Sleep Posture Recognition With a Dual-Frequency Microwave Doppler Radar and Machine Learning Classifiers

Abstract

An automated, robust, noncontact sleep posture recognition technique is proposed in this letter, which uses optimizable (Bayesian hyperparameter tuning) machine learning (ML) classifiers applied to dual-frequency (2.4 GHz, 5.8 GHz) monostatic continuous-wave radar-measured effective radar cross section and chest displacement. The technique is demonstrated to accurately recognize three different key sleep postures categories for 20 participants, with greater accuracy and computational efficiency than prior published research involving either a custom ML model or threshold-based assessment. Three ML classifiers (K-nearest neighbor, support vector machine (SVM), and decision tree) were assessed, with an SVM using a quadratic kernel achieving an accuracy of 85 and 80%, at 2.4 and 5.8 GHz, respectively, and the decision tree classifier recognizing sleep postures in less than 2 min with 98.4% accuracy for dual-frequency combined measurements.