Implementation of IndoBERT Model in Predicting Anxiety Disorders from Comments on Social Media

Abstract

The development of social media as a means of emotional expression has opened up new opportunities in the early detection of mental health disorders, particularly anxiety disorders, which are still rarely analyzed using Indonesian-language computational approaches. To implement and evaluate IndoBERT model in detecting indications of anxiety disorders based on Indonesian-language social media comments. The method of study used an experimental quantitative approach with a total of 6,075 comments collected from Twitter, Instagram, and TikTok, which were classified into two categories: anxiety and normal. Pre-processing processes were carried out through text cleaning, slang normalization, and stopword removal before the IndoBERT model was trained using fine-tuning techniques for three epochs. Model performance was tested using accuracy, precision, recall, and F1- score metrics, and evaluated through confusion matrix analysis and k-fold cross-validation to ensure consistency of results. The results show that IndoBERT achieved 99.67% accuracy, 0.98 precision, 0.96 recall, and 0.97 F1-score in the anxiety class, with very low classification errors. This performance demonstrates that the model is able to effectively recognize linguistic patterns of anxiety despite data imbalance between classes. These findings confirm IndoBERT’s potential as a basis for developing a text-based early detection system for anxiety disorders in Indonesia. It is recommended for future studies to expand the data sources, add other psychological disorder categories, and compare performance with other algorithms to improve the model’s reliability.