Comparative Study of Signal Processing and Machine Learning Algorithms for Contactless Respiratory Rate Estimation Using FMCW Radar
Respiratory rate (RR) is a vital physiological marker whose accurate and continuous monitoring holds particular clinical significance. Conventional contact-based sensors, while reliable, impose physical constraints that may cause discomfort and skin irritation. Radar-based sensing, particularly frequency-modulated continuous-wave (FMCW) radar, offers a compelling contactless alternative: it operates through clothing, remains unaffected by lighting conditions, and enables unobtrusive continuous monitoring without compromising patient comfort or privacy. Nevertheless, radar-based RR estimation remains technically challenging due to signal noise, motion artifacts, and the inherent subtlety and variability of breathing patterns. Although prior work has demonstrated viable RR estimation using either signal processing or machine learning approaches, a direct comparison of both paradigms on pediatric radar data under a unified preprocessing framework has not been established. This study addresses that gap through a systematic and comprehensive comparative evaluation of classical signal processing and machine learning (ML) techniques for RR estimation using a publicly available FMCW radar dataset collected from 50 children under 13 years old. Six methods were evaluated: fast Fourier transform (FFT)-based frequency-domain analysis, time-domain peak detection, convolutional neural networks (CNN), multilayer perceptron (MLP), LightGBM regression, and KalmanNet. All were benchmarked against the ground truth obtained from a clinically validated reference system. The frequency-domain analysis method combined with multichannel averaging achieved the best overall performance, with a mean absolute error (MAE) of 2.60 breaths per minute (bpm), a root mean square error (RMSE) of 3.39 bpm, and an average inference time of 44.2 ms per analysis window, outperforming time-domain peak detection (MAE 3.66 bpm), CNN (MAE 2.75 bpm, RMSE 3.39 bpm), MLP (MAE 4.01 bpm, RMSE 4.64 bpm), LightGBM (MAE 5.91 bpm, RMSE 6.41 bpm), and KalmanNet (MAE 5.97 bpm, RMSE 6.73 bpm). These results indicate that, in the limited-data pediatric setting, a carefully designed classical signal processing method provides highly competitive performance, achieving accuracy statistically comparable to CNN while outperforming the remaining handcrafted feature-based learning methods, and remaining computationally efficient and interpretable.