DNN 기반 우울증 위험 예측 성능 향상을 위한 하이퍼파라미터 최적화 방법 연구

초록

This study proposes a method to improve the predictive performance of a DNN‑based depression‑risk model by optimizing its hyperparameters using biometric data from wearable devices and PHQ‑9 survey responses. After rigorous preprocessing, including quality enhancement and class-imbalance correction with SMOTE, we compared several traditional machine-learning algorithms with a deep neural network (DNN). The DNN, optimized with MinMax scaling, six hidden layers, and the Nadam optimizer, achieved the best performance, recording an F1-score of 0.8953 and demonstrating superior ability to capture the non-linear and hierarchical patterns intrinsic to wearable biosignals. These findings provide empirical support for developing wearable-based systems for early depression screening and continuous monitoring.

키워드