A Two-Stage Deep Learning Framework for Enhanced Radio Frequency Sensing Reliability

This study introduces a two-stage machine learning (ML) framework for robust and accurate respiration rate estimation using radio frequency (RF) signals. The proposed system employs two custom convolutional neural networks (CNNs): the first model classifies and detects interference in the input spectrograms, while the second model estimates respiration rates from interference-free data. By leveraging transfer learning, the models are optimized for high performance. The classification model achieves an impressive accuracy of 92.0%, effectively identifying and filtering out interference-prone samples. The regression model demonstrates exceptional precision, with a Root Mean Squared Error (RMSE) of 1.53 bpm on the test data. This system underscores the potential of RF-based contactless monitoring for vital sign sensing in healthcare, especially in scenarios that require non-invasive and continuous observation. These findings highlight the significant role of advanced ML techniques in enhancing the accuracy and reliability of RF sensing for healthcare applications.