Message Recommendation Strategies for Tailoring Health Information to Promote Physical Activities

In many behaviour change interventions, computer-tailored health information has proven to be more effective than general health information. However, the majority of these studies have only achieved small effect sizes and the effectiveness of computer-tailored health communication (CTHC) remains inconsistent across different populations and behaviours. Since most CTHC studies measure a behaviour difference (e.g., steps per day) or biological difference (e.g., blood pressure), it is challenging to determine whether the intervention's success is due to the quality of message tailoring or other factors (e.g., user interface design). This paper presents a study that assesses the performance of various algorithms for tailoring health information. These algorithms include a rule-based approach, based on behaviour change theories and machine learning algorithms. Despite limited data, the evaluated algorithms significantly outperform random message selection, achieving a 1.7-fold increase in precision for predicting participants' preferred messages, and a 1.38-fold improvement in overall accuracy for anticipating participants' preferences.