Author(s): LU, N., LAU, W., Institution: HONG KONG BAPTIST UNIVERSITY, Country: HONG KONG, Abstract-ID: 770

INTRODUCTION:
Social cognitive theory (SCT) has been widely applied to explain adolescents’ physical activity (PA) through constructs such as self-efficacy, outcome expectations, social support, and goal-related processes. However, limited evidence exists regarding whether these mechanisms operate similarly within generative artificial intelligence (GenAI)–based interventions. This study aimed to test SCT as a theoretical framework for explaining PA changes in a 4-week GenAI-based intervention among healthy adolescents using a structural equation modeling (SEM) approach.
METHODS:
A 4-week pre–post intervention was conducted among 200 adolescents aged 12–15 years recruited from three middle schools in Xinxiang, China. Participants attended an SCT-informed intervention that included an in-person component emphasizing peer competition to enhance perceived social support and engagement. Meanwhile, the online component was delivered via a GenAI-driven app developed by our team, providing personalized PA prescriptions, daily reminders, self-monitoring, and adaptive feedback to target self-efficacy, outcome expectations, and goal-setting. PA was assessed using the Physical Activity Questionnaire for Adolescents (PAQ-A). SCT constructs, including self-efficacy, outcome expectations, social support, and planning (as a proxy for goals), were measured using validated Chinese instruments. A cross-lagged SEM with single-indicator latent variables was estimated using maximum likelihood, controlling for baseline PA, gender, age, and family socioeconomic status. Model fit was evaluated using chi-square per degrees of freedom (chi2/df), comparative fit index (CFI), and standardized root mean square residual (SRMR).
RESULTS:
Results suggested a modest improvement in PA from baseline to post-intervention. Cross-lagged SEM supported key SCT pathways, indicating that baseline self-efficacy and goal-setting were positively associated with post-intervention PA after controlling for baseline PA and covariates. Outcome expectations were positively associated with PA primarily through goal-setting, whereas social support showed mainly indirect effects. Overall model fit was acceptable.
CONCLUSION:
This study provides support for SCT as a useful framework for explaining PA change within a GenAI-based intervention among healthy adolescents. These findings suggest that GenAI-based PA interventions may be more effective when explicitly targeting self-regulatory and motivational SCT constructs. The results inform theory-driven optimization of scalable AI-supported interventions to promote adolescent physical activity.