Human Cues in Self-help Lifestyle Interventions: an Experimental Field Study (Preprint)

Abstract

Background: Self-help eHealth interventions are generally less effective than human-supported ones, as they suffer from a low level of adherence. Nevertheless, self-help interventions are useful in the prevention of non-communicable diseases, as they are easier and cheaper to widely implement. Adding humanness in the form of a text-based conversational agent (TCA) could provide a solution to non-adherence. In this study we investigate whether adding human cues to a TCA facilitates relationship-building with the agent, and makes interventions more attractive for people to adhere to. We will investigate the effects of two types of human cues, which are visual cues (eg, human avatar) and relational cues (eg, showing empathy). \nObjective: We aim to investigate if adding human cues to a TCA can help increase adherence to a self-help eHealth lifestyle intervention and explore the role of working alliance as a possible mediator of this relationship. \nMethods: Participants (N=121) followed a 3-week app-based physical activity intervention delivered by a TCA. Two types of human cues used by the TCA were manipulated, resulting in four experimental groups, which were (1) visual cues-group, (2) relational cues-group, (3) both visual and relational cues-group, and (4) no cues-group. Participants filled out the Working Alliance Inventory Short Revised form after the final day of the intervention. Adherence was measured as number of days participants responded to the messages of the TCA. \nResults: One-way ANOVA revealed a significant difference for adherence between conditions. Against our expectations, the groups with visual cues showed lower adherence compared to those with relational only or no cues (t(117) = -3.415, P = .001). No significant difference was found between the relational- and no cues-groups. Working alliance was not affected by cue-type, but showed to have a significant positive relationship with adherence (t(75) = 4.136, P < .001). \nConclusions: We hypothesize that the negative effect of visual cues is due to a lack of transparency about the true nature of the coach. Visual resemblance of a human coach could have led to high expectations that could not be met by our digital coach. Furthermore, the inability of TCAs to use non-verbal communication could provide an explanation for the lack of effect of relational cues or the effect of cue-type on working alliance. We give suggestions for future studies to test these potential mechanisms. Clinical Trial: Pre-registration: OSF Registries, https://osf.io/mgw2s