Most adults do not engage in sufficient physical activity for good health [1]. Noncommunicable diseases related to sedentary lifestyles are one of the leading causes of death worldwide, costing societies approximately US $67.5 billion per year according to conservative estimates [2]. Although the core reasons and dynamics of insufficient physical activity vary among societies, it is a global public health priority to support individuals in increasing their physical activity levels and improving their health [3]. To design effective interventions, a better understanding is needed of the factors that determine individuals’ physical activity in everyday life and the techniques that are effective in targeting those determinants [4].

Smartphone apps and activity trackers offer many key features for supporting active lifestyles, and the wide reach and cost-effective dissemination of digital interventions hold promise for population-level behavioral support [5]. Sensor technology in smartphones provides many advantages for physical activity interventions, such as the automatic measurement of activity and individually tailored support messages [6,7]. Another benefit is the opportunity to collect real-time data on the association between individuals’ cognition and behavior in their natural environments, minimizing mistakes because of memory bias.

The effectiveness of mobile device–based interventions for physical activity varies, but overall, mobile health apps have led to small increases in physical activity [8,9] and decreases in sedentary time [10]. However, smartphone intervention effects on daily steps have not been demonstrated [11]. To develop smartphone interventions that reach their potential and lead more systematically to health-enhancing levels of physical activity, we must identify the factors that determine physical activity and lead to successful intervention engagement.

Typically, physical activity apps are built on self-monitoring of behavior, a behavior change technique (BCT) [12] established as a key ingredient of successful physical activity interventions [13]. However, the fact that self-monitoring is an effective BCT for an average participant does not mean that it will help all individuals or that everyone will use this technique. Tracking physical activity may not motivate all users and can even undermine motivation [14,15]. To increase the uptake and impact of physical activity apps, self-monitoring and tracking features may need to be combined with other behavioral or motivational techniques [16].

One of the factors influencing the uptake and commitment to self-monitoring and other self-regulatory BCTs is the quality of motivation for physical activity [17,18]. Individuals are more likely to set goals, make plans, and follow their progress when physical activity meets their psychological needs, corresponds to their life goals or identity, or brings them pleasure [17,18]. By addressing these determinants of autonomous motivation, interventions may help individuals more actively engage in self-regulatory BCTs.

A meta-analysis on physical activity motivation found that interventions that included any digital component produced significant cumulative effects on intention, stage of change, and autonomous motivation, but the meta-analysis included too few studies to specifically examine the motivational effects of smartphone-based interventions or identify which digital components most effectively increased motivation [19]. To optimize the motivational efficacy of smartphone-based interventions, it could be useful to digitally replicate versions of techniques drawn from face-to-face interventions, such as interpersonal interaction, sense of relatedness, empathy, and encouragement. These are central elements in satisfying the psychological needs of self-determination theory (SDT) and are systematically used in the interaction method known as motivational interviewing (MI). Supporting the needs of autonomy, competence, and relatedness and creating an empathetic environment could also increase the impact of smartphone-delivered interventions.

Most evidence on the determinants of physical activity comes from group average–based between participant studies. Studying individual participants at the intraindividual level may help detect effects that differ from those found between participants [20,21]. Individuals can demonstrate associations of different strength and even in opposite directions when examining key variables of interest, a phenomenon that would be missed when observing group averages only. In addition, different predictors can be more influential for different individuals, as shown by Smith et al [22], who found that different social cognitive theory–based determinants predicted physical activity in the 6 adults they studied.

Studies conducted at the within-person (or idiographic) level include N-of-1 studies, which analyze each individual as their own study unit [23]. Such studies are increasingly being used in health psychology [22,24-27]. In addition to observing associations between variables, active manipulation of research conditions can be conducted with N-of-1 randomized controlled trials (RCTs), which may assess more than one intervention element during the same trial using a factorial design, as in the studies by Nyman et al [28] and Sniehotta et al [29]. These within-person RCTs use individuals as their own controls comparing periods with intervention elements with control periods with no active intervention elements [23]. In the past, N-of-1 RCTs in the field of health behavior change have required a research team member to actively deliver the interventions, prompting the participants to choose the right intervention with daily SMS text messages [29], delivering intervention allocation envelopes to the participants once a week in person [28], and collecting data from the participants every week in person [28,29]. This is time-consuming for both researchers and participants, leading to possible selection bias and increasing the risk of errors. Automating intervention delivery via a smartphone platform may improve measurement precision and data quality. N-of-1 RCTs are recommended by experts, especially for an individual’s treatment decisions [30] and for testing theoretical mechanisms within individuals [31]. However, within-person designs remain underused in behavior change research [32].

N-of-1 designs do not limit analyses to the individual level but also allow for aggregating data across all participants [29,33]. In aggregated multilevel models, individual-level variance is incorporated into the error terms, losing its informational value about individuals [34]. Conversely, aggregated multilevel models enable some level of generalization from the whole sample while adjusting for the individuals’ differences in the dependent variable [20]. Combining idiographic N-of-1 analyses with aggregated models can offer different perspectives on the same scientific question as all methodological approaches have their own biases [35].

Smartphones and wearable technology enable the continuous collection of individual-level data, suiting within-person studies [36]. Smartphone apps also enable N-of-1 RCTs that deliver prespecified intervention techniques to users at randomly allocated times.

The Precious mobile app was designed building on theory and evidence to increase users’ physical activity using motivational and self-regulatory techniques [37]. The core functions of the app were as a tracking tool with self-regulatory BCTs, such as behavioral goal setting and self-monitoring, and motivational tools aimed at increasing uptake of the tracking features.

The motivational features of the Precious app draw on MI, a person-centered communication style that supports behavior change by increasing the salience of values and goals related to the desired behavior within an atmosphere of acceptance and compassion [38]. MI has shown promise in increasing physical activity with face-to-face or telephone-delivered interventions [39] and with computer-based interventions [40,41], but automated delivery of MI via smartphone apps has not been studied.

The Precious app combines MI with heart rate variability–based biofeedback to strengthen the mental link between an individual’s actions and well-being. Biofeedback has been found to reduce stress and anxiety about physical activity, removing a barrier for some inactive individuals [42]. A pilot study also found that biofeedback improved quality of life and reduced fatigue in a small sample of women with chronic fatigue syndrome [43]. Feasibility tests with the Precious app found promising participant engagement with the MI features [37]. In a 3-month usability RCT, the Precious app was found to be acceptable among persons with obesity, and participants were particularly satisfied with the app’s biofeedback report and physical activity modules [44].

Several well-established theories suggest that individuals’ physical activity is shaped primarily by 2 key modifiable psychological factors: self-efficacy and motivation.

Self-efficacy is defined as one’s beliefs in their capability to successfully perform courses of action and achieve desired effects [45]. It is characterized as a key factor determining intentions to be physically active within health psychological theories including the health action process approach [46], social cognitive theory [47], the theory of planned behavior (as perceived behavioral control) [48], and SDT (as competence) [49]. High self-efficacy predicts physical activity across different study populations, especially when initiating activity [50,51].

Motivation is a key predictor of physical activity within, for example, SDT [49]. Motivation refers to the desire, urge, energy, or reason to perform a specific behavior [52]. One of the prerequisites for sustained motivation is the sense of competence, or self-efficacy, as the fulfillment of this psychological need is suggested to lead to an internalized desire to act [53]. This internalized or autonomous motivation is an established predictor of physical activity [54-56].

Despite the central position of self-efficacy and motivation in behavioral theory and interventions, few studies have observed whether changes in these predictors are followed by immediate changes in physical activity in everyday life [22,25,57]. Studies measuring behavioral variables typically compare very few time points and summarize the average effect of the predictors for the sample rather than providing effects for each participant [34,58]. Finding an association between self-efficacy, motivation, and physical activity within individuals repeatedly over time in everyday life would provide stronger evidence for predictive models [25].

This study is the first in a series of factorial N-of-1 experiments conducted using the Precious app, and it specifically sought to test the impacts of the digitalized MI (dMI) and biofeedback intervention components under study. The aims of this study were to (1) test whether the participants’ daily steps increased on intervention days when the app delivered motivational interventions and (2) examine the associations among self-efficacy, motivation, physical activity, and daily steps.

The analyses addressed the following research questions (RQs): (1) Do people take more steps on days when they are offered motivational smartphone-based interventions (intervention 1: MI components; intervention 2: biofeedback) compared with nonintervention days? (RQ 1) and (2) Do daily self-efficacy and motivation predict daily steps in individuals? (RQ 2).

This study was a 40-day 2 × 2 factorial N-of-1 RCT testing the effects of dMI and biofeedback and involved twice-daily (morning and afternoon) ecological momentary assessments (EMAs) of psychological and environmental predictors of physical activity. The study has been reported following the Consolidated Standards of Reporting Trials Extension for reporting N-of-1 Trials (CENT) [59]. The trial was not formally preregistered as conventions for the registration of factorial N-of-1 experimental studies had not been established before this study’s commencement in 2016. However, a dated publicly available version of the study protocol was published just after the start of data collection and before any data analyses [60].

The 40-day trial included 12 three-day active study periods (ie, 2 days during which one of the study conditions described in the Randomization section was implemented plus a wash-out day in which dMI features were hidden from the app and biofeedback notifications ceased), 2 additional control days (1 each after the fourth and eighth active study periods), and a 2-day lead-out period in which all app features were available (Figure 1). Wash-out days were included in the study design to eliminate any carryover effects of dMI or biofeedback on cognition or behavior on subsequent days [61].

Each active study period tested the effects of one of the following conditions: (1) both dMI and biofeedback, (2) dMI alone, (3) biofeedback alone, or (4) a control period in which neither intervention was delivered. The 4 conditions were repeatedly block randomized to one of the 12 three-day active study periods using a computer-generated code. This led to each of the 4 conditions being assigned to three 3-day study periods overall. Randomizing the timing of repeated interventions helps avoid potential time-based confounders that might systematically coincide with intervention delivery [62]. A code for activating the individually randomized trial was printed and sealed in an opaque envelope. Each participant drew one of the envelopes at the baseline meeting with a researcher (JN) and entered this code into the app to initiate the trial procedure.

As we aimed to test the impact of motivational features that require active cognitive engagement with the tasks [37], it was not possible to blind participants to the interventions they received on a given day. However, participants were not aware of the specific study hypotheses or analysis methods, and during the study, participants were blinded to the sequence in which intervention components would be delivered or available in the Precious app. Tests for blinding were not conducted.

On MI intervention days, participants received a morning notification of new content: “A new test period has started. Come see what the Precious app has to offer.” The notification remained visible until it was touched to open the app or swiped away. The digitalized elements of MI (Figure 2) are described in detail in the study by Nurmi et al [37] and include the techniques listed in Table S1 in Multimedia Appendix 1. To imitate the interactivity of face-to-face MI, dMI tools were offered in a stepwise manner: 3 dMI tools in 3 consecutive periods. Set A included “What do I want” and “Choose Favorite PAs,” set B included “Importance Ruler” and “What’s Next (stage of change),” and set C included “Time Machine” and “Confidence Ruler.” Thus, the tools offered slight variability over the course of the trial to maintain user interest.

The biofeedback interventions started with a preparation notification the day before each measurement: “Tomorrow is a Firstbeat assessment day. Be sure to have the device charged and ready to wear tomorrow.” On the first intervention day, participants received a notification stating the following: “Today is the day! Please wear your Firstbeat device today and upload the data tomorrow.” On the second intervention morning, the notification read the following: “We hope you slept well. Please upload the data from your Firstbeat device to get a report on your activity, sleep, and stress levels.” To access the report, participants needed to plug the wearable sensor into their computer’s USB port and upload the data to the trial website. The biofeedback report was provided using the Firstbeat Bodyguard 2 heart rate variability monitor (Firstbeat Technologies Ltd). These data were first pushed securely to the Firstbeat company servers for analysis and then passed securely to the Precious server for storage. This intervention used BCTs 2.4 (self-monitoring outcomes), 2.6 (biofeedback), 2.7 (feedback on outcomes of behavior), 7.1 (prompts or cues), and 12.5 (adding objects to the environment) [12].

On control and wash-out days, the dMI and biofeedback tools were hidden from the app. The participants could still freely access all the self-regulatory BCTs, including behavioral goal setting and self-monitoring (the full list is available in Table S1 in Multimedia Appendix 1). They also received communication related to their goal progress at 5 PM every day: “You’ve taken [step total] steps today—that’s (100*[step total]/[goal amount])% of your goal. Keep going!” (BCT 2.2, feedback on behavior and goal progress [12]). If participants set a step goal and reached it, they received a congratulatory message: “Good job! You’ve achieved your step goal for today. Click here to see your progress.” (BCT 10.4, social reward [12]).

Participants were recruited from the general population using advertisements in the Metro newspaper of the Helsinki area, Finland, and a Facebook page and targeted advertisements in October 2016. People who responded to the advertisement were contacted by the research team via email or phone to establish whether they met the following eligibility criteria: age of >18 years, ability to speak Finnish, ability to read and understand English, no contraindications to engaging in physical activity as assessed using the Physical Activity Readiness Questionnaire [65], ownership and use of a smartphone with a compatible operating system (Jelly Bean version 4.1 or higher for Android), willingness to install the Precious app on said smartphone for a 40-day period, no use of any physical activity trackers (eg, Fitbit, Garmin, or Misfit) or physical activity apps in the previous 6 months, no participation in other trials or behavior change programs in the previous 6 months or during the trial, levels of physical activity below the recommendation of 150 minutes per week of moderate-intensity physical activity [66], and willingness to wear an activity tracker for the duration of the study. In addition, participants were excluded if they were seeking to be enrolled in the trial concurrently with a friend or relative. This exclusion criterion was applied to avoid exposure to randomly timed intervention materials during control days through the other person’s phone.

Of the 147 people who responded to the advertisement, 48 (32.7%) did not respond to efforts to contact them, 24 (16.3%) were too active, 24 (16.3%) already used a physical activity tracker or health app, 16 (10.9%) did not have the necessary technical set-up, 6 (4.1%) were excluded as they wished to participate with a friend or spouse, 5 (3.4%) were undergoing other interventions, 4 (2.7%) had poor health, and 2 (1.4%) had scheduling conflicts. The remaining 17 people met all the inclusion criteria, but 2 (12%) declined to participate as they did not wish to do so without a friend or spouse who did not meet the inclusion criteria. This resulted in a final sample of 15 participants.

The number of participants was limited by available resources (eg, activity bracelets, heart rate variability sensors, and technical support). The intervention length was limited partly by the estimated battery life of Mi Band activity bracelets (approximately 40-50 days without charging) and partly to avoid major public holidays that might affect participants’ physical activity. The planned sample of 15 participants with 40 observation days would yield 600 observation days in total, which was similar to that of earlier factorial N-of-1 RCTs that included 8 participants for 62 days and 10 participants for 60 days [28,29].

Participants were invited for a face-to-face intake session from October 2016 to November 2016 in which they were asked to read the study information sheet (which had also been sent to them by email) and could ask questions. Individuals wishing to be enrolled in the study then provided informed consent to participate and began the study.

To begin the study, participants had a 60- to 90-minute long individual instruction session with a researcher (JN) at a university office. After receiving information on the study and signing informed consent sheets in person, participants randomly chose an envelope from a bag. Each opaque envelope contained a study code, which was entered into the Precious app to carry out the study period randomization procedure described previously.

Participants received help to install the Precious and EMA apps on their phones and entered their study code into the Precious app. The researcher then instructed participants on how to use these apps. Participants then received the Mi Band activity bracelet and learned how to pair it with the Precious app. For biofeedback measurements, they then received a Firstbeat Bodyguard 2 device and were taught how to conduct heart rate variability measurements and read these reports at home on their PCs. Participants were advised to follow the instructions of the apps until their individual follow-up meetings. Printed instructions and researchers’ contact details were provided with the material pack, and participants were encouraged to contact the researchers in case of any technical difficulties. All participants received a portable power bank as a gift for participating and to help keep their phones charged during the trial.

After the 40-day trial, participants returned to the university for an individual follow-up meeting, debriefing, and exit interview and to return the activity bracelet and heart rate variability monitor. A researcher (JN or KK) downloaded the wristband data that were stored locally on participants’ phones and helped participants uninstall the study apps. Participants were then asked about their experiences using the Precious app over the course of the study during a semistructured interview. At the end, participants were rewarded with 3 movie tickets and thanked for their participation.

The University of Helsinki Ethical Review Board in the Humanities and Social and Behavioural Sciences granted a favorable decision for this study (statement 3/2016).

A total of 15 healthy adults (n=4, 27% male and n=11, 73% female) took part in the study, with ages ranging from 28 to 57 (mean 42.33, SD 9.82) years. All participants reported wearing the activity bracelet continuously for the duration of the trial; however, some participants were missing step data because of technical errors, including participant 9, who had no step data, and participant 2, who had step data for only 53% (21/40) of the measurement days. This left a total of 88.3% (530/600) of usable data points. Missing values were not imputed as the missing completely at random test [71] found the study values not to be missing completely at random (χ²₂₅=85.3; P<.001) and common multiple-imputation methods cannot reliably handle this potential bias [72,73].

All participants (15/15, 100%) finished the trial, returned for the follow-up meeting, reported wearing the activity bracelet continuously for the entire duration of the trial, and kept using features of the Precious app throughout the trial. Participants engaged with dMI features during an average of 5.10 (SD 1.0; range 3-7) out of the 7 intervention periods and conducted biofeedback measurements during an average of 5.67 (SD 1.4; range 2-7) out of 7 intervention periods. All participants (15/15, 100%) conducted biofeedback measurements, but 4 of them partly missed the suggested intervention days: 3 (75%) conducted a measurement a day later than suggested, and 1 (25%) conducted all their biofeedback measurements during control days. Unforeseen technical problems prevented 13% (2/15) of the participants from accessing their biofeedback reports during the trial (details in Table 1 and Table S1 in Multimedia Appendix 2). Among the 13 participants for whom data were available, step count goals were set on 42.7% (222/520) of the trial days, with a wide range of 0 to 38 days out of 40. These data are presented visually in Figure S1 in Multimedia Appendix 3.

Visual analysis of sequence charts (Multimedia Appendix 4) confirmed that individuals’ steps varied sufficiently over time to conduct analyses [32]. Individual participants’ steps did not show statistically significant time trends except for participant 4, whose steps slightly decreased over time (B=−84, SD 39, 95% CI −163 to −5; R²=0.12; P=.04). However, visual assessment showed a decline in participant 4’s steps only during control days (Multimedia Appendix 4). Multimedia Appendix 4 also shows how participants’ average activity levels and progress over time differed between individuals. Weekday-related patterns in physical activity were detected and controlled for in participants 10, 11, and 13; autoregression was detected in participants 7, 10, 11, 14, and 15; and their individual lag was adjusted with a pertinent step lag variable in the dynamic regressions.

The overall mean daily steps were 10,786 (SD 5393) on dMI intervention days, 11,125 (SD 5360) on biofeedback intervention days, and 11,053 (SD 5922) on control days (including wash-out days). See Table 2 for additional step count data. The dMI and biofeedback interventions did not show any statistically significant associations with daily steps for any individual participant (all P>.05; Table 3).

The average number of steps of participants across all time points (excluding wash-out days; fixed effect in the null model) was 11,137, and participants’ overall steps did not significantly change over the course of the study. The data showed no advantage of a quadratic fit compared with a linear fit. The differences between participants’ step slopes explained 21% of the variance in steps over time. The ICC was high, with 57.1% of the variance attributable to differences between participants. The covariance between slope and intercept was nonsignificant (P=.53), suggesting that the initial level of steps did not affect changes over time.

Multilevel models of the intervention effects did not identify any significant associations between condition and steps (dMI: B=−246, 95% CI −1012 to 520, SE 389, t₃₁₂=0.63, and P=.53; biofeedback: B=67, 95% CI −688 to 821, SE 384, t₃₁₁=−0.17, and P=.86). Adjusting for autoregression did not significantly improve the model fit, and there was no association between participants’ steps at adjacent time points (B=0.12, SE 0.07; P=.11). Table 4 presents the series of models used, and Table 5 shows the details of the model that best fit the data. The day of the week variable revealed that participants were most active on Mondays and Wednesdays, taking 2303 and 1609 steps more, respectively, than on Sundays. Multimedia Appendix 4 shows that intercepts and slopes varied substantially between participants on both intervention and nonintervention days.

Morning self-efficacy predicted a higher number of steps during the day in 27% (4/15) of the participants (Table 6), and morning motivation predicted a higher number of steps during the day in 20% (3/15) of the participants (Table 7). Self-efficacy and motivation were analyzed and are presented separately because of their theory-based association [46,48,49] and high correlation in the sample (r=0.597).

Table 8 presents the series of models used to investigate associations between motivation and self-efficacy and daily steps, and Table 9 shows the details of the models that best fit the data. When including wash-out days, the ICC was high—approximately 61.4% of the variance in steps was attributable to differences between participants. The average number of steps across all participants and time points (fixed effect in the null model) was 11,185. The average starting level of steps across participants was 11,789. The fixed effect of time on steps was statistically significant and negative (B=−0.32, 95% CI −58 to 5; P=.02); however, adding time to the model did not improve the model fit, and the linear change over time explained only 1.1% of the variability within participants in their steps. The data showed no advantage of a quadratic fit compared with a linear fit. Adding a random effect of time did not improve the model fit. However, the variation in the growth model between participants seemed significant—adding a random effect of time when also letting the intercept and slope correlate (UN) explained 13% of the between-person intercept variance (difference between individuals). This would indicate that participants had different starting step levels and trajectories over time, as shown in Multimedia Appendix 4. The covariance between slope and intercept—UN (2,1)—was nonsignificant (P=.23), suggesting that the initial level of steps did not affect how much they changed over time. Adding the control variables day of the week, within- and between-person perceived barriers, and pain or illness clearly improved the model fit (P<.001) and explained 5.4% of the within-person variance and 3.3% of the between-person variance in steps.

Adding the predictor, fixed effect of within- and between-person self-efficacy, improved the model fit (P<.001) and explained 3.9% of the residual variance (individuals’ change over time) in steps. The fixed effect of within-person self-efficacy was statistically significant (B=462, 95% CI 296-628, SE 84; P<.001), suggesting that, when participants’ morning self-efficacy increased by 1, their daily steps increased by 462. The within-person perceived barriers were not statistically significant (B=−48, 95% CI −209 to 114, SE 82; P=.56; Table 9). Within-person pain or illness scores indicated that participants took, on average, 1524 steps less when they reported 1 score higher on a scale of 0 to 2 (B=−1524, 95% CI −2378 to 669, SE 435; P<.001). Parameter estimates for all variables in this model are shown in Table 9.

The fixed effect, within- and between-person morning motivation, improved the model fit and explained 2.7% of residual variance (individuals’ change over time) in steps. The fixed effect of within-person motivation was statistically significant (B=390, 95% CI 201-578, SE 96; P<.001), suggesting that, when participants’ morning motivation increased by 1, their daily steps increased by 390. Within-person perceived barriers were not statistically significant (B=−93, 95% CI −257 to 71; P=.27). Within-person pain or illness scores indicated that participants took, on average, 1828 steps less when they reported 1 score higher on a scale of 0 to 2 (B=−1828, 95% CI −2676 to –980, SE 431; P<.001). Parameter estimates for all variables in this model are shown in Table 10.

No statistically significant differences were detected between steps on intervention and control days, neither in the N-of-1 analyses nor when the data were aggregated. The average of the aggregated steps on biofeedback intervention days was approximately the same as the control day average, and on dMI intervention days, the step average was somewhat lower than the control day average, although this difference was not statistically significant. These findings should be interpreted with caution because of the following features of the pilot trial.

First, the availability of self-regulatory BCTs may have reduced the intervention effectiveness. To maintain user interest in the app over time, the participants had continuous access to several self-regulatory BCTs also during control days (Table S1 in Multimedia Appendix 2). Meta-regressions have found a combination of self-monitoring and other self-regulatory BCTs to be effective in increasing physical activity [13] and intentions for physical activity [19]. Behavioral goal setting and self-monitoring were also found to increase steps in some participants in a factorial N-of-1 RCT [29]. Of the MI techniques used in this study, meta-analyses have found support only for the techniques “BCT 15.2: mental rehearsal of successful performance” in increasing intentions [19] and “SDT3: provide a rationale” in increasing autonomous motivation for physical activity [74]. These meta-analyses did not detect statistically significant effects of MI on autonomous motivation [19,74], intention, or stage of change [19]. Thus, the continuous availability of self-regulatory BCTs may have overridden the possible effects of the motivational interventions.

Second, the impact of the dMI intervention may have been diluted by the daily EMA questions as answering questions on motivation, self-efficacy, perceived barriers, and pain or illness also requires self-reflection on one’s motivation, capability, and opportunity to be physically active. Although the dMI intervention had more substantial content and used BCTs to evoke reflection on the reasons, life goals, and positive memories with physical activity [37], it is possible that users associated the daily EMA questions with the Precious app content more generally.

Third, the second intervention element, heart rate variability–based biofeedback, has the potential to be highly motivational content as it is personally relevant and responds immediately to changes in behavior [75-77]. The downside of this personal tailoring is that the content cannot be standardized or predetermined (ie, the “active ingredient” of the feedback may change). For instance, inactive participants may find it discouraging to see their activity levels not meet the recommendations, and participants with high stress scores may decide to focus on recovery instead of exercise. Thus, providing biofeedback to participants whose behavior and stress levels differ means that the intervention content also differs. Interestingly, anecdotal feedback in the follow-up interviews (not reported in this study) revealed that the participants may have actively varied their activity levels on biofeedback measurement days to receive a comprehensive picture of their well-being. Thus, the biofeedback measurements may have encouraged participants to even decrease their steps to receive a baseline reading on a recovery day.

Finally, the motivational interventions may also have met a ceiling effect. Despite self-reporting physical activity levels low enough to be included in the study, our participants were nevertheless those who had actively contacted the research team after seeing the newspaper advertisement, and they reported relatively high motivation and had high daily step averages overall. This possible ceiling effect and self-selection bias issue are a wider problem in physical activity promotion research.

The MI-based relational features in the Precious app aimed to support users’ need for relatedness [37] but did not seem to reach the effectiveness of face-to-face MI [39]. This is in line with the results of a meta-analysis that found that face-to-face delivery is a key factor in interventions targeting motivation and intention for physical activity [19]. Future studies are needed to investigate how digital interventions could approach the effectiveness of personal contact, exploring, for instance, the amount of contact [78] or the depth of engagement with intervention content [37]. The simple automated messages of the Precious app may not be perceived as MI, and more sophisticated, artificial intelligence–based solutions could improve the user experience and service effectiveness.

The observational analyses revealed that morning self-efficacy predicted daily steps in 27% (4/15) of the participants, whereas motivation predicted steps in 20% (3/15) of the participants. Self-efficacy and motivation were also statistically significant predictors when aggregating data from all participants. This acted as a validating element for the pilot trial data collection as daily steps and their predictors followed the theory-based hypotheses. It also provides further support to the models that suggest self-efficacy and motivation as key determinants of physical activity. The positive associations suggest that strong enough self-efficacy and motivation will help some individuals overcome everyday hurdles and find ways to add more steps to their everyday lives. Thus, self-efficacy and motivation remain central intervention targets.

Only 4 (29%) out of 14 participants showed an association between self-efficacy or motivation and steps when N-of-1 analyses were undertaken. For some, this may be explained by the low statistical power. These associations may also be hidden by time lags of varying lengths as motivation may not translate to physical activity immediately. Physical activity is time-consuming, and motivation can only translate to activity when environmental factors allow that [34]. For instance, participant 13 reported that her daily steps mainly depended on the availability of a car as her job included moving between different locations either by car or on foot. Exercising and sports may also require special clothing or equipment, and certain activities may be available only on certain weekdays, whereas others require a partner or team to play with. The perceived barriers variable may have controlled for some of these environmental factors. More accurate control variables might be created by interviewing participants on their personal barriers and facilitators before the start of the trial. Instead of daily steps, a more proximal outcome of increased motivation might be an increase in action planning [18]. Future studies could focus on detecting the effects of motivation on planning and follow the enactment of these plans.

Individual differences were found in the associations between the EMA predictors and steps. For participants 13 and 14, even high motivation could not overcome the impact of perceived barriers and pain or illness. For participants 7 and 14, pain or illness may have decreased daily steps independently of their motivation, but the effect was not detected when self-efficacy was included in the model, suggesting that their beliefs about their capability to be active could provide a better estimate of their daily activity than their desire to be active. Interestingly, participant 14 seemed to take more steps on days with higher self-efficacy but also on days with higher perceived barriers. A mixed methods study could identify possible reasons behind these associations by interviewing participants using their data as a starting point.

This trial aimed to test the immediate impact of specific intervention elements on individuals and did not hypothesize a lasting change in an individual’s steps over time. The data showed a slight overall decrease in steps over time, possibly influenced by the timing of the intervention as weather conditions often become increasingly challenging in Finland from October to December [79].

This feasibility study revealed many useful considerations for future studies in this emergent field of within-person RCTs.

To our knowledge, this study is the first to test smartphone delivery of dMI features in a fully automated factorial N-of-1 RCT. Both approaches are important avenues for future research—they incorporate the possibilities that new technologies offer for personalized, ubiquitous support to individuals and for rapid testing of the impact of several intervention components in small study populations. Such solutions with reduced costs and easy delivery are needed to tackle major public health challenges.

The procedures of this pilot field trial were acceptable judging by the high intervention uptake and high participant adherence to the activity bracelet and daily EMA measurements. All participants (15/15, 100%) completed the trial, which is in line with other N-of-1 studies lasting a maximum of 3 months with no dropout [27,29,81] or a low dropout [28,82].

The data collection strategy was another strength of this trial as the cognitive correlates of physical activity were collected in a real-life environment and physical activity was objectively measured. Daily EMA measurements minimized the biases associated with retrospective questionnaires. The interventions in the Precious app were multifaceted, including techniques from MI and physiological, heart rate variability–based biofeedback to increase the salience of the immediate consequences of behavior on physiological well-being. Unlike interventions delivered in person, smartphone delivery with a modular study design allowed for testing for specific, isolated intervention techniques and their immediate impact on behavior. This study design may help determine the “active ingredients” in interventions and, thus, advance the understanding of behavioral determinants. This approach could also help determine the individual “dosage” that each user needs [5,83,84]. The use of digital technology also helped track the exact timing of the interventions and recognize that some biofeedback measurements were conducted outside the suggested times.

This factorial N-of-1 RCT with the Precious app showed high acceptability and adherence in an ecologically valid setting. The detected daily within-person associations among self-efficacy, motivation, and steps provided further support for central behavior change theories but also highlighted the possible differences between individuals as these associations were only detected in less than one-third of participants. In addition, this pilot study identified several suggestions to improve the implementation of future N-of-1 RCTs, which come with their specific challenges [61].

The automated N-of-1 delivery was mainly successful. Delivery of the dMI elements of the Precious app was deemed acceptable and feasible, whereas the biofeedback interventions faced some technical and practical challenges. As 27% (4/15) of the participants missed some biofeedback measurement days, smartphone notifications alone seem to have been an insufficient nudge to start the measurement. As N-of-1 RCTs are very sensitive to the accurate timing of interventions [61], future studies using app-controlled intervention delivery need to ensure that notifications are received on time, possibly using sound and vibration alarms in addition to text-based notifications.

This study conducted intention-to-treat analyses studying whether intervention availability is associated with increased step count. Participants typically only engaged with the dMI interventions on the first day they were available. Owing to the pilot nature of the study, the relatively low number of days of engagement with the service, and the spillover of the biofeedback measurements, no per-protocol analyses were conducted. Future studies could explore whether the days in which participants actively engaged with the intervention materials are associated with changes in behavior as behavior change interventions are most effective for people who actively use the available BCTs [85-87].

To minimize participant burden and the risk of confounding motivational effects, all EMA questions were single items [88]. However, single items offer no same-day reference points for imputing missing values. As participants showed high adherence to the twice-a-day EMA measurement [89], future studies could consider assessing more motivational variables that, for example, distinguish qualitatively different motivations from SDT [49].

This study used a factorial design to test 2 different interventions in a short field trial. The duration of this trial was limited to 40 days, primarily because of the estimated battery life of the accelerometers used to collect step data. This means that the study was underpowered to assess the interaction effects of the 2 interventions and, according to recent evidence [90], perhaps even underpowered to study the effects of single interventions in which longer time-series data are needed.

This paper presented an automated N-of-1 factorial RCT delivered via smartphone testing 2 types of physical activity interventions: dMI and heart rate variability–based biofeedback. High intervention uptake and high adherence to daily EMA measurements indicated a good level of acceptance of the Precious app and the automated factorial N-of-1 design, but no intervention effects were found. Daily self-efficacy and motivation were associated with daily steps in 27% (4/15) and 20% (3/15) of the participants, respectively, and in the aggregated data from all participants. The novel use of randomly timed, preprogrammed smartphone notifications for the delivery of the intervention components may decrease the risk of human errors in intervention allocation or data collection. The automated delivery may be sensitive to other challenges, such as missed smartphone notifications and technical problems. With careful selection of intervention content and improved focus on the uptake of digital interventions during the allocated intervention days, an automated N-of-1 RCT can become a valuable tool for testing the impact of specific intervention techniques.