Linguistic Feature-based Depression Prediction Using Hierarchical Transformer-attentive Model for Mental Disabilities Diagnosis

Abstract

Mental health disorders and disabilities are among the most pressing global challenges, particularly in high-stress sectors such as the technology industry. Mental health conditions can often be predicted based on linguistic and psychological features, as individuals frequently express emotional and cognitive states through textual communication. This paper presents a comprehensive review of relevant linguistic features for detecting mental health disorders, focusing on markers related to emotional state, cognition, and social interaction. Building on these insights, the study introduces hierarchical transformer-attentive depression analyzer (HiT-ADA), a novel deep learning (DL) framework that integrates convolutional neural network, recurrent neural network, long short-term memory, and transformer architectures. A key innovation in HiT-ADA is the triple-axis radius arith-metic optimizer (TARAO) mechanism, which dynamically adjusts model parameters to reduce overfitting and improve adaptability across datasets. The model was evaluated on benchmark datasets such as CLPsych and Mental Health in Tech, demonstrating superior performance over traditional machine learning and prior DL methods, with an F1-score of 0.99. By combining linguistic feature analysis with the TARAO optimizer, HiT-ADA enables more accurate and reliable prediction of mental health conditions. This high-precision early-identification system shows potential for deploy-ment in corporate environments and broader mental health screening applications.