We conducted qualitative in-depth interviews with family caregivers of people with dementia, policymakers involved in planning dementia programs, clinicians, and community dementia service providers to identify desired features for a mobile app designed for caregivers of individuals with dementia. Using purposive sampling, we recruited family caregivers through referrals from community dementia organizations and an ongoing study, whereas policymakers, clinicians, and service providers were approached through key contacts in relevant organizations. A trained qualitative researcher conducted interviews, which were audio recorded, transcribed, and analyzed on NVivo 11 (Lumivero) through content analysis.

Two informal caregivers of individuals with dementia guided the research team on the design and content of the CareBuddy app. Study team members met with them frequently to elicit feedback.

Guided by the stress process model [31], we developed a logic model to develop the “CareBuddy” app (Figure 1). The stress process model, an established framework, describes the relationship between caregiving stressors, psychosocial resources, and caregivers’ psychosocial well-being [31,32]. The framework thus provides a foundation to develop an integrated caregiving ecosystem that addresses caregiving stressors and delivers comprehensive support by enhancing caregivers’ access to tailored resources, coping strategies, and social support networks.

We conducted pilot testing in 2 phases using a mixed-method design incorporating both qualitative feedback and quantitative assessments for app usability.

In this phase, we aimed to gather feedback on the app’s content and functionality. Participants interacted with the prototype version of the app using the interviewer’s smartphone. The interviewer then guided the participant through the app’s various components using a topic guide to prompt open-ended discussions. This phase involved 3 rounds of in-depth interviews, with the app iteratively updated after each round based on participants’ feedback. The interview guide, developed by the study team, explored key aspects of user experience, app functionality, the relevance of the app’s solutions for managing behavioral and physical symptoms of individuals with dementia, perceived usefulness, and suggestions for improvement. We also collected demographic data and assessed usability using the 10-item System Usability Scale (SUS) [33].

We purposively sampled family caregivers and health care providers. Eligible caregivers were Singaporean or permanent residents aged ≥21 years, involved in making treatment decisions for a family member with dementia, had been providing care for at least 6 months and were planning to do so for the next 2 months, able to read and write English, and had intact cognition as determined through the Abbreviated Mental Test (only for those aged ≥65 y). Health care providers included those providing health or social care within community organizations or hospitals. The interviews were conducted in English by a trained researcher with no prior relationship with the participants. Sessions were held at a time and place convenient for the participants, lasting 1 to 2 hours.

This phase assessed caregivers’ app engagement, usability dimensions including satisfaction and acceptability, and subjective experiences. Caregivers downloaded the app (iOS or Android version) onto their own mobile phones, with a researcher providing an orientation on its features and navigation. If the care recipient was mobile, caregivers additionally received a GPS tracker for tracking their location.

Eligibility criteria for family caregivers remained the same as in phase 1, with the additional requirement of owning an Android or iOS mobile phone. Caregivers of individuals residing in long-term care facilities were excluded.

Following informed consent, we collected demographic data for the caregivers (eg, age, gender, relationship, and education) and the individual with dementia (eg, age and gender) through a survey. At the end of the 1-month trial, app usage was tracked throughout the month, including average engagement time per week, time spent on each component, and frequently used components. We administered the Mobile Health App Usability Questionnaire (MAUQ), English stand-alone patient version, for the intended audience, family caregivers of individuals with dementia [34]. MAUQ, a validated 18-item questionnaire, assesses usability in 3 dimensions (ie, ease of use, interface and satisfaction, and usefulness) on a 7-point Likert scale, ranging from 1 (strongly disagree) to 7 (strongly agree) [34]. Caregivers also participated in open-ended interviews exploring their experiences and subjective feedback on the relevance and perceived usefulness of the app. The open-ended portion was audio recorded and transcribed. Interviews lasted 10 to 25 minutes.

We summarized demographics and calculated SUS scores both overall and by interview round to assess the app’s usability evolution. For the phase 1 pilot, we compared the SUS scores to the benchmark mean of 68, which indicates good usability [35,36]. In the phase 2 pilot, we calculated total and average scores assessing MAUQ usability dimensions, that is, ease of use, interface and satisfaction, and usefulness, including satisfaction and acceptability, with higher total and average scores indicating greater usability [34]. An average score lower than 4 meant poor usability [37]. We also summarized app usage data in terms of the number of hours used, components used, and time spent on each component in phase 2.

Using NVivo 11 software, 2 team members independently coded transcripts and conducted content analysis [38]. We developed a code book based on the User Version of the Mobile Application Rating Scale [39], categorizing feedback into key domains: engagement, functionality, aesthetics, information quality and quantity, and subjective feedback. We synthesized user feedback within these domains and incorporated suggestions for refining the app.

We triangulated quantitative and qualitative findings on usability and engagement.

On the basis of the user requirements and the stress process framework–based logic model, we developed app components to address both primary and secondary stressors (Figure 1).

The app targets primary caregiving stressors such as behavioral symptoms (eg, aggression, agitation, mood swings, hallucinations, disorientation, and refusal of care) and physical challenges (eg, mobility issues, eating difficulties, dry skin, toileting problems, communication difficulties, and sleep disturbances). Upon logging in, caregivers are guided through an assessment using a series of questions to identify current behavioral and physical symptoms, as well as their potential underlying causes. On this input, the app offers tailored, practical management strategies. Specifically, for the care recipient’s behaviors, the approach uses the unmet needs model (which views behaviors as expressions of unarticulated needs), the progressively lowered stress threshold model (which emphasizes the reduced capacity of the care recipient to deal with stress or stimuli), and the ABC (antecedent, behavior, and consequences) models. The solutions are further derived using the DICE (describe, investigate, create, evaluate) and P.I.E.C.E.S. (Physical. Intellectual. Emotional. Capabilities. Environment. Social.) frameworks [40,41]. The content was developed in consultation with dementia care providers and from existing resources published globally. All content was reviewed and validated by 2 geriatricians on the study team with expertise in dementia care.

A particularly difficult behavioral symptom, wandering, is addressed through a GPS location monitoring function integrated into the app. This feature allows caregivers to monitor the real-time location of their care recipients using GPS-enabled trackers. Caregivers can view the last detected location of their care recipients directly through Google Maps. Additionally, the app provides step-by-step navigation to the tracked location. The app monitors the GPS tracker’s battery level and provides reminders to recharge the battery.

To address primary stressors, CareBuddy also facilitates collaborative care management through a care plan function, which supports information sharing between family caregivers and health care professionals. The Care Plan includes a section on the individual’s hobbies and interests to promote person-centered care. Caregivers can also use a centralized messaging tool to share updates and discuss care plans with providers. The app also allows caregivers to apply for care center transfers, with approvals managed by health care providers. A secure QR code–based system enables multiple caregivers (eg, family members or friends) to coordinate care efficiently. By streamlining communication and care planning, CareBuddy strengthens coordinated, person-centered care.

To ensure timely access to health services, CareBuddy integrates telemedicine access via deep links to third-party provider apps. Users can initiate virtual consultations with established general practitioners directly through CareBuddy, eliminating the need to search for separate telemedicine apps. This integration enhances continuity within CareBuddy’s digital ecosystem. Instructional content is also provided to guide caregivers on using these services effectively.

In moments of acute distress, caregivers can also access a crisis support helpline within the app, connecting them directly to trained social workers who provide emotional support, practical guidance, and crisis intervention.

A large language model (LLM)–powered chatbot is integrated into the app to provide personalized, real-time responses to caregivers’ queries by offering advice, summarizing key strategies, and encouraging engagement with relevant app components. This chatbot uses GPT-4 as the default language model for generating responses and GPT-3.5 for contextual analysis based on the knowledge database. The system is built on a retrieval-augmented generation architecture, where relevant knowledge items are retrieved from the database and used as context in prompts to the language model. This enables the chatbot to generate responses that are both contextually accurate and conversationally coherent. The chatbot integrates both an internal knowledge repository and external knowledge sources to generate accurate, contextually relevant responses. The internal repository comprises a structured question-and-answer (Q&A) database, curated by the study team to address key caregiving stressors. To enhance response accuracy, we implemented natural language processing to semantically map user queries to the most relevant Q&A entries, ensuring precise information retrieval. A tiered, confidence-based retrieval system further optimizes response generation, with confidence scores derived from cosine similarity between user queries and knowledge-based questions. Queries scoring above 95% are directly matched to Q&A entries with LLMs adding empathetic context, scores between 90% and 95% synthesize information across multiple sources, and scores below 90% engage ChatGPT with domain-specific context for broader, tailored responses. These thresholds were empirically determined through user testing to balance accuracy for specific queries with generalization for broader or ambiguous ones. Testing confirmed that high-confidence matches yield precise answers, whereas lower-confidence matches generate broader, contextually relevant summaries. To ensure safety, the chatbot responds strictly within its knowledge scope at high confidence and, in urgent or sensitive cases, directs users to immediate resources such as helplines rather than attempting diagnostic or therapeutic advice. Personal information is coded, protected, and delinked from individual users, ensuring all chatbot conversations are completely anonymous.

To address secondary stressors such as financial burden and future care planning, the app offers curated content on financial resources and subsidies, along with guidance on how to do advance care planning, power of attorney, making a will, and Central Provident Fund nominations. Caregivers also receive information on the anticipatory trajectory of illness for dementia so that they can plan for the evolving needs of their care recipient as the condition progresses.

To reduce internal caregiver stress and improve emotional well-being, CareBuddy includes a dedicated self-care module featuring psychologist-developed worksheets. These resources help caregivers reflect on their caregiving identity and journey, recognize signs of burnout, engage in self-care, and build coping strategies for anticipatory grief and emotional resilience.

The app also provides access to the user’s phone calendar, which allows caregivers to create personalized event reminders. This function helps users keep track of important dates, appointments, or caregiving tasks and provides timely alerts to ensure they receive the necessary prompts to take action. This feature responds directly to caregiver feedback and is designed to support daily organization.

Finally, CareBuddy facilitates peer support through an in-app online forum, available both on the app and via a web portal. This space allows caregivers to seek help, share experiences, offer support, get connected with other caregivers, and build community. Users can flag inappropriate posts, block others if needed, and benefit from a backend “Censor Curse Words” function that ensures respectful communication. An LLM-based artificial intelligence (AI) bot named “Sparky” is available in the online forum to stimulate user interactions and facilitate engagement with caregivers. The AI bot “Sparky” generates initial replies to caregivers’ new posts and provides comments when requested by caregivers.

The CareBuddy app interface is intentionally designed to be user-friendly, with frequently used functions placed in prominent positions and enhanced with high-resolution icons. A text-to-speech feature allows users to listen to information instead of reading, supporting accessibility. Onboarding instructions and reminders are integrated to support user navigation and regular engagement.

The CareBuddy caregiver app is connected to a dedicated health care provider version that supports them in collaborating with family caregivers. Within the health care provider version interface, users can approve new clients with dementia, update client information, and engage in care planning discussions with caregivers. The health care provider version also enables providers to participate in the online forum, exchanging insights and offering support to caregivers and other stakeholders.

The CareBuddy caregiver app is also connected to a moderator version. Moderators of the online forum can manage the forum via the CareBuddy web portal, with tools to pin or feature posts and share multimedia caregiving resources.

CareBuddy includes an embedded system usage log that captures detailed user activity across the app. This feature tracks user actions to better understand caregiver needs and app engagement patterns. In addition, a data-capturing component tracks interactions within the online forum, including posts, threads, and social dynamics among users. These built-in analytics features support future improvements and enable ongoing research into user behavior.

CareBuddy integrates a comprehensive notification system using Firebase Cloud Messaging to keep users engaged and informed. The app supports system-scheduled notifications that provide timely nudges for user engagement and updates within the CareBuddy environment. In-app notifications are triggered by specific user interactions, such as messaging between family caregivers and health care professionals, ensuring users stay informed of meaningful exchanges and collaborative care efforts. Additionally, push notifications are delivered even when the app is closed, allowing critical information, such as new messages, care plan updates, or forum interactions, to reach users without delay. These notifications are customized based on user activity and context, including upcoming appointments or scheduled reminders.

This feature was designed to support comprehensive storage, retrieval, and analysis of past user interactions, incorporating user metadata, conversation sessions, and analytical reporting to continuously refine chatbot performance and personalization.

Data privacy and security measures were implemented for proper handling of sensitive features and data within the app. GPS location data and chatbot records are stored in an anonymized format and not linked to personally identifiable information in the backend database. Thus, data are decoupled from patients’ profile data, and identifiable information is separately stored and managed by the research team.

Before the CareBuddy mobile app was made available for download (on both the Apple App Store and Google Play Store), the app was reviewed by both platforms. CareBuddy’s adherence to best practice measures ensured it satisfied both platforms’ prevailing compliance and regulatory requirements and ensured privacy and security protection for end users. Furthermore, the privacy policy and End User License Agreement were also made available to end users to elaborate on how content filtering and moderation are implemented to ensure rigor, appropriateness, and security of the app. CareBuddy app users also provided their consent and acceptance of the service terms during the account registration process.

Moreover, all user data and interaction records (including data related to sensitive features) are stored securely in the Amazon Web Services cloud, with its access strictly limited to the designated system administrator from the research team and further protected with multifactor authentication for enhanced security. These measures ensured that user privacy, security, and ethical considerations were upheld throughout the development and pilot testing.

The CareBuddy system structure is presented in Figure 2, and sample screenshots of the components, interface, and features are presented in Figure S1 in Multimedia Appendix 1.

The analysis of substantive interview feedback and suggestions from phases 1 and 2 is summarized under 5 domains of the User Version of the Mobile Application Rating Scale, with representative quotes from participants.

Participants in phase 1 found the information credible, high quality, and easy to understand, covering relevant caregiving topics. Users in round 1 found the text wordy, suggesting bullet points, visuals, and a ‘view more’ option for details. With progressive iterations and changes made in response to suggestions from earlier rounds, most participants in round 3 were satisfied with content clarity and organization. They perceived the content to be rich, straightforward, containing extensive resources, and providing solutions for managing common behavioral symptoms. One health care provider mentioned:

Caregivers in phase 2 valued the content on dementia trajectory, management of behavioral and physical symptoms, predeath grief, stress management, and self-care content, finding it relatable. They found the content well organized and easy to use. In comparison to existing local apps, they found CareBuddy comprehensive and eliminated the need for search engines such as Google.

While most participants in phase 1 found the app easy to use, some struggled with navigation and chatbot performance. Participants suggested making the chatbot responses more empathetic. In response, we enhanced navigation by adding subheaders for each section and incorporating hyperlinks to external resources. We also improved the chatbot’s vocabulary, ensuring sensitivity for better responsiveness. A health care provider noted:

In the 1 month of using the app in phase 2, caregivers found the search function in the directory intuitive. They also used the interactive app elements such as a GPS tracker, worksheets, a chatbot, and a directory extensively. Most participants found no issues with navigation; however, some reported missing links to external sites and noted that the tracker function required frequent refreshing to update the care recipient’s location.

Early feedback from phase 1 cited small fonts, cluttered layout, and dull colors. In response, we made adjustments to text size, formatting, and color contrast for improved readability. We added an audio option for users who preferred listening to the content. Later feedback was very positive. A health care provider reported:

Many participants in phase 1 showed interest and eagerness to read the app content and suggested providing empathetic responses to ensure appropriateness. Moreover, providing practical solutions rather than theoretical responses ensured engagement with the app.

In phase 2, at least 40% of caregivers used the app every day. Many participants had positive feedback and felt happy to connect with fellow caregivers on the community portal. Others preferred to just hear others and found information shared by other caregivers to be helpful. They reported that the moderators did a good job of initiating discussions and engaging users. One caregiver reported:

Feedback from participants in early phase 1 was mixed. Some felt the app to be more relevant for early dementia diagnosis, whereas others believed it was suited for later-stage dementia care. In response, we enhanced the content of the app, incorporating more examples of how care can be provided. Caregivers in round 3 perceived the app to be more useful, functioning as a valuable guide and offering real-time support with practical solutions for addressing behavioral and physical symptoms. Some of the participants were keen and excited about the launch of the app due to the lack of a comprehensive app such as CareBuddy. A caregiver noted that:

A health care provider also reported that:

Over a 1-month trial, caregivers found the app solutions helpful in managing behavioral challenges (eg, agitation). The easy access to resources and content enabled a better understanding of dementia progression. Caregivers reported that the curated content saved time, providing quick access to care resources such as nursing homes, respite care, and mental health helplines:

Participants appreciated the CareBuddy as comprehensive and were keen to recommend the app to others. One of the caregivers reminisced that having the CareBuddy app would have been immensely valuable when their loved one had been newly diagnosed with dementia.

The changes made to the app based on participants’ suggestions by phases are described in Table S2 in Multimedia Appendix 1, and qualitative analysis with illustrative quotes is detailed in Multimedia Appendix 2.