<![CDATA[Mental health disorders, particularly depression and anxiety, have become increasingly prevalent in Indonesia, affecting millions of individuals across all age groups. However, limited access to mental health professionals and persistent social stigma frequently prevent timely early detection and intervention. Social media platforms serve as digital diaries where Indonesian users express their emotional states, presenting a unique opportunity for automated mental health risk screening. This study proposes a deep learning approach using Bidirectional Long Short-Term Memory (BiLSTM) for classifying Indonesian social media text into three mental health risk categories: Normal (low risk), Depression (high risk), and Anxiety (moderate risk). Unlike standard LSTM, which processes text in a single direction, BiLSTM captures contextual information from both forward and backward directions — a critical advantage for understanding nuanced expressions in Indonesian informal language. A dataset of 1,488 Indonesian text samples (500 Normal, 494 Depression, 494 Anxiety) was collected from Twitter and labeled by expert annotators with an inter-annotator agreement (Cohen's Kappa) of 0.87. Comprehensive text preprocessing was applied, including case folding, noise removal, stopword elimination using NLTK, and stemming using Sastrawi. The proposed BiLSTM model was evaluated against three baseline methods: Naïve Bayes, Support Vector Machine (SVM), and standard LSTM. Experimental results demonstrate that BiLSTM achieves superior performance with 87.5% accuracy, 86.8% precision, 86.2% recall, and 86.5% F1-score, outperforming standard LSTM by 4.2% and SVM by 12.1%. A desktop application with a graphical user interface was developed for practical deployment, featuring real-time detection, confidence scoring, and prediction history logging. This research contributes an effective, reproducible, and deployable deep learning-based screening tool for mental health risk detection from Indonesian social media text.]]>
