Using Machine Learning to Predict Young People’s Internet Health and Social Service Information Seeking

Machine learning creates new opportunities to design digital health interventions for youth at risk for acquiring HIV (YARH), capitalizing on YARH’s health information seeking on the internet. To date, researchers have focused on descriptive analyses that associate individual factors with health-seeking behaviors, without estimating of the strength of these predictive models. We developed predictive models by applying machine learning methods (i.e., elastic net and lasso regression models) to YARH’s self-reports of internet use. The YARH were aged 14–24 years old (N = 1287) from Los Angeles and New Orleans. Models were fit to three binary indicators of YARH’s lifetime internet searches for general health, sexual and reproductive health (SRH), and social service information. YARH responses regarding internet health information seeking were fed into machine learning models with potential predictor variables based on findings from previous research, including sociodemographic characteristics, sexual and gender minority identity, healthcare access and engagement, sexual behavior, substance use, and mental health. About half of the YARH reported seeking general health and SRH information and 26% sought social service information. Areas under the ROC curve (≥ .75) indicated strong predictive models and results were consistent with the existing literature. For example, higher education and sexual minority identification was associated with seeking general health, SRH, and social service information. New findings also emerged. Cisgender identity versus transgender and non-binary identities was associated with lower odds of general health, SRH, and social service information seeking. Experiencing intimate partner violence was associated with higher odds of seeking general health, SRH, and social service information. Findings demonstrate the ability to develop predictive models to inform targeted health information dissemination strategies but underscore the need to better understand health disparities that can be operationalized as predictors in machine learning algorithms.