Digital Biomarkers for Depression Monitoring Using Multisensing Wearable Wristband

Depression has become one of the most prevalent mental disorders among adolescents. However, its clinical etiology remains unclear, and unlike many other diseases, it lacks objective biochemical or medical imaging indicators for assessment and severity evaluation. To address this gap, this study presents a multi-sensing wearable wristband that integrates audio, activity, and heart rate sensors. Furthermore, this system is used to capture physiological and behavioral data for investigating how these parameters evolve in depressed adolescents during hospitalization. To identify meaningful digital biomarkers from the collected data, a multi-step analytical method has been constructed. Firstly, statistical methods such as correlation analysis, variance analysis, and non-parametric tests are employed to eliminate redundant, unrelated, and non-differential features, thereby narrowing the feature space. Moreover, a variational autoencoder is applied to address class imbalance, and focal loss is incorporated to enhance model sensitivity to difficult samples. Secondly, an XGBoost-based framework is developed to classify depression improvement levels and the model achieve a validation accuracy of 0.8. Finally, feature contribution analysis based on the proposed model is conducted using Shapley values and linear model coefficients, providing both individualized and global interpretations of differential feature sets before and after treatment. By integrating insights from these analyses, a set of objective digital biomarkers has been identified that may aid in screening and diagnosing adolescent depression.