Hybrid Deep Learning Approach for Classifying Anxiety and Stress in Adolescents through Speech and Text Data

Today, adolescents are exposed to a multitude of challenges, caused both in part by academic competition, dismissal from peer pressure, social media divulgence, and the shifting picture of the family and the social surroundings that young people are familiar with. The stressors of these can turn them particularly vulnerable to psychological conditions such as anxiety and stress, and if not readily identified and treated in the long run could be severe mental long term health consequences. However, these traditional assessment methods are typically limited by subjectivity interpretation, social desirability bias, and scalability in real-life situations. To address these limitations, this study puts forward a novel hybrid deep learning framework which employs both Convolutional Neural Networks (CNN) and Bidirectional Long Short Term Memory (BiLSTM) networks for the purpose of detecting anxiety and stress levels in adolescents. Both emotional tone and linguistic patterns are captured by the system, processing multimodal inputs, from acoustic features extracted from speech and from semantic information, from transcribed text. To exploit spatial hierarchies saved in Mel-frequency cepstral coefficients (MFCCs) of speech signals, CNNs are employed. The dependencies in the textual data are modeled using BiLSTM layers. The model successfully combines these complementary representations to gain an overall view of the user’s mental state. Experimental evaluations on a labeled dataset of adolescent speech text pairs show better performance than baselines on using each modality separately. Results demonstrated that combined speech and text can be employed for reliable, automated mental health evaluation. Not only does this improve diagnostic accuracy but it also creates a real-time, scalable screening tool for early intervention and continuous mental well being monitoring in youth populations.