AI-Driven Content for Gamified Digital Health Applications: Providing healthcare personnel tools for the creation, refinement, and structuring of personalized content

Individuals in rehabilitation or treatment have a treatment plan that describes the goals that they are going to work on, which can be used as personalized content in gamified health applications. However, the creation, refinement, and structuring of personalized goals (and broader content) in gamified health applications is a time-consuming process in need of a systematic approach that augments the ability of healthcare professionals and individuals in these tasks. In a randomized trial, patients with SMI preferred personalized goals but found them more stressful, likely due to participants not experiencing flow, due to a mismatched difficulty progression despite case manager involvement. Through an iterative process, we developed and evaluated the GOALS system, an LLM-based goal refinement system for individuals to create well-defined goals in collaboration with their healthcare professionals, for use in gamified digital health applications. Using prompt- and software-engineering, we found that LLMs can be used to create well-defined, measurable goals that are grounded in behavior change and goal-setting theory. In evaluations, healthcare professionals evaluated LLM-generated goals significantly higher than the goals they had created in collaboration with their patients, and considered the LLM-based goal-setting process significantly more time-efficient. When LLM-generated goals were empirically evaluated in an RCT with students and staff members of universities, there was no significant difference in engagement between participants given LLM-generated lifestyle goals compared to those given one-size-fits-all lifestyle goals. The results of the RCT indicate that LLM-based goal-setting is more time-efficient for health interventions, without the risk of participants losing engagement with the intervention content. However, the RCT also indicated a need for goal refinement, as LLM-generated goals were considered generic over time. To ethically enable LLM-based goal refinement, workshops revealed that stakeholders expressed willingness to share context-sensitive data with LLMs if there is control over their data, transparency in data handling, local app storage, and a clear value proposition in return for sharing such data. We also developed and verified the Automated Planning of Level Systems (APLES) system, a tool to aid in the systematic structuring of balanced level systems in digital health interventions. A technical feasibility study demonstrated that structuring health intervention content in a level system can be modelled as a state space problem. In verification, APLES demonstrated its ability to create balanced level systems in terms of difficulty and fun ratio when given predetermined difficulty, fun graphs, and content. When comparing level systems in a health intervention conducted with students and staff members of universities, participants, when given the option, chose to do activities within their comfort zone and avoid types of activities that are healthy. However, there was no significant difference in user engagement between level systems that gave participants less autonomy and required them to do healthy activities outside of their comfort zone. In future work, we recommend that researchers explore health promotion within already existing entertainment games by introducing the Health Intervention Minigame Framework, designed to facilitate the replacement of dark patterns in entertainment games with health activities found in health interventions as a starting point. In conclusion, leveraging AI-based tools to support the creation, refinement, and structuring of digital health content can help deliver more effective and personalized experiences for individuals, particularly in vulnerable populations. The tools explored augment the ability of healthcare professionals and researchers in promoting healthy behavior change.