Deep learning for sleep quality assessment: A CNN‐based approach outperforming traditional algorithms in wearable accelerometer data analysis

Abstract Objective This study evaluates and enhances wearable sleep monitoring by comparing two feature extraction methods: traditional activity counts and deep learning‐derived features. By identifying optimal machine learning architectures, we aim to improve sleep stage classification accuracy, providing a robust, noncontact tool for clinical chronic disease management. Method We employed dual feature extraction strategies: (1) traditional activity counts (triaxial acceleration signal summation and windowed averaging) and (2) a two‐tier 1D‐CNN architecture with batch normalization/ReLU/max‐pooling for hierarchical feature learning. The extracted features were subsequently modeled using four machine learning approaches (convolutional neural network [CNN], XGBoost, SVM, RF) to assess their relative performance in sleep stage classification. Results The CNN‐derived features consistently outperformed traditional features across all classifiers, achieving superior sleep quality assessment accuracy (Macro AUC: CNN 0.97 vs. XGBoost 0.98 with CNN features). Notably, the CNN model showed particular advantages in capturing transient movement patterns and cross‐channel physiological relationships critical for sleep staging. Conclusion This comparative analysis validates that although conventional activity counts maintain interpretability, CNN‐based feature extraction provides significant performance gains for wearable sleep monitoring, especially when integrated with ensemble methods. The findings establish an optimized framework for clinical sleep assessment in chronic disease management, balancing computational efficiency with diagnostic accuracy.