Anxiety-Focused Classification Using Smartphone-Based Gait Analysis and Deep Learning Models

Abstract

Mood disorders, such as anxiety, significantly impact individual well-being and present challenges to healthcare systems globally. Traditional screening methods, including self-administered symptom questionnaires, provide valuable insights into symptoms but may miss dynamic fluctuations. This study explores the potential of smartphone inertial sensors and deep learning models for anxiety risk classification. Using gait data from 36 participants, collected via smartphone accelerometers and gyroscopes, we evaluated three models: Attention-CNN, Attention-LSTM, and a hybrid CNN-Attention-LSTM. The hybrid model achieved the highest accuracy of 87.5% in classifying at-risk for anxiety versus no-risk groups, outperforming individual models by effectively capturing spatial and temporal gait features. While the findings highlight the feasibility of smartphone-based gait analysis for anxiety risk classification, the study's limitations include a small sample size and controlled data collection settings. Future research should focus on larger, diverse datasets and real-world scenarios, with an expanded scope to other mood disorders. This approach offers scalable, non-invasive mental health monitoring and early classification.