish baseline step counts. Participants who wore their Fitbit for at least 3 days (≥8 hours per day) during the baseline week transitioned into the intervention phase the following Monday. Those who did not meet this wear threshold were sent reminders and given up to 7 additional days to meet the baseline criteria.

The WalkToJoy intervention package was evaluated for feasibility and adherence using a pretest-posttest study design. To maximize what we could learn from this study and assess the preliminary efficacy of each intervention component in improving participants’ affective valuations and attitudes toward walking over the 6-week study, a 3 × 2 factorial design was implemented. For walking suggestions, the factor levels were receiving walking suggestions with cute animal GIF images (factor on) or receiving walking suggestions without the GIF images (factor off). For salience messages, the factor levels were receiving the salience messages (on) or not receiving the salience messages (off). For planning, the factor levels were receiving the planning prompts (on) or not receiving the planning prompts (off). Upon enrollment, participants were randomized across the 3 factors, resulting in assignment to 1 of 8 unique intervention conditions (Table 1).

In addition, the study included microrandomization of the walking suggestions and salience messages to evaluate their near-term impact on step counts within a 4-hour window and on daily affective reflection about their recent walks, respectively.

The WalkToJoy system microrandomized the delivery of GIF images and text-only walking suggestions at 2 decision points per day (ie, morning and afternoon) determined by participants’ wake-up times and time zones, with a 50% probability of activation at each decision point. This meant that participants were expected to receive, on average, 1 walking suggestion per day. Randomization at each decision point was independent of intervention provision at previous decision points. Given the study duration of 42 days, each participant could be randomized up to 84 times.

For participants in the salience conditions, salience messages were microrandomized with a 50% probability once per day at noon based on participants’ time zone. As a result, participants would receive, on average, a salience message every 2 days. Participants in the nonsalience conditions did not receive any salience messages.

There was no embedded MRT for the planning component. Participants in the planning conditions received a planning prompt each Monday morning once a week with an activation chance of 100% but randomly received 1 of the 2 types of planning prompts with 50% probability: a prompt to try a new planning strategy or a prompt to keep the same strategy as the previous week. After completing the planning prompt, participants received two reminders during the week (on Tuesdays and Thursdays) reiterating the specific plans they had written. Participants in the nonplanning group did not receive any planning prompts or reminder messages.

In line with the Obesity Related Behavioral Intervention Trials model for behavioral treatment development [64], this study was designed as a proof-of-concept trial with the goal of evaluating whether the WalkToJoy intervention package showed promise that it could meaningfully impact a key affective mechanism—affective attitudes—that mediate sustainment of PA. Therefore, our aim was to explore whether the intervention showed signs of potential efficacy to guide future refinement and scaling efforts. This approach aligns with the Multiphase Optimization Strategy framework, which encourages early-phase experimentation to assess feasibility and identify promising intervention components to inform package optimization before larger trials, such as randomized controlled trials [54,65]. At the time of planning the study, there was no clear empirical evidence for what magnitude of change in affective attitudes should be considered clinically meaningful. As a result, we relied on an observation that a change of half an SD constitutes a meaningful impact across a range of health measures [66,67]. Starting with the SD for affective attitudes of 1.21 reported by Kiviniemi et al [60], we calculated that an impact of half an SD would correspond to a moderate standardized effect size of 0.5. To detect this effect size with 90% power and type I error rate of .05, we calculated that we needed a minimum of 36 participants. A recently published paper by Dunton et al [68] has since provided further support for these assumptions by finding that even a change of a third of an SD in affective response to exercise can meaningfully improve exercise behavior. Assuming a 10% attrition, as in our recent mHealth studies, the minimum required sample size was 40 participants. We note that this sample size would have also enabled us to detect a small standardized effect of 0.12 on the proximal outcome of the microrandomized walking suggestions and a standardized effect of 0.17 on the proximal outcome of the microrandomized salience messages, both with 90% power at the α level of .05.

In addition, while step count was included as an exploratory behavioral outcome, this study was not powered to detect statistically significant between-group differences at the conventional α level of .05. Rather, the aim was to identify potential signs of effect that could inform future hypothesis-driven trials, consistent with the exploratory goals of early-phase behavioral intervention research.

To examine changes in affective attitudes (primary outcome), affective valuations (secondary outcome), and step counts (exploratory outcome) between the start and end of the study, we used descriptive statistics and paired 2-tailed t tests. In addition, to assess the main effects of each of the 3 baseline-randomized factors on the weekly assessments of the affective measures as well as daily measures of step count, we used the generalized estimating equations (GEE) model. This allowed us to assess the overall (average and across time) effect of delivering the GIF version of the walking suggestions versus delivering the non-GIF, text-only version of the suggestions; the overall effect of delivering salience messages about positive aspects of walking versus not delivering salience messages; and the overall effect of prompting weekly planning of enjoyable walks versus not prompting planning. To decrease noise, covariates in our analyses included time in the study and baseline values of the outcome variables (affective measures and step counts, respectively). For the GEE analyses of the factorial experiment, 1-sided P values were used to test directional hypotheses that the presence of each intervention component (GIF prompts, salience messages, and planning prompts) would lead to improvements in affective outcomes and step count. This decision reflects the study’s proof-of-concept nature, where the aim was to maximize sensitivity to detect expected beneficial effects rather than assess bidirectional changes.

The analyses of microrandomized components examined their impact on their respective proximal outcomes. For walking suggestions, our primary interest was the average effect of sending a prompt versus not sending a prompt on the subsequent 4-hour step count. The 4-hour window of the proximal outcome was anchored in the time of message randomization. For the salience messages, our primary interest was the average effect across time of sending a salience message versus not sending a salience message on the affective reflection (ie, enjoyment) on the day’s walking experience and anticipated affect toward upcoming walks measured at 8 PM each evening. To assess these effects, we used the centered and weighted least squares method [12,69], the standard statistical approach for examining data from MRTs. This approach enables robust estimation of causal treatment effects while accommodating longitudinal data correlations through the nesting of outcomes within participants [69]. To reduce noise, the analyses of the walking suggestions included a covariate of the 1-hour prerandomization step count.

In addition to the statistical analyses, the data from the postintervention interviews were thematically analyzed using an inductive approach, beginning with a detailed review of each transcript to identify patterns and assign unique codes. Emergent themes and subthemes were synthesized through systematic rereading and joint discussions among the research team to ensure rigor and resolve discrepancies. The analysis concluded when data saturation was reached, indicated by the absence of new themes or variations.

Participant recruitment began in May 2023, with 77 individuals initially consented via the university online platform. However, of these 77 individuals, 21 (27%) were lost to follow-up during onboarding to the WalkToJoy app and never progressed to the baseline phase. In total, 56 participants were onboarded and randomized to the intervention groups, but 7 (12%) were lost to nonadherence (ie, they did not wear the Fitbit activity tracker during the baseline phase). Ultimately, 49 participants received the intervention, but 4 (8%) more were withdrawn due to nonadherence to wearing the Fitbit activity tracker (ie, 10 consecutive days of no activity data despite reminder messages and emails; n=3, 75%) or because their device broke (n=1, 25%). A total of 45 participants successfully completed the study (Figure 7). Data collection concluded in September 2023. The demographic characteristics of the 49 participants who received the intervention are summarized in Table 3.

The recruitment for exit interviews involved a purposive sampling strategy to select participants who (1) were allocated to 1 of the 8 intervention arms and (2) had recently (within the previous 30 days) finished taking part in the study. Of the 14 participants invited, 9 (64%) completed the interview process. Notably, an additional participant was invited from the subgroup that received the comprehensive intervention package comprising all 3 components. The demographic details of the interview participants are summarized in Multimedia Appendix 2.

Of the 45 participants who completed the study, 40 (89%) completed the affective measures for the primary and secondary outcomes (affective attitude and affective valuations) at both baseline and the final intervention week; pretest-posttest statistics are reported for this subset. In addition, 93% (42/45) of the participants provided valid Fitbit wear data—defined as at least 3 days per week with ≥8 hours per day of wear, confirmed via heart rate minute data—at both baseline and the final intervention week (week 6). Pretest-posttest analysis for average daily step count is reported for this subset.

Among the 45 participants enrolled in the study, adherence to the weekly outcome assessments varied across the 6-week intervention period. A total of 270 weekly assessments were expected across all participants (45 participants × 6 weeks = 270 total assessments). Overall, 94.1% (254/270) of the expected assessments were completed, indicating a high level of adherence. The percentage of participants who completed all 6 assessments was 76% (34/45), whereas 98% (44/45) completed at least 4 out of the 6 assessments.

The total possible decision points for the microrandomized GIF and text-only walking suggestions were 3240 (45 participants × 36 days × 2 per day), where participants received suggestions 2 times per day excluding Sundays. Of these 3240 possible decision points, 25 (0.77%) were not randomized due to technical issues, resulting in a total of 3215 (99.23%) decision points randomized. Across the 3215 decision points, both heart rate data and step count data for the proximal outcome were completely missing in 217 (6.75%). This missingness likely resulted from device nonwear during the entire proximal outcome time frame (ie, a 4-hour window). Among the remaining 2998 decision points, heart rate data were partially missing in 335 (11.17%), indicating intermittent wear or data transmission issues during that window. To address wear adherence missingness, an imputation method was applied for the MRT analyses of the walking suggestions. For each minute within the 4-hour proximal outcome window following a decision point, heart rate data were assessed within a 5-minute radius. If data were detected within this window, the minute was classified as an adherent minute even if no heart rate data were available for the exact minute.

The total possible number of decision points for the microrandomized salience messages was 792 (22 participants × 36 days × 1 per day), where the 22 participants randomized to the salience group received salience messages once per day excluding Sundays. Of these 792 possible decision points, 46 (5.8%) were not randomized due to technical issues, resulting in a total of 746 (94.2%) decision points randomized. Of the 746 decision points, 381 (51.1%) were randomized to send salience messages, and 365 (48.9%) were randomized to not send any salience messages that day. Of the 381 activated salience messages, 1 (0.3%) failed to be delivered due to mobile carrier network issues. In addition, across these 745 decision points, missingness occurred in 103 (13.8%) for the measure of affective reflection and anticipated affect when participants failed to complete the daily end-of-day survey for that day. Therefore, the MRT analysis of the salience factor was conducted on the 642 decision points with a valid proximal outcome.

As walking prompts, salience messages, and planning were baseline randomized in a factorial experiment, we examined the effects of each of the 3 intervention components on both affective measures and steps. A GEE model was used to account for within-subject correlations over the 6-week intervention period.

The MRT analyses of walking suggestions did not show a significant difference in step counts in the 4-hour window following the decision points, either between providing any kind of walking prompt and providing no prompt or between sending a GIF prompt versus a text-only walking prompt (P=.55). For the salience group, the MRT analysis revealed a nonsignificant increase in daily affective reflection scores of 0.142 points when salience messages were sent compared to when no message was sent (P=.26). Similarly, there was a moderate but nonsignificant increase in daily anticipated affect toward upcoming walks of 0.16 points (P=.15).

The primary aim of this proof-of-concept study was to evaluate the feasibility of the WalkToJoy intervention and gather initial evidence for its ability to have clinically meaningful impact on affective association with PA and on PA levels among middle-aged to older adults. The intervention package was grounded in the ART of physical inactivity and exercise, and it leveraged both automatic and reflective affective processes to foster intrinsic motivation [20,28]. Our secondary aim was to gather preliminary evidence for the efficacy of each of WalkToJoy’s 3 intervention components—smile-inducing walking suggestions, salience messages, and planning prompts—to inform optimization decisions for their use in future versions of the WalkToJoy intervention.

Table 5 combines both the quantitative and qualitative findings. Overall, the results suggest that WalkToJoy is both feasible and promising for fostering positive affective associations with walking and promoting engagement with the walking behavior itself. The intervention package as a whole demonstrated robust pretest-posttest improvements in affective outcomes. Affective attitude, the primary outcome, significantly increased by 0.547 points (P<.001), whereas affective valuations, the secondary outcome, showed an even greater improvement of 0.718 points (P<.001). Both of these impacts are over half an SD in magnitude, suggesting—based on recent findings [68]—that they may have a meaningful impact on engagement in PA itself. Exploratory measures, including affective reflection and anticipated affect, also demonstrated substantial gains, both increasing by 0.692 points (P<.001 in both cases). These findings underscore the intervention’s potential for impacting participants’ affective associations with walking, consistent with ART’s claim that repeated positive affective experiences can increase intrinsic motivation for PA.

Behaviorally, the pretest-posttest analysis of valid average daily steps showed a nonsignificant increase of 80 steps (P=.79). However, due to the presence of extreme outliers and a noticeable peak in step count during week 5, we conducted a post hoc exploratory analysis comparing baseline to the average of weeks 4 to 6. This revealed a larger increase of 506 steps (P=.07). While this finding suggests a preliminary signal of a positive behavioral trend, it should be interpreted with caution given its post hoc nature and the lack of significance at the conventional threshold for statistical significance. Nonetheless, the observed trend suggests a potentially meaningful signal worth exploring further in future optimization trials. An increase of 500 steps per day has been associated with a 6% reduction in cardiovascular disease risk and a 7% reduction in cardiovascular mortality [70,71], suggesting that this magnitude of change may have a meaningful implication for health promotion, particularly in midlife and older adults. Qualitative data further supported the positive impact on affective mechanisms and PA as well. Participants described heightened awareness of the benefits of walking; increased positivity, confidence, and motivation; and a diminished perception of psychological barriers.

Among the intervention components, the GIF walking suggestions emerged as the most impactful driver of affective and behavioral outcomes. While GEE revealed nonsignificant results for affective attitude and affective valuations, GIF prompts significantly increased anticipated affect toward future walks by 0.345 points (P=.046) compared to identical text-only prompts. These findings provide initial support for our hypothesis that evaluative conditioning—pairing a behavioral stimulus (eg, a walking prompt) with a positively valenced stimulus (eg, smile-inducing GIF images)—fosters positive emotional associations over time through automatic affective processing [28,37,72]. This conditioning mechanism likely contributed to the observed increases in participants’ positive anticipation and motivation to walk. In addition, GIF prompts significantly boosted steps, with GIF group participants averaging 1834 more steps per day across the entire study (P=.05) compared to those in the text-only group. This substantial behavioral impact highlights GIF messages’ potential as engaging and burden-resistant reminders that can impact PA behavior itself.

These results align with the qualitative insights, where participants showed a clear preference for the smile-inducing GIF walking prompts over identical text-only prompts for maintaining engagement and motivation. While text-only prompts were initially effective as reminders to walk, they became burdensome over time. In contrast, participants consistently described the GIF images as humorous, visually engaging, emotionally uplifting, and even evoking a sense of anticipation. These qualities helped sustain engagement and mitigate message fatigue. A recurring challenge in current mHealth interventions is achieving sustained user engagement, which is often corroded by treatment fatigue [14,73,74]. High user burden from repeated treatment delivery frequently results in low intervention adherence, thereby limiting the long-term effectiveness of such interventions. Our findings suggest that evaluative conditioning provides a promising framework for designing fatigue-resistant and positively valenced message delivery strategies. This approach has the potential to mitigate message fatigue and sustain engagement, especially in interventions requiring long-term adherence. Moreover, this affect-based approach could be extended to other components of mHealth interventions, such as planning, goal setting, and monitoring behaviors. Future studies should explore ways to integrate affective strategies into interventions that have traditionally relied on reflective cognitive processes to further enhance intervention effectiveness.

Collectively, these results highlight the potential of GIF walking suggestions to support behavior change and foster intrinsic motivation for PA. These findings also build on and extend previous evaluative conditioning studies conducted in controlled laboratory environments. While previous research has demonstrated that pairing PA cues with positively valenced stimuli can improve implicit attitudes and affective evaluations [36-38], such studies have typically been brief and conducted with young adult populations in highly controlled settings. In contrast, WalkToJoy applied this mechanism in a real-world mHealth context among middle-aged and older adults. Compared to HeartPhone [39], a similar mHealth intervention that passively presented exercise imagery on smartphone lock screens, WalkToJoy delivered emotionally engaging stimuli as part of an active behavioral cue—delivering GIF images alongside walking suggestions via SMS text message or Multimedia Messaging Service. This design more directly integrated affective elicitation with the behavioral prompt itself, potentially allowing for a stronger affective-behavioral link and the influencing of behavior.

This final point about how WalkToJoy implemented affective conditioning may also offer an explanation for why we saw large increases in steps for this factor but not in impact on affective attitudes and valuations. Affective associations take time to shift, and given our sample size and the 6-week study duration, we likely did not have enough time to detect a statistically significant impact of a single component on affective associations. However, given that the GIF images were embedded in walking suggestions, they may have had a more immediate impact in that they made participants more receptive to the walking suggestions themselves, resulting in more near-term activity. The interview findings we presented offer support for this conjecture that increased receptivity to walking suggestions may have played a key role in mediating the effects of this intervention.

The MRT analysis revealed no significant increase in a 4-hour step count immediately following the delivery of GIF versus text-only prompts (P=.55). This null finding may be due to design-related factors, including the potential misalignment between the analysis window and the suggestive—not directive—nature of the message framing in the walking suggestions. Affective processes—the primary mechanisms of change in this intervention—are known to evolve gradually and accumulate over time. As such, the lack of immediate impact on the short-term proximal outcome may reflect this slower trajectory. In addition, participants noted in the interviews that GIF prompts were interpreted as flexible, day-long reminders rather than immediate calls to action. This perception appeared to stem from the suggestive and nonurgent wording of the prompts (eg, Could you go for a short walk this afternoon?), which allowed participants to incorporate walking into their routines on their own terms. By reducing immediate pressure, the framing used in WalkToJoy seemed to align with participants’ preferences, fostering positive engagement with the intervention while supporting their ability to make choices about when to walk.

This observation highlights the nuanced role of message framing in health interventions [40,41,43], emphasizing that both directive and suggestive framings have distinct advantages and limitations depending on the context and user group. Directive messages, which emphasize immediate action (eg, Take a short walk now to feel energized!), may be better suited for individuals requiring a clear prompt to overcome inertia or decision fatigue as they provide explicit guidance and can create a sense of urgency. Such framing has been shown to be effective in other interventions such as HeartSteps, where more assertive prompts successfully encouraged immediate walking [75]. However, directive framing can also risk inducing psychological resistance, particularly in individuals who value autonomy and flexibility. In contrast, suggestive, nondirective framing—as used in this study—offers a more autonomy-supportive approach. By presenting walking as an open-ended opportunity rather than an immediate demand, this type of framing may reduce pressure, increase perceived control, and encourage participants to act in ways that align with their routines and preferences. This aligns with self-determination theory, which posits that behaviors motivated by autonomy and perceived choice are more likely to be internalized and sustained over time [76]. However, the trade-off is that this framing might not elicit immediate behavioral responses, as observed in the lack of significant proximal step count increases in this study. Future mHealth interventions could benefit from leveraging both approaches, tailoring message framing to the specific goals of the intervention and the needs of diverse user groups. For instance, directive prompts could be used strategically in moments requiring immediate action, whereas suggestive prompts could support longer-term habit building and user autonomy.

Our analyses found no significant impacts of salience messages and planning prompts on affective and behavioral outcomes, suggesting that these components may require refinement or that larger or longer studies would be needed to demonstrate impact given that affective processes take time to evolve. Despite the lack of statistical significance, qualitative data suggest that salience messages facilitated meaningful attitudinal shifts by reframing walking as an enriching and mindful activity. Through the emphasis on the positive aspects of walking via message framing [40,41,43], participants felt that salience messages transformed walking from an obligation into an intrinsically rewarding experience. Similarly, planning prompts, though not statistically significant, were highly valued by participants for their role in enhancing the enjoyment and personalization of walking. These observations are consistent with previous research underscoring the importance of personalization in enhancing user engagement and satisfaction in health interventions [77,78].

In contrast to the more automatic mechanism of evaluative conditioning used in the GIF walking suggestions, the salience and planning components of WalkToJoy aimed to support belief updating, which is theorized to be slower and more reflective in nature. Previous interventions have used message framing or action planning to promote walking among older adults [41,51-53], but few have explicitly targeted affective attitudes as a mechanism of change. WalkToJoy advances this approach by using positive framing and planning prompts to elicit emotionally positive walking experiences and then foster reflection on what these experiences mean for how participants feel about walking. Thus, these components slowly nudged participants to revise how they thought they felt about PA using their own experiences of evidence that walking may be something that they actually like. Our qualitative findings indicate that this process unfolded for at least some of our participants, providing initial evidence for its promise.

This study, as a proof-of-concept investigation, faced several inherent limitations. First, this study was not fully powered for the factorial experiment, limiting the ability to detect smaller or more nuanced effects of the individual intervention components. In addition, the study’s duration of 6 weeks was relatively short, precluding the ability to assess WalkToJoy’s long-term impact on affective mechanisms and behavior change.

Another notable limitation is the potential sampling bias of requiring participants to own a Fitbit, which may have resulted in a more health-literate and self-motivated sample than the general population. However, this was consistent with the intervention’s target audience—individuals who are interested in becoming more active but face barriers to maintaining motivation—making this sample appropriate for evaluating the intervention’s initial promise. Participants were generally moderately active and reported favorable attitudes toward walking at baseline, potentially limiting observable improvements but also suggesting higher receptivity to the intervention. Future work should examine how such interventions perform among individuals with more negative affective associations with PA and lower baseline activity levels.

This proof-of-concept study of the WalkToJoy intervention demonstrates the feasibility and potential of affective-based strategies for promoting PA. The smile-inducing walking suggestions in particular played a significant role in enhancing engagement and motivation, underscoring the value of affect-driven and visually engaging approaches as a fatigue-resistant intervention strategy. While salience and planning prompts showed promise, further refinement and extended deployments may be necessary to optimize their effects. These findings highlight the importance of integrating affective dimensions into digital health interventions to foster sustained engagement and meaningful health behavior change.