Overcoming Data Scarcity in Non-Contact Respiratory Monitoring: A DTW-Enhanced CNN-LSTM Approach

This study investigates non-contact respiratory pattern classification using Ultra-Wideband (UWB) radar sensors and deep learning. A CNN-LSTM hybrid architecture was developed combining spatial feature extraction through convolutional layers with temporal pattern recognition via LSTM networks. To address data scarcity in the minority class, a two-stage augmentation strategy incorporating Dynamic Time Warping-based SMOTE-TS was implemented. The experimental evaluation utilized 700 respiratory recordings from seven healthy volunteers performing controlled breathing exercises. Under controlled laboratory conditions, the system achieved 94.3% accuracy and 0.969 AUC, with an average inference time of 45.3 ms per sample (SD: 8.7 ms), demonstrating computational feasibility for real-time applications. This preliminary investigation establishes technical proof-of-concept, though validation with clinical populations remains necessary before medical deployment.