Christian Kent Cabilan Abay-Abay, Thristan Jay Escasinas Nakila, Christine Flores Peña
Mental health disorders, such as depression and anxiety, constitute significant global public health challenges, yet resource-intensive clinical diagnostics impede scalable early detection. This study develops a hybrid Natural Language Processing (NLP) model to classify seven mental health states, Normal, Anxiety, Depression, Suicidal, Stress, Bipolar, and Personality Disorder, from user-generated textual data. The model integrates VADER for rule-based sentiment scoring [1], BERT for deep contextual embeddings [2], BiLSTM for sequential text modeling [3], and CRF for structured label prediction [4] to accurately identify explicit and nuanced emotional cues [5] [6]. Using a dataset of 51,093 text entries [7], the model was evaluated through 10-fold stratified cross-validation with SMOTE applied to address class imbalance [8]. The hybrid model achieved an overall accuracy of 0.81, macro F1-score of 0.78, and weighted F1-score of 0.81. Performance was strongest for Normal (F1=0.94), Anxiety (F1=0.85), and Bipolar (F1=0.84) classes, while comparatively lower for underrepresented classes such as Personality Disorder (F1=0.68) and Stress (F1=0.67). The results demonstrate that combining lexical, contextual, and sequential components yields a robust framework for mental health text classification, offering a scalable approach to support early detection and public health interventions [9] [10]. © 2026 Copyright held by the owner/author(s)
Department of Computer, Information, Science, and Mathematics, University of San Carlos Cebu, Cebu, Philippines
