Early Warning of Adolescent Depression: A Joint Detection Model Combining Social Media Speech Behaviour and Psychological Scales

Na Wang

PHD candidate, Faculty of Engineering, Universiti Putra Malaysia, UPM Serdang, Selangor 43400, Malaysia & PHD candidate, School of Information and Physical Sciences, The University of Newcastle, Callaghan, NSW 2308, Australia

Raja Kamil

Associate Professor, Faculty of Engineering, Universiti Putra Malaysia, UPM Serdang, Selangor 43400, Malaysia

Syed Abdul Rahman Al-Haddad

Professor, Faculty of Engineering, Universiti Putra Malaysia, UPM Serdang, Selangor 43400, Malaysia

Normala Ibrahim

Professor, Faculty of Medicine and Health Sciences/ Hospital Sultan Abdul Aziz Shah (HSAAS), Universiti Putra Malaysia, 43400 UPM

Abstract

Adolescent depression represents a mounting public health challenge, with the widespread use of social media presenting novel opportunities for the early identification of depressive symptoms through language-based analysis. This research proposes and assesses an integrated detection model designed to recognise depressive tendencies among adolescents by examining their linguistic behaviour on social media, using annotated tweet data sourced from real-world contexts. Utilising a quantitative methodology based on secondary data, three machine learning algorithms-Logistic Regression, Support Vector Machine (SVM), and DistilBERT-were trained and evaluated on a dataset comprising 10,310 tweets. Among the models, DistilBERT demonstrated superior predictive capability, achieving an F1-score of 0.9981, an accuracy rate of 99.81%, and an AUC value of 0.9975, thereby surpassing the performance of the conventional classifiers. The results affirm that patterns in adolescent language use on social media platforms exhibit detectable signs of depression, which can be effectively identified using transformer-based architectures. The study advocates for the extension of this methodological framework to encompass multilingual and temporally varied datasets to enhance its applicability across broader populations. Practical applications include the early deployment of artificial intelligence (AI)-enabled systems within educational settings and digital environments for non-intrusive mental health surveillance. One principal limitation is the absence of real-time and multi-platform validation, which warrants further investigation in subsequent studies.

Adolescent Depression, Machine Learning, Social Media Analysis, DistilBERT, Digital Mental Health