Low-cost Embedded Breathing Rate Determination Using 802.15.4z IR-UWB Hardware for Remote Healthcare

TL;DR

The paper tackles non-invasive breathing-rate monitoring for remote healthcare using a low-cost IEEE $802.15.4z$ IR-UWB radar. It proposes a CNN-based pipeline that operates on CIR-derived range-frequency maps, and situates it against rule-based baselines, showing superior accuracy in unseen environments. A compact, 8-bit quantized CNN enables deployment on a resource-constrained platform, supported by an open dataset and a practical energy model that suggests long-term operation on portable power sources. The results indicate the approach is viable for real-world deployments (e.g., bed monitoring, vehicle presence) with robust generalization and substantial battery-life gains, while highlighting future directions like multi-person scenarios and explainable AI.

Abstract

Respiratory diseases account for a significant portion of global mortality. Affordable and early detection is an effective way of addressing these ailments. To this end, a low-cost commercial off-the-shelf (COTS), IEEE 802.15.4z standard compliant impulse-radio ultra-wideband (IR-UWB) radar system is used to estimate human respiration rates. We propose a convolutional neural network (CNN) specifically adapted to predict breathing rates from ultra-wideband (UWB) channel impulse response (CIR) data, and compare its performance with both other rule-based algorithms and model-based solutions. The study uses a diverse dataset, incorporating various real-life environments to evaluate system robustness. To facilitate future research, this dataset will be released as open source. Results show that the CNN achieves a mean absolute error (MAE) of 1.73 breaths per minute (BPM) in unseen situations, significantly outperforming rule-based methods (3.40 BPM). By incorporating calibration data from other individuals in the unseen situations, the error is further reduced to 0.84 BPM. In addition, this work evaluates the feasibility of running the pipeline on a low-cost embedded device. Applying 8-bit quantization to both the weights and input/ouput tensors, reduces memory requirements by 67% and inference time by 64% with only a 3% increase in MAE. As a result, we show it is feasible to deploy the algorithm on an nRF52840 system-on-chip (SoC) requiring only 46 KB of memory and operating with an inference time of only 192 ms. Once deployed, an analytical energy model estimates that the system, while continuously monitoring the room, can operate for up to 268 days without recharging when powered by a 20 000 mAh battery pack. For breathing monitoring in bed, the sampling rate can be lowered, extending battery life to 313 days, making the solution highly efficient for real-world, low-cost deployments.

Figures (6)

• Figure 1: Samples were collected across various environments with 16 participants. Environment 1 had subjects lying down. Environments 2 and 3 had them sitting. In the final 2 environments, subjects stood at multiple distances and angles from the radar. Defining a setup as a distinct combination of environment, distance, angle, and posture resulted in 15 different setups.
• Figure 2: Visualization of processing pipeline. iq samples coming from the receiver will serve as input for the pipeline. Each sub-representation of the data is described in Section \ref{['sec:methodology']} and is indicated with a letter at the bottom of the subfigure.
• Figure 3: The breathing rate estimates of the rule-based algorithms and the proposed CNN are visualized on a range-frequency map. In a simple scenario with a visible peak, shown in (a), the predictions from all methods closely align. However, in noisier scenarios where the peak is less distinct, as shown in (b), the cnn predictions are significantly closer to the ground truth.
• Figure 4: Breathing rate prediction errors in different situations (environment, posture, distance and body angle). The green, yellow and purple plots are the results when applying the rule-based methods described in \ref{['sec:peak_finding']}. The red plots correspond to the L1 errors of the cnn for fully unseen situations. The blue box plots correspond to the L1 errors of the cnn for predicting the breathing rate for a new unseen person when other persons have already been observed in the specific situation (e.g. some form of pretraining on other individuals was possible). The cnn achieves a mae of 1.73 bpm in unseen situations, significantly outperforming rule-based methods (3.40 BPM). When pretraining is possible, the mae of the cnn is further reduced to 0.84 BPM.
• Figure 5: Median L1 error for different sampling rates. Lowering the sampling rate negatively affects L1 error. For easy situations (situation 1), the sampling rate can be lowered to 4 Hz. When also considering more difficult situations, sampling rate should remain above 20 Hz.