This study was designed as a 2-week, single-arm proof-of-concept study to evaluate the acceptability, adherence, and technical implementation of a smartphone-based, holistic just-in-time adaptive digital lifestyle intervention delivered by a rule-based CA. The interventions focused on promoting daily health behaviors across 7 domains relevant to cognitive and overall health, including physical activity, nutrition, sleep, cognitive exercises, hydration, and social engagement. The target population consisted of older adults with either SCD or MCI.

As a proof-of-concept study, the primary objectives were to assess plausibility, acceptability, and adherence patterns to the CA-initiated dialog sequences, and participants’ subjective evaluation of the intervention’s usability and acceptability. Secondary objectives included exploring the perceived working alliance with the CA and collecting qualitative feedback to inform future iterations of the intervention.

We adopted the JITAI framework to design a low-burden, chatbot-delivered digital lifestyle intervention tailored to the behavioral needs and capacities of older adults with SCD or MCI—individuals in the early stages of cognitive decline who generally retain the ability to provide reliable, momentary self-reports when prompted with low-burden, structured questions and limited recall windows [45].

Following the conceptual framework by Nahum-Shani et al [20], which defines key elements such as decision points, tailoring variables, intervention options, and decision rules, and proximal or distal outcomes (Table 1), the intervention was designed to assess a participant’s momentary states and only deliver a lifestyle suggestion when two criteria were met: (1) the participant had not yet engaged in the suggested lifestyle behavior (ie, determining the state of vulnerability or opportunity), and (2) the participant indicated availability and willingness to engage in the suggested lifestyle behavior (ie, state of receptivity).

Each day’s JITAI was structured around a decision point, occurring at a time predefined by each participant (1 PM, 2 PM, or 3 PM). At this time, the CA first assessed the participant’s state of vulnerability—a key tailoring variable—by asking whether they had already completed the targeted activity (eg, physical activity, hydration, or social contact).

If vulnerability (=an opportunity) was detected, because the response was “no,” a second tailoring variable—the participant’s state of receptivity was assessed to determine a second decision point. Therefore, the CA asked whether the individuals felt able and willing to complete the suggested activity later that day. Only when both vulnerability and receptivity were affirmed, the intervention option was triggered: a low-burden, tailored suggestion encouraging engagement in a target behavior (eg, “Why don’t you try and go for a 30-min walk. I’m sure you’ll feel great after”).

A third decision point occurred in the evening, where the CA inquired whether the participant had completed the suggested behavior. This response was used to capture the proximal outcome of adherence—that is, whether the support had led to real-world action that day.

Given the short duration and early-phase nature of this proof-of-concept, no distal outcome (eg, change in the QoL or long-term cognitive functioning) was assessed at this stage.

The complete mapping of each JITAI component to its operationalization in this study is also summarized in Table 1 and visualized in Figure 1. This table further clarifies how each core construct—decision points, tailoring variables, intervention options, decision rules, and outcomes—was instantiated within the JITAI logic of the Elsa/Erik study.

The smartphone-based JITAI was developed with the latest version of the MobileCoach software platform (April 2024) [46,47], which has already been used for various health interventions [31,37,39,40,48-51]. MobileCoach offers data collection capabilities (eg, via self-reports and interaction logs) and the delivery of JITAIs with the help of a rule-based CA. The user interface follows the principles of messaging apps, such as Apple’s iMessage or WhatsApp. However, answer options are predominantly predefined with a few exceptions (eg, when the CA asks for the nickname of the study participant). MobileCoach was hosted by Pathmate Technologies. Participants could choose from two CAs, Elsa or Erik, to be their digital coach over the 2-week study period. Figure 2 demonstrates the three conversational turns identifying the state of vulnerability, state of receptivity, and adherence to the CA-suggested health-promoting activity.

Participants were recruited in the Bern Memory Clinic and local cognitive training groups [52] in July 2024. First, a telephone screening for a first check of inclusion and exclusion criteria and a detailed explanation of the study was performed. An investigator explained to each participant the nature of the study, its purpose, the procedures involved, the expected duration, the potential risks and benefits, and any discomfort it may entail. As inclusion criteria, participants were required to be aged between 60 and 85 years, exhibit cognitive decline classified as SCD or MCI, possess a smartphone (either iOS or Android operated), demonstrate fluency in German, and have the capacity to provide written informed consent. Participants were excluded if they had severe psychiatric disorders or a history of substance abuse. Cognitive performance was assessed using the Montreal Cognitive Assessment (MoCA) score, with >25 indicating SCD, and scores equal to or below 25 indicating MCI [53,54] by expert investigators in a standardized clinical setting. Detailed study information and consent forms were sent to potential participants several days prior to the on-site visit. Participants gave written informed consent and were informed of their right to withdraw from the study at any time. Participation was voluntary, and no incentive for participation was provided.

During the on-site preintervention visit, demographic data, general cognitive performance (assessed using the MoCA Score), and technology readiness were evaluated. Questions on technological readiness included the duration, modality, and frequency of smartphone use (Multimedia Appendix 1).

Participants then downloaded the study app, MobileCoach, and received 2 weeks of JITAI promoting healthy lifestyle behaviors through their CA. The CA started a maximum of two daily dialogs with the participant, one in the afternoon (to assess the state of vulnerability and receptivity) and one in the evening (to assess whether the health-promoting activity was implemented) if the participant indicated that they were vulnerable in the afternoon. Participants preselected their preferred times for these dialogs with the CA: 1 PM, 2 PM, or 3 PM for the afternoon dialog and 7 PM, 8 PM, or 9 PM for the evening dialog. The notifications prompted health-promoting activities and assessed the participant’s vulnerability and receptivity to these activities. In total, the CA suggested health-promoting activities from 7 aspects of health, twice each during the 2-week study period (Table 2). The health-promoting activities were decided among three authors (EB, RV, and PJ), based on evidence-based domains affecting positive health outcomes in the literature [55].

After the 2-week study period, a follow-up telephone call was conducted to confirm study completion and gather feedback on app use. Participants were sent (digitally or by post) a postintervention questionnaire (Multimedia Appendix 2).

The primary end point is adherence to the CA-initiated conversational turns, measured objectively via app interaction logs. Data from the app were collected using the MobileCoach software platform in anonymized form using participant IDs. The platform allowed the data to be exported and further used for analyzing the measured metrics reported in the results section. States of vulnerability, receptivity, and adherence to both the CA-initiated conversational turns and CA-suggested health-promoting activities, as well as the daily response rates to vulnerability questions, were measured.

Secondary end points include perceptions of technology acceptance, working alliance with the CA, self-reported adherence to the CA-suggested health-promoting activity, and suggestions for improving the JITAI (conducted through a questionnaire and short telephone interview). As a clarification, adherence to the CA-initiated conversational turns refers to whether the participants responded to the first notification assessing their state of vulnerability. Adherence to the CA-suggested health-promoting activity refers to whether the participant was able to complete the suggested activity.

A postintervention questionnaire was used to assess technology acceptance constructs, based on prior work on CA [31]. Constructs measured in this study included perceived ease of use, enjoyment, usefulness (as a reminder of the health-promoting behavior and as a motivator to complete the activity), control over the CA, and also the intention to interact again with the CA [56-58]. Working alliance, an important relationship quality, has previously been shown to relate strongly to treatment outcomes [59-62] and was measured with the 6-item session alliance inventory [63]. A 7-point Likert scale was used for each item and anchored from 1=strongly disagree to 7=strongly agree.

During the follow-up telephone call, a short semistructured interview was conducted to ask participants: (1) What did you particularly like about the interaction with Elsa/Erik? (2) What needs to be improved about the interaction with Elsa/Erik? (3) Do you have any further comments or suggestions for improvement? The postintervention questionnaire also included these three open-ended questions, which were then updated and added to their interview responses. Additionally, participants were also provided with space to provide open feedback and details regarding their smartphone and mobile operating system.

Based on the mobile app data collected by MobileCoach, metrics for vulnerability, receptivity, and both adherences (ie, CA-initiated conversational turns and CA-suggested health-promoting activity) could be analyzed. Adherence to the CA-initiated conversational turns was calculated from the response to notifications on the status of vulnerability—the first afternoon message. Participants were categorized as vulnerable (who had not yet completed the activity), not vulnerable (who had already completed the activity), or nonresponsive. Receptivity was assessed only among vulnerable participants, who were further categorized based on their intent to complete the activity. Participants were either classified as receptive (intended to complete the activity), not receptive (did not intend to complete the activity), or nonresponsive. Proportions were calculated by dividing the number of participants in each category by the total number of vulnerable participants. Adherence to the CA-suggested health-promoting activity was measured by tracking whether participants who were vulnerable in the afternoon completed the activity by the evening notification time. Among these participants, those who completed the activity were considered to have adhered, while those who did not were categorized as not adhered. The proportion of adherence was calculated by dividing the number of participants who completed the activity by the total number of vulnerable participants who responded to the evening message.

For the self-report items of the postintervention questionnaires, we conducted descriptive statistics to report mean and SD calculations. Shapiro-Wilk tests indicated that the self-reported data were not normally distributed, and therefore, Wilcoxon signed rank tests were conducted to assess whether the median of the evaluations significantly differed from the test value of 4, representing a neutral response on the 7-point Likert scale. An overall score for the 6-item session alliance inventory was computed by averaging the six individual items, as Cronbach α of 0.90 demonstrated high internal consistency [9].

Inductive thematic analysis was conducted with the interview responses using MAXQDA (VERBI) [64]. Codes were determined from the raw questionnaire and interview data (translated into English) by authors EB and RV and further classified into overarching themes based on their trends. Two rounds of thematic analysis were conducted, and the results are presented as code and thematic maps, also using MAXQDA.

The study was reviewed by the local Ethics Committee in Bern (identifier 2024‐01029) and categorized as not falling under the Human Research Act [65]. Written informed consent was obtained from all participants in the study language, German. Participants were sent a study information sheet that included the inclusion and exclusion criteria, detailed information about the study, and their rights as participants, as well as a consent form for participation in the study and a consent form for further use of data. All data collected in this study were anonymized, and no identification of individual participants in the data, study, or supplementary files is possible. Collection, processing, and storage procedures for this study strictly adhered to Swiss legal and ethical frameworks, including the Swiss Federal Act on Research involving Human Beings (Human Research Act) and the new Federal Act on Data Protection. Participants were not provided compensation for the study.

Overall, 17 participants were recruited for the study. One participant was excluded due to outdated hardware, and another participant due to technical issues (notifications could not be delivered), resulting in a total analyzable sample of 15 participants. All participants were assigned to the CA-delivered app intervention in this single-arm study, and reporting followed the CONSORT (Consolidated Standards of Reporting Trials) checklist (Checklist 1). Participant sample characteristics indicated 67% (n=10) identified as female, and 33% (n=5) as male, where the mean age was 70.3 (SD 5.01) years, and the mean MoCA Score was 27.73 (SD 2.01). Based on the suggested cut-off [53,54], 12 participants were classified as having SCD and 3 as MCI. There was no significant difference between SCD and MCI participants in any of the technology readiness outcomes described in Table 3.

On average, 81% (mean 12.14) of participants responded to the afternoon notification each day over the 14-day study period (SD 1.51 responses per day to the state of vulnerability dialogue), demonstrating adherence to the CA-initiated conversational turns. Conversely, 19% (mean 2.86, SD 1.51) of participants, on average, did not respond each day. A detailed breakdown of the daily response rates according to the response type is presented in Figure 3.

For each aspect of health, participant engagement was analyzed across three key stages: determining vulnerability, identifying receptivity, and measuring adherence to the CA-suggested health-promoting activity. Table 4 presents these metrics as percentages, with results grouped by specific activities. On average, 27% (mean 4.07, SD 2.27) of participants reported being vulnerable, and of these, 100% responded to the state of receptivity notification. An average of 83% (mean 3.14, SD 1.61) of participants were receptive to completing the health-promoting activity, with 69% (mean 2.86, SD 1.70) of participants self-reporting adherence to the activity. An average of 95% (mean 3.93, SD 2.27) of participants responded to the adherence evening notification.

However, engagement patterns varied significantly across health domains. For example, the cognitive exercise domain showed the highest combined engagement: 43% (n=6.5) of participants reported being vulnerable, 70% (n=4.5) were receptive, and 55% (n=3.5) adhered to the suggested activity. In contrast, sleep was the domain with the lowest proportion of vulnerability: only 1 (7% of all participants) person indicated being vulnerable at all who was receptive on both days and consistently reported that they had already engaged in the recommended sleep behavior by the time the just-in-time prompt was delivered. This suggests that many participants already had stable evening routines or did not perceive a need for behavioral adjustment in this area.

Participants were asked to rate their experience interacting with the CAs Elsa/Erik on a 7-point scale (where 1=strongly disagree and 7 =strongly agree). Further, participants rated their perceived relationship with Elsa/Erik on a 7-point scale (where 1=never and 7=always). The results are displayed in Table 5.

Results indicate that all mean values lie above the neutral scale value of 4. Moreover, Wilcoxon signed rank tests indicate that the median of perceived ease of use and two items of the session alliance inventory (items 1 and 2) were significantly higher compared to the neutral test value of 4. Additionally, the median values of perceived usefulness as a reminder (PU1), perceived enjoyment, perceived control, and the overall session alliance inventory items were also rated highly. The remaining construct items (perceived usefulness as a motivator [PU2] and intention to interact) still had medians above the value of 4. Therefore, it can be concluded that participants evaluated the CAs Elsa/Erik favorably.

To complement the quantitative findings, all participants were contacted postintervention for a short semistructured interview (approximately 10 min). Results focusing on (1) positive feedback, (2) potential for improvement, and (3) open remarks (cf questions mentioned in the Methods) were thematically analyzed and coded. Figure 4, therefore, presents a thematic map of two overarching themes—app enjoyment and enhancing engagement—along with related subthemes and their frequency counts. Multimedia Appendix 3further provides direct participant quotes to illustrate each subtheme and clarify sentiment.

Participants generally appreciated the overall functioning of the app (both visually and technically), with several describing it as “easy to use” (P06, P07, P12, and P17), “appealing, nothing to criticize” (P14), or “the design was ok” (P10, P11, and P12). Participants also responded positively to the tone and conversational style of the agent: “There were no accusations if something wasn’t done, which was very pleasant” (P07), but also highlighted the need for adaptations for older adults: “The language was quite youthful, it should be adjusted since the app is for older people” (P09). The timing and delivery of notifications also emerged as an area for improvement, with several users noting they had missed prompts or found the timing suboptimal: “It might be better if the notifications came in the morning […] especially if you are still working” (P03, P04, and P05).

This proof-of-concept study was deliberately designed as a formative, discovery-oriented investigation rather than a hypothesis-driven trial. This said, no a priori benchmarks for adherence or acceptability were defined, given the novelty of the target population and the limited existing evidence on smartphone-based and CA-driven JITAIs in individuals with SCD or MCI. Instead, this design allowed for open-ended observations of use patterns and participant responses to interventions. While this approach supported initial insights into the plausibility and user engagement of the JITAI, future studies will benefit from the use of literature- or theory-informed benchmarks to support clearer interpretability and “go/no-go” decision-making as suggested, for instance, by emerging frameworks such as the Digital Therapeutics Real-World Evidence framework [68].

Another key limitation of the current JITAI design lies in its reliance on self-reported data only to assess both states of vulnerability, receptivity, and adherence. While the target population—individuals with SCD and MCI—is generally capable of providing momentary self-reports over short timeframes, this approach may not scale to individuals with more advanced cognitive decline [45]. To minimize cognitive burden, we limited the recall window to same-day behaviors and used simple, structured prompts. Nevertheless, future iterations of the intervention could benefit from the integration of passive sensing modalities (eg, step count, sleep patterns, location-based context, or even social interaction cues from Bluetooth proximity) to reduce reliance on self-reports and enable more accurate, inclusive data collection. Wearable devices, such as smartwatches or fitness bands, hold particular promise for older adults with cognitive impairments [69]. In addition to enabling continuous, real-time tracking of health behaviors like physical activity, sleep wearables offer a more reliable notification interface: unlike smartphones, which may be overlooked or out of reach, wearables provide discreet, wrist-based notifications through visual cues or haptic feedback, increasing the likelihood that prompts are noticed and acted upon [70].

Further, this early-stage proof-of-concept study was designed to investigate plausibility-related outcomes such as acceptability, usability, and engagement, rather than to demonstrate broad effectiveness. The outcomes observed here reflect the experiences of a predominantly female and technology-aware sample, reflecting common patterns in research study participation, but limit generalizability across genders.

Additionally, embedding the JITAI into a stepped-care program or study with in-person contacts, as a form of blended treatment [71], would allow for additional collection of proxy (or “other”) reports from caregivers or clinicians, which could serve as alternative data source to strengthen the reliability of the decision logic. A blended approach could also offer the necessary support to enhance the overall effectiveness and user experience with the hardware [72,73].

The use of a rather simple, low-burden rule-based CA further constrained the depth and flexibility of user interactions. Participants could only select from predefined response options, and the CA delivered one daily activity suggestion based solely on binary assessments of vulnerability and receptivity. This approach did not accommodate individual preferences or provide alternative options if the suggested behavior was, for instance, irrelevant (cf the low vulnerability reported for days with a sleep intervention). Indeed, several participants expressed a desire for more contextual flexibility and choice. While this JITAI was intentionally designed to reduce cognitive load and mitigate confusion or misinformation, its limited adaptability also constrained the intervention’s alignment with the defining characteristics of a “true” JITAI, which requires nuanced, context-sensitive tailoring over time. Future research may address this limitation by exploring hybrid architectures that retain the safety and clarity of rule-based logic while incorporating the adaptive potential of generative artificial intelligence (eg, large language models). Such models may allow for richer personalization, deeper therapeutic engagement, and more responsive intervention delivery—provided that appropriate safeguards and oversight mechanisms are in place to preserve safety and therapeutic integrity for cognitively vulnerable users [74,75].

The multicomponent approach used in this study, which integrates various aspects of health such as sleep, hydration, physical fitness, and cognitive exercises, is particularly promising for enhancing the QoL in older adults. By addressing multiple dimensions of well-being simultaneously, this approach has the potential to produce more comprehensive and sustained benefits compared to interventions that focus on a single health aspect. For older adults, maintaining a balance across these domains is crucial, as deficits in one area, like poor sleep or inadequate hydration, can negatively impact other aspects of health, such as cognitive function or physical mobility. By promoting holistic health behaviors, the JITAI not only supports cognitive health but also contributes to overall well-being, potentially delaying the onset of more severe impairments and enhancing daily functioning. This integrated strategy aligns with current geriatric care principles, which emphasize the importance of a well-rounded approach to aging and hold great promise for improving the QoL in this population.