Developing a Computational Theory of Linguistic Alignment to Measure Engagement in Mhealth HIV Prevention Interventions

Introduction: Adolescent men who have sex with men (AMSM) account for 76% of HIV diagnoses among all young people and continue to face increasing incidence. Evidence-based mobile Health interventions (mHealth intervention) are designed to reach high risk HIV populations, sending scripted text messages of medication reminders and HIV prevention information. Yet, many mHealth interventions suffer from low usage, which reduces their public health impact. Text messages often ignore the participant’s linguistic background, a key component demonstrated to enhance participants’ satisfaction, which can potentially amplify mHealth interventions effects. In this study, we demonstrate how to optimize mHealth text-messaging interventions by analyzing linguistic style and alignment of mHealth users, as a first step to tailor the intervention messages. Such tailoring and linguistic analysis can increase engagement to and satisfaction with the intervention in an efficient, scalable, and non-obtrusive manner.

Methods: We developed a computational linguistic method to analyze the linguistic style of text messages (N=17784) exchanged by AMSM participants (N=132) aged 14 to 18, who are communicating in a peer-to-peer platform, developed for a completed, randomized controlled trial, Guy2Guy (G2G). This mHealth intervention sent prevention scripted messages related to safer sex and provided an interactive platform that matched AMSM into dyads for peer support and skill practice. Dyads were matched on age, geographical location, and sexual experience. We extracted linguistic features automatically and computed a linguistic alignment score between participants. Dyads were dichotomized in high or low engagement based on the number of messages exchanged. We trained a regression model to predict engagement and satisfaction based on dyad’s linguistic alignment.

Results: The messages exchanged within dyads varies widely (mean=269, sd=457). Preliminary evidence suggests that linguistic alignment is positively correlated with the number of messages exchanged in the peer support platform (p<0.002).

Conclusions: Computational linguistics methods have the potential to efficiently monitor engagement and satisfaction to the intervention. These methods can facilitate the implementation of future generation text-based mHealth HIV interventions in two ways: First, the analysis of linguistic style can reveal the linguistic inter-personal differences useful for tailoring the content and timing of text-messages. Second, detecting linguistic alignment is not easily done by independent observers even if they are trained. This computational algorithm is efficient, scalable, non-obtrusive, and automatic, which reduces the need to hire and train personnel dedicated to measuring key features of the mHealth interventions delivery and implementation. Guy2Guy mHealth intervention improves HIV testing and entry into prevention services, particularly for high risk adolescent MSM population, hence, it helps reduce HIV health disparities.