Annual IEEE EMBS Benelux Chapter Leuven, 28 th November 2019 18 th National Day on Biomedical Engineering Brussels, 29 th November 2019 EVALUATION OF PERSONALIZED PERSUASIVE MESSAGING FOR HEALTH BEHAVIOR CHANGE Jens d’Hondt, Raoul Nuijten, and Pieter Van Gorp * Eindhoven University of Technology, Eindhoven 5612 AZ, The Netherlands [ * as presenting author ] Keyword(s): Artificial Intelligence for Behavior Change, Persuasive Health Technologies 1. INTRODUCTION Tailoring persuasion strategies to the personality of the recipient (adaptive persuasion) has shown to be effective in a variety of contexts (i.e., not only in marketing but also in healthcare [1]). However, most studies only measure the effectiveness of persuasive approaches on short- term behavior and attitudes (e.g. direct feedback or click-through rates) without measuring their effect on long-term behavior. Therefore, we have evaluated whether in the context of health promotion, adaptive persuasion not only improves attitudes and intentions, but also has the potential to induce actual behavioral change. 2. MATERIALS AND METHODS We conducted a randomized controlled trial for evaluating an adaptive persuasive messaging system in the context of a digital health promotion intervention for employees and students of the University of Technology in Eindhoven (N=149, duration: 6 weeks). The persuasive messaging system was built as a personalized e-mail agent as an extension to the mature mHealth platform GameBus [2]. The extensions were based on state-of-the-art requirements of adaptive persuasive technologies [3]: (1) the system extension incorporated user-level profile building support, (2) persuasive messages were based on a theory-based psychological profiles, (3) the approach was algorithmic, (4) objective measures were used to assess user appreciation for the messages delivered and (5) the system incorporated rules for continuous learning. The trial control group (n=74) received messages based on a random profile from the Big-Five personality types while the trial intervention group (n=75) received messages based on the personalized system. Multiple generalized linear (mixed effect) models were fitted on the data for different effectiveness metrics and compared to each other using χ 2 tests. The following effectiveness variables were chosen: (1) Message Feedback (MF) as collected systematically for all messages delivered, (2) Performed Activities (PA) as the number of activities logged in GameBus, and (3) Message Success (MS) as the percentage of messages causing an activity being performed. Multiple models were compared to a so-called “null” model, which expressed individual-level model intercepts as a random effect [4]. 3. RESULTS AND DISCUSSION Analysis of the linear models suggest that the treatment group did evaluate MF higher than the control group but at the same time the treatment group did not perform more physical activities PA. We therefore advise a more skeptical towards the longer-term effectiveness of adaptive persuasive techniques and we aim to design technical extensions for longer term changes in behavior. In future work we will extend sample sizes and trial durations while also improving learning algorithms and message delivery technologies. References [1] M. Morris and F. Guilak. “Mobile Heart Health: Project highlight”. In: IEEE Pervasive Computing (2009), pp. 57–61. [2] A. Shahrestani, P. Van Gorp, P. Le Blanc, F. Greidanus, K. de Groot and J. Leermakers, “Unified Health Gamification can significantly improve well-being in corporate environments," 39th Int. Conf. of the IEEE EMBC, Seogwipo, 2017, pp. 4507-4511. [3] M. Kaptein and A. Van Halteren. “Adaptive persuasive messaging to increase service retention: Using persuasion profiles to increase the effectiveness of email reminders”. In: Personal and Ubiquitous Computing 17.6 (2013), pp. 1173–1185. [4] J. d’Hondt et al. “Evaluation of computer- tailored motivational messaging in a health promotion context”. 11th Int. Conf. on Modeling and Using Context. Nov. 2019.