Only 14 studies reported adopting a theoretical base in designing their JITAIs. The most used theory was Social Cognitive Theory (n=9) [26-31,39,40,43], followed by Kanfer’s Theory of Self-Regulation (n=7) [26-30,39,40], Fogg Behavior Model (n=2) [49,51,59], Transtheoretical Model (n=2) [52,59], and Health Action Process Approach (n=1; Tables 1 and 3) [25]. Although Kanfer’s Theory of Self-Regulation was used in 7 different studies [26-30,39,40], they were all derived from the development and evaluation of the SMARTER application, compared to other theories, which were applied across multiple, distinct interventions.

All studies shared weight loss as the distal outcome. Weight loss measures included weight (n=8) [31,43,45-48,50,52], percent weight loss (n=12) [10,26-30,33-37,39,40,43,52], BMI (n=8) [10,33-37,52,56], and waist circumference (n=1) [52]. Proximal outcomes varied across dietary behavior (n=25), physical activity (n=20), and self-weighing (n=17). Dietary behaviors were further subcategorized into (1) dietary structure (eg, measurement of caloric intake and food types; n=13) [26-30,39,40,47,48,51-54], (2) eating rate (n=4) [56-59], and (3) dietary lapses (n=8) [10,32-37,50]. Physical activity was divided into (1) increases in physical activity (n=19) [25-31,39-44,47-49,51-53] and (2) decreases in sedentary behavior (n=3) [41,42,55]. Self-weighing was used to increase users’ weight consciousness (Tables 1 and 3) [26-31,39,40,43,45-48,51-53,56].

To measure tailoring variables, 14 studies used active assessments only [10,31-37,43,45,46,54], 12 studies used passive assessments only [25,41,42,44,49,52,53,55-59], and 9 studies used both (Tables 1 and 3) [3,26-30,39,40,47,48,50,51].

Active assessments included EMA methods such as user-initiated, time-based, and event-based EMA (n=8) [10,32-37,54]. Other studies incorporating active assessments also used self-reported weight measurements, and questionnaires focused on dietary intake, dietary lapses, physical activity, and location (n=15) [26-31,39,40,43,45-48,50,51]. Passive assessments primarily used accelerometers (n=19) [25-30,39-42,44,47-49,51-53,55,57] to measure physical activity (n=18) [25-30,39-42,44,47-49,51-53,55] and eating rate (n=1) [57]. Other sensors used were location sensors (n=4) [44,49,50,52], galvanic sensors (n=2) [41,42], thermal flow sensors (n=2) [41,42], skin temperature sensors (n=2) [41,42], air temperature sensors (n=2) [41,42], image capturing monitors (n=2) [53,58], sound sensors (n=1) [59], and digital weighing devices (n=2) [52,53], Time (n=10) [10,33-37,44-46,49], weather (n=2) [44,49], and personal calendars (n=2) [49,52] were also passively sampled to contextualize user behavior and enhance intervention relevance (Tables 1 and 3).

The JITAI systems used either basic “if-then” rule-based algorithms (n=24) [25-31,39-48,50,51,53,55-58] or machine learning algorithms (n=11; Tables 1 and 3) [10,32-37,49,52,54,59]. Among the 5 interventions using machine learning, Step Up Life [49] and ThinkSlim [54] used decision tree–based models, DietAlert [10,32-37] used ensemble methods that combined multiple models to enhance predictive accuracy, Sweetch [52] used an adaptive reinforcement learning approach to optimize message compliance for individual users in specific contexts, and ADM-IPA [59] implemented a deep belief network to classify dietary intake. These systems leveraged machine learning to recognize activity patterns, user routines, and contextual information, thereby enhancing personalization and improving predictive accuracy, allowing for adaptation to individual needs in real time. Two interventions reported model performance: DietAlert demonstrated accuracies ranging from 0.67 to 0.80, with specificity between 0.68 and 0.84 and sensitivity between 0.45 and 0.77 [32-35], while ADM-IPA reported a best accuracy of 0.74 [59] (Table 2). In contrast, the remaining 14 interventions [25-31,39-48,50,51,53,55-58] used “if-then” rule-based algorithms, in which predefined conditions triggered specific feedback or interventions (Table 2). These algorithms relied strictly on fixed rules and did not consider user behavior.

Intervention delivery methods varied across the included studies. Application messages were most common (n=24) [10,25-30,32-37,39-42,44,47,48,50-52,54] followed by text messages (n=7) [31,43,45,46,55,57,58], audio feedback with or without an earpiece (n=3) [53,57,59], mobile phone haptic vibrations (n=1) [49], and vibrotactile feedback via an augmented fork (n=1) [56]. Smartphones were the primary device in all studies except one [56], with 16 out of 35 studies (46%) also using at least one wearable device, such as smartwatches, in addition to the primary device (Tables 1 and 3) [25-30,39-42,47,48,53,55,57,59].

The included JITAIs offered a range of types of support, including prompts, feedback messages, recommendations for coping strategies, and educational information. Most studies used a combination of prompts and feedback (n=17) [25-30,39,40,44,47-49,53,54,57-59]. Four studies used prompts exclusively [41,42,50,51], 2 studies relied solely on feedback [52,56], and 7 studies integrated prompts with recommendations for coping strategies [10,32-37]. Finally, 5 studies combined prompts, feedback, and educational information (Tables 1 and 3) [31,43,45,46,55]. Prompts were used in various ways, including reminders for physical activity or self-monitoring of weight or exercise [25,27,41,55], and served as alerts to notify users when their dietary behavior was approaching undesirable thresholds [43,50,54,56]. Feedback messages were designed to inform users about their achievements and areas needing improvement. For instance, messages might offer positive reinforcement for successfully limiting calorie intake while encouraging further reduction in sugar consumption [27]. Feedback could also include actionable suggestions tailored to the user in a real-time context, such as recommending physical activities based on location and time of day [44,52]. Recommendations for coping strategies included techniques such as cognitive restructuring and problem-solving [33,34]. Educational information was presented as factual content to users [31,43].

Retention rates were high across studies, ranging from 74% [29] to 100% [32,43,51]. Although 100% retention was most frequently reported in studies that provided participant compensation [32,51], one noncompensated study also achieved 100% retention [43]. Among noncompensated studies, retention ranged from 74% [29] to 100% ([43] Table 2). Definitions of compliance varied across studies but were most commonly based on whether participants viewed, opened, or responded to prompts or feedback. For example, the DietAlert studies defined compliance as EMA survey responses, which ranged from 62.9% to 94.6% [32-37]. In contrast, the SMARTER trial measured compliance using the proportion of feedback messages opened, which ranged from 42.2% to 65.3% [26,28,29,39]. Across all interventions, the reported compliance rate ranged from 15.1% to 94.6% (Table 2) [26,28,29,31-37,39,42,45,47,51,54]. Participants rated interventions as effective, useful, and motivating, with high satisfaction scores ranging from 84.6% to 90% and a strong willingness to recommend their use (Table 2) [31,33,34,41,44,45,49]. Five studies found that greater user engagement was associated with better weight loss outcomes [26,30,31,34,40]. Twenty-three studies incorporated elements to increase user engagement [10,25-31,33-37,39-42,45,46,49,50,52,55], such as creating diverse intervention options and addressing user availability. Preliminary feedback indicated that users preferred interventions with lower demands (eg, prompting a 3-minute break more frequently vs a 12-minute break less frequently [41,42]), milder feedback [40,48], and greater context-sensitivity [40] (Table 2).

Significant weight loss was reported across multiple interventions [26,28-31,33-37,39,40,43,45-47,52], ranging from 0.5% [36] to 3.7% [37]. Reductions in BMI, waist circumference, and blood pressure were also reported [28,29,52,56]. Improvements in physical activity were consistently observed across studies, including greater adherence to physical activity [39,52], decreased sedentary time [41], and increases in step count or walking minutes [25,31,42,55]. Dietary behaviors also improved, with decreases in dietary lapses [33,37], increased adherence to calorie goals [26], improvements in overall diet quality, such as higher fruit and vegetable intake [30,46], and slower eating rates [56] (Table 2).

Consistent with previous JITAI reviews, this review identified a lack of behavioral theories incorporated in weight management JITAIs, potentially stemming from application developers’ limited familiarity with health behavioral theories [8,14]. Multidisciplinary collaborations between engineers, behavioral scientists, and clinicians could bridge this gap [61]. In the absence of dynamic theories, the lack of even static ones is a major shortcoming, as they are essential for comprehending the intricacies of real-world behavior and the dynamics of behavioral change [62]. Theory plays a critical role in the design, implementation, and evaluation of behavior change interventions [63], and DBCIs that integrate theories have shown better outcomes than those that do not [8,62,64]. Theories are valuable in informing the selection of tailoring variables, decision rules, decision points, and intervention options [6,65].

Weight loss was the distal outcome for all studies, given that it was the inclusion criterion. The most common health behavior proximal outcome was dietary behavior. This is likely because dietary behavioral change has been found to have a more significant and consistent effect on weight loss outcomes than physical activity change, although the combination of both is more effective than either alone [66-68].

Other proximal outcomes included physical activity and self-weighing, which reflect the pathways through which weight management JITAIs can exert their effects. Notably, studies that used JITAI approaches to target both physical activity and dietary lapses simultaneously yielded more conclusive results regarding their effectiveness in promoting weight management behaviors [69]. Moreover, diet-focused JITAIs appear to be more successful when paired with other strategies, such as food logging and self-weighing [34]. Collectively, these findings highlight that while diet remains a primary factor in weight loss, a more integrated approach that incorporates physical activity, self-monitoring behaviors, and supportive strategies may maximize the effectiveness and sustainability of JITAIs.

Physical activity and eating rate were mostly measured with passive assessments, but for dietary structure assessment, most studies relied on EMAs. This suggests that passive food intake monitoring technology may not be as accessible as those for physical activity or eating rate [70]. However, a limitation of relying on EMAs over passive assessments is that they may increase cognitive burden and reduce engagement, especially if presented too frequently, as users grow weary of responding [6,10,12]. Possible alternatives for dietary structure assessment with lower user burden include the employment of image processing technology in active assessment, where users simply need to upload photos of their meals, or passive assessment with automatic image recognition through continuous video monitoring [58,70].

Most studies opted for rule-based algorithms as opposed to machine learning. Machine learning can optimize the strengths of JITAIs, enhancing predictive functions and increasing the accuracy of tailoring processes. In addition, the increased information relevance can increase perceived usefulness and user engagement [71]. Machine learning can also support the cost-effective upscaling of JITAI deployment [52]. However, integrating machine learning can be resource-intensive due to computational demands (eg, handling missing data, optimizing parameters, and model testing) and may not always be necessary or effective [10]. For example, machine learning has been effective in predicting lapses at a group level, but its strength falters at an individual level [72,73]. Machine learning also loses its significance if a causal relationship between 2 variables has already been established in the literature [10].

Furthermore, the no free lunch theorem posits that no algorithm is universally optimal for all problems, meaning machine learning is not a 1-size-fits-all solution, but rather a collection of approaches that can be tailored or personalized at the individual level to improve upon prespecified algorithms [74]. Therefore, while machine learning has significant potential to enhance JITAI effectiveness, developers should only use it when practical and likely to improve user experience. Developers should prioritize designing theory-based JITAIs based on feasibility and user needs.

The types of support identified in the included studies were prompts, feedback, recommendations for coping strategies, and educational information. There were differing user preferences regarding the intervention options, with some users preferring more concise instructions [58] and others preferring more detailed ones to guide them [40]. A major challenge in designing effective intervention options is the need to consider variables including boredom, habituation, cognitive overload, and contextual restraints [6].

Given the numerous potential intervention options and delivery methods, as well as the ambiguous efficacy of each component, researchers could consider using microrandomized trials. Unlike traditional randomized controlled trials, participants are not assigned to groups. Instead, each JITAI intervention, and how the participant interacts with it, would be considered a trial. Components of the intervention, such as content, timings, and frequencies, would be randomized based on the study. Data on proximal outcomes, or immediate reactions to the intervention, would be collected and analyzed using regression models at the decision-point level. Information gathered would then be linked to distal outcomes, analyzed at a participant level instead [75].

Microrandomized trials allow researchers to account for the element of time in their study, while also collecting a large number of data points both between and within participants. In addition, insights can be gained from qualitative research, and reinforcement learning could be used to learn user preferences [75].

Ultimately, one of the greatest challenges JITAIs face is optimizing user engagement [10], as evidenced by the highly variable compliance rates found in the included studies, which are attributed to the continuous requirement for lengthy user inputs. Intrinsically, the JITAI design integrates elements accounting for states of receptivity (eg, availability, whether the user is driving, or sleeping), which should increase user engagement [6]. However, many of the studies did not include, or included insufficient, elements to address this. This includes both studies that termed themselves JITAIs and those that did not. We were unable to evaluate sustained engagement due to a lack of evidence, similar to findings from another review on JITAIs for physical activity [14]. This is likely because JITAIs are still in development, and intervention durations were relatively short. In addition, previous literature indicated that JITAIs often have short lifespans [50].

One way to improve user experience and engagement is by focusing on promoting long-term weight loss by enhancing self-efficacy and providing coping strategies, rather than relying solely on short-term prompt effects [76,77]. Similarly, incorporating passive sensing, especially for dietary habits, could reduce user burden and has been shown to reduce attrition rates [78]. User engagement can also be further improved by developing user-centered design elements, such as interfaces or interventions that have been personalized based on the social context of the users [79]. By integrating such features, JITAIs are likely to be perceived as more useful, engaging, and ultimately more effective in sustaining behavior change.

Across interventions, JITAIs demonstrated significant impacts on distal and proximal outcomes, such as weight loss, reductions in BMI, waist circumference, and blood pressure, as well as improvements in physical activity and dietary behavior. These outcomes reflect the ability and effectiveness of JITAIs to deliver adaptive, personalized, and timely support that targets the behavioral and contextual factors most relevant to each individual. By leveraging personalized data, JITAIs can improve weight management strategies by reducing sedentary time [41,42,55], encouraging adherence to calorie goals [26], and providing prompts that mitigate dietary lapses [10,32-37]. The dynamic and individualized features of JITAIs further help to sustain engagement and adherence, which are crucial for long-term behavior change [34,36,39,42,51]. Hence, these findings highlight the role of JITAIs in translating behavioral theory into practice, producing measurable improvements in lifestyle habits and health outcomes.