Development of Tailored Early Mental Health Interventions Leveraging Achievement-Related Traits in University Students

Abstract

Early mental health intervention is vital for university students due to increased psychological vulnerability during this transitional period, yet many underutilize available resources. This underuse is due to students’ varying stages in the Transtheoretical Model of Change (TTMC), suggesting the need for stage-appropriate intervention. Although personalization could improve effectiveness, most studies emphasize symptomatology or algorithmic accuracy over delivery methods. Key factors such as psychological traits of students, user interaction, and contextual relevance remain underexplored. To fill this gap, this study designs digital interventions tailored to students’ TTMC stages, incorporating motivational, behavioral, and self-monitoring strategies to promote mental health management (MHM).

For precontemplation stage in TTMC, a study to explore strategies to enhance motivation for mental healthcare. Group-tailored feedback was developed based on data from 174 university students, including measures of depression, anxiety, attention, perfectionism, procrastination, and sleep hygiene. This feedback was validated through interviews with 13 counselors. Students with Patient Health Questionnaire- 9 (PHQ-9) or Generalized Anxiety Disorder-7 (GAD-7) scores above 14 were categorized as a group recommended for direct professional care. The remaining students were clustered into three subgroups using k-means clustering to provide tailored feedback. Procrastination and perfectionism were discrim- inative factors in the clustering. Counselors expected the group feedback to be effective in enhancing motivation and reducing reluctance by fostering a sense of belonging. They recommended providing contextual explanations based on responses from questionnaires to increase acceptance. To generalize the feedback, a larger dataset was collected from 627 students, identifying five psychological profiles characterized by varying degrees of emotional stability, procrastination, and perfectionism. Feedback derived from these new profiles was implemented in a digital counseling program involving 46 participants. 96% reported increased interest in mental health and expressed a willingness to recommend the service to others, although only 57% of participants agreed with the suggested content. This suggests that group labeling, by satisfying a human need for belonging, can increase interest and promote openness to self-care, regardless of the accuracy of the feedback.

As procrastination was identified as a key factor in non-clinical students, the next study targeted procrastination for early intervention in the preparation-action stage of TTMC. A cognitive behavioral therapy-based semi-generative chatbot named Moa to respond interactively but maintain the content, was developed and integrated into a general to-do list application. Moa classified procrastination fac- tors with the probability matrix (PM) using a multi-layered convolutional neural network. Trait-based recommendations were generated from a PM cumulated over a week with inferred current state, while temporal recommendations used a adjusted PM without inferred context. A mixed method randomized controlled trial compared the Moa-integrated app (treatment group, n = 47) with a standard to-do app (control group, n = 38) over two months. Participants used the app three days per week in the first month, followed by an optional second month. Linear mixed model revealed a significant group×time interaction on procrastination (p = .014). Only the treatment group represented significant decreases in irrational procrastination and stress in post-hoc tests (p < .001). Piecewise linear mixed model confirmed longer sustained performance-related engagement in the treatment group (β = .003, p = .043). Thematic analysis identified the Moa’s additional role as an emotionally supportive companion, due to its empathetic and caring responses, alongside its advisory function. All participants accepted trait-based recommendation with context of prediction, whereas only 39% of users accepted temporal recommendations, due to limited awareness of the context of current state. However, the acceptance of the temporal recommendation did not affect psychological outcomes except for practical strategies. These represent relational empathy with Moa and contextual explanation before recommendation can be important in effectiveness as well as personalization precision.

Although Moa significantly reduced procrastination, app usage still declined. Users with high procrastination levels tended to disengage from self-monitoring mental health apps. To support sustained self-evaluation, the third study introduced a nudging system with a classification model optimized to reduce false alarms in mood tracking. Data collected from 241 participants over six weeks were used for model development. Model performance was assessed using integrated recall, defined as the proportion of valenced moods correctly classified or labeled as neutral. Machine learning models, such as logistic regression, were compared with a Hierarchical Bayesian Logistic Regression (HBLR) to account for both population-level trends and individual variability. Significant coefficients (p< .05) from logistic regression were used as model predictors, including prior day assessment completion, PHQ-9 and GAD-7 scores, exam periods, physical activity, and weather conditions. The HBLR model without online learning achieved the highest integrated recall for positive mood (0.977), while standard logistic regression demonstrated better that for negative mood (0.932) with lower computational cost, and was therefore recommended for real-world implementation. Assessment reminders were designed with empathetic messages based on predicted mood. The level of empathy could be adjusted using the probability from the model to allow for greater personalization. External contextualization, incorporating stressors such as exams or air quality, was necessary to mitigate delayed negative affect, as the previous day’s mood was the most effective predictor of the next day’s mood. This study proposes a novel approach to sustaining engagement by using emotionally resonant alerts, positioning predictive systems as reflective companions rather than prescriptive monitors.

Three key mechanisms emerged as critical to promoting university students in early mental health intervention: achievement-related traits such as procrastination and perfectionism should be considered for students to understand and be interested in mental health; relational empathy—built through group-tailored feedback, the supportive relationship with chatbot, and empathetic alarming—plays a decisive role in engagement as well as algorithmic precision; acceptance of one’s current state through contextual explanation must precede personalized recommendations for them to be effective. Especially, contextualizing feedback with longitudinal observations are effective for acceptance and various situational data is necessary for temporal recommendation. This work advocates a shift in digital mental health intervention design—from optimizing algorithmic precision to emphasizing achievement-related traits, emotional resonance, and contextual understanding of current status. It proposes prevention-focused framework to support mental well-being and engagement in university students.