Transformer-Based Heartbeat Monitoring with FMCW Radar Under Random Body Motion
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: We're looking at the paper: "Transformer-Based Heartbeat Monitoring with FMCW Radar Under Random Body Motion".
Dev: Here's the gist of the paper. Millimeter-wave Frequency Modulated Continuous Wave (FMCW) radar enables contactless cardiac monitoring, but heartbeat estimation becomes challenging when respiration and random body motion (RBM) distort the radar signal <ref:2605.28432#pg2> This paper proposes a hybrid framework for 77 GHz FMCW radar that combines model-based signal processing with a Convolutional Neural Network (CNN)–Transformer network <ref:2605.28432#pg2>.
Rosa: The proposed method consists of two main components: a model-based signal processing block and a CNN–Transformer network <ref:2605.28432#pg2> The first component processes raw mmWave FMCW radar measurements into physiologically meaningful features (time series) <ref:2605.28432#pg2>.
Dev: The second component, the CNN–Transformer network, processes these extracted features to reconstruct the photoplethysmography (PPG) signal <ref:2605.28432#pg2>.
Rosa: Signal Processing Pipeline The model-based block involves several steps to convert raw radar data into features <ref:2605.28432#pg2> 1.
Title and authors: Rosa: We’re diving into "Transformer-Based Heartbeat Monitoring with FMCW Radar Under Random Body Motion" and the folks who wrote it, Matteo Pardi, Amir Hosein Oveis, Saba Kharabadze, Ajeet Kumar. This paper tackles that core issue of heart monitoring using millimeter wave radar when things get physically messy.
Dev: It’s interesting how they framed the problem right up front by focusing on the challenge: when respiration and random body motion distort the radar signal, it makes heartbeat estimation really difficult.
Taro: They propose a hybrid framework, combining model-based signal processing with a CNN Transformer network to handle that distortion.
Rosa: So, instead of just throwing raw data at a deep learning model, they first use the model-based block to extract features like chest displacement time series from the radar measurements.
Dev: That feature extraction step is crucial because it’s what converts the raw physical wave data into something that has some physiological meaning for the next stage.
Taro: They detail how this involves things like range processing and then estimating chest velocity using wrapped phase differences, which is key to getting that displacement estimate.
Rosa: And they spend time on refining those velocity samples because they recognize that if you just use raw measurements, it can lead to errors when the range bins migrate.
Dev: That interpolation step is what allows the system to handle the inherent noise from random body motion better by replacing unreliable samples with estimates from reliable neighbors.
Taro: After getting that displacement estimate, they construct eight different feature channels, including amplitude and a coarse random body motion indicator, which feeds into the neural network.
Rosa: Then this whole set of features goes into the CNN Transformer network to reconstruct a photoplethysmography signal that mimics what we see in a real heartbeat.
Dev: So it’s not just one black box; it’s a structured pipeline where you have these physical estimations feeding into a powerful sequence model for reconstruction.
Taro: And the architecture involves a convolutional stem to map those features to some latent embedding, followed by positional encoding, then the Transformer encoder handling the temporal relationships.
Rosa: It sounds like they are trying to give the network both a structural understanding of the data and an understanding of how things change over time.
Dev: That’s what I mean. Giving it context on where those features came from before letting it figure out the complex temporal patterns is a smart design choice for this problem.
Taro: The architecture shows they are modeling both the static structure of the signal and its dynamic evolution, which is important when you're dealing with breathing and motion.
Rosa: And they used synchronized photoplethysmography signals as their ground truth during supervised training to guide that reconstruction process effectively.
Dev: That ground truth synchronization is a big piece of information because it gives the AI a clear target to aim for during the learning phase.
The paper's summary: Taro: So, what they actually say in their summary is that this hybrid approach successfully combines the model-based processing with the CNN Transformer network to solve the problem of heart rate estimation under random body motion.
Rosa: Essentially, they argue that separating the tasks—feature extraction and signal reconstruction—allows each component to focus on its specific job better than a single end-to-end model could.
Dev: They’re trying to show that this separation leads to more physiologically meaningful features derived from the raw radar data before feeding them into the AI for pattern recognition.
Taro: The paper summarizes that they are extracting chest displacement and constructing high-level motion features from the raw radar data in the first block, while then using these features to reconstruct a PPG-like signal in the second block.
Rosa: They’re emphasizing that this two-block structure is what makes the entire method capable of handling both breathing and random movement effectively.
Dev: They highlight how this method is designed to be robust by explicitly addressing motion corruption through the feature engineering part before letting the deep learning component do its heavy lifting.
Taro: So, when you read the summary, it’s clear they are focusing on creating features that are less susceptible to noise introduced by those environmental factors before they get processed by the network.
Rosa: They’re essentially saying that this structured approach is superior because it forces a physical interpretation of the data before letting the AI get involved.
Dev: And this structural approach is what allows them to achieve results across all three test conditions: stationary, deep breathing, and random body motion.
The paper's improvements: Rosa: They suggest improvements mainly in how they refine those model-based features to better separate the motion components from the cardiac signals when looking at Figure four <ref:2605.28432#pg2>.
Dev: They are focusing on refining that feature construction block because they think making those handcrafted features more physically meaningful will make them less susceptible to noise from random body motion.
Taro: Specifically, they want to improve how the features are built so that they can better isolate the heart signal from the movement artifacts in a way that’s more robust.
Rosa: So, if you look at Figure four they’re suggesting tweaks there to make sure those features are cleaner inputs for the deep learning block <ref:2605.28432#pg2>.
Dev: That means refining the feature engineering part so that when we move through the world, the resulting features stay closer to what a heartbeat actually is.
Taro: The goal of these improvements is to ensure that whatever motion profile you encounter, the system recovers gracefully instead of completely breaking down its estimation capabilities.
Rosa: So they are looking at making those handcrafted features more robust inputs for the deep learning component when dealing with real-world movement.
Dev: And from a practical standpoint, if those features are cleaner, it means we might be able to use lighter models in the second block while still getting high accuracy.
Conclusion: Rosa: So, wrapping up the paper "Transformer-Based Heartbeat Monitoring with FMCW Radar Under Random Body Motion," the main conclusion is that their hybrid framework reliably reconstructs the PPG signal across all tested scenarios.
Dev: They’ve shown that this method outperforms existing techniques for metrics like average heart rate and heart rate variability when conditions are nonideal or affected by unwanted motion.
Taro: They demonstrated that it still follows the overall trend of the true photoplethysmography and correctly detects main peaks even in the random body motion scenario, which is a strong result.
Rosa: The implication is that this provides a reliable tool for radar-based vital sign monitoring in environments where you can't guarantee perfect stillness.
Dev: It’s practical because it moves us closer to having non-contact cardiac monitoring that can actually be used in more realistic, dynamic settings than just a lab bench.
Taro: The future work they mentioned is testing other neural networks to see if those variations can improve the performance and also reduce the computational cost of this entire system.
Rosa: It’s a solid result showing that the proposed architecture works well under motion-corrupted conditions, confirming its practical value for tracking vital signs in realistic environments.
Department of Information Engineering, University of Pisa · Radar and Surveillance Systems (RaSS) National Laboratory, CNIT
eess.SP
Submitted: 2026-05-27
Updated: 2026-05-27
Comments: 6 pages, 5 figures. Accepted for publication in the Proceedings of the 2026 IEEE Radar Conference (RadarConf26)
Journal ref: 2026 IEEE Radar Conference (RadarConf'26)
DOI: 10.1109/RadarConf2663773.2026.11674955
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 83/100
The gist: Millimeter-wave Frequency Modulated Continuous Wave (FMCW) radar enables contactless cardiac monitoring…
Key concepts
- FMCW Radar
- Frequency Modulated Continuous Wave radar uses radio waves with a continuously changing frequency to measure distance and velocity. In this study, 77 GHz FMCW radar is used to detect chest movements without physical contact, making it suitable for contactless heart monitoring.
- Model-Based Signal Processing
- This initial step converts raw radar measurements into meaningful features. It involves processing the signal to remove static clutter, identify the chest reflection, and estimate key physiological parameters like amplitude and velocity. This creates a set of physically relevant data points for the neural network.
- CNN-Transformer Network
- This deep learning component takes the extracted features from radar as input. The CNN part processes these features to learn spatial patterns, while the Transformer part models complex temporal dependencies to reconstruct the original PPG signal, which represents heartbeat activity.
Terminology
Summary
Millimeter-wave Frequency Modulated Continuous Wave (FMCW) radar enables contactless cardiac monitoring, but heartbeat estimation becomes challenging when respiration and random body motion (RBM) distort the radar signal <ref:2605. p. 028432#pg2> This paper proposes a hybrid framework for 77 GHz FMCW radar that combines model-based signal processing with a Convolutional Neural Network (CNN)–Transformer network <ref:2605. p. 028432#pg2>.
Methodology Overview
The proposed method consists of two main components: a model-based signal processing block and a CNN–Transformer network <ref:2605.28432#pg2> The first component processes raw mmWave FMCW radar measurements into physiologically meaningful features (time series) <ref:2605. p. 128432#pg2>. The second component, the CNN–Transformer network, processes these extracted features to reconstruct the photoplethysmography (PPG) signal <ref:2605. p. 128432#pg2>.
Signal Processing Pipeline
The model-based block involves several steps to convert raw radar data into features <ref:2605.28432#pg2>
-
Range Processing and Clutter Removal: This step involves multiplying each CPI by a Hamming window and applying a one-dimensional FFT, followed by subtracting the mean value across the slow-time dimension for each range bin to suppress static clutter <ref:2605. p. 128432#pg2>.
-
Chest Range Bin Selection: The dominant chest reflection is identified by selecting the range bin corresponding to the maximum signal power for each slow-time frame <ref:2605. p. 128432#pg2>.
-
Phase, Amplitude, and Velocity Estimation: From the extracted chest IQ signal, instantaneous amplitude A[t] and wrapped phase are computed <ref:2605. p. 128432#pg2>. Phase differences between consecutive frames are calculated to convert them into instantaneous radial velocity v[t] <ref:2605. p. 128432#pg2>.
-
Velocity Refinement: Velocity samples affected by range bin migration are replaced using linear interpolation over neighboring reliable samples <ref:2605. p. 128432#pg2>.
-
Displacement Reconstruction: The corrected velocity vˆ[t] is integrated over slow time to obtain the chest displacement Rˆ[t] <ref:2605. p. 128432#pg2>.
-
Feature Construction: Eight complementary feature channels are constructed, including Chest displacement Rˆ[t], Chest signal amplitude A[t] in dB, and a Coarse RBM activity indicator <ref:2605. p. 128432#pg2>.
Deep Learning Architecture
The feature tensor F ∈ R C×T is fed into a Transformer-based network to estimate the PPG signal yˆ ∈ R T <ref:2605.28432#pg2>. The model architecture includes four blocks:
a) Convolutional Stem: This processes the input tensor F using a one-dimensional ReLU convolutional stem to map the feature space to a latent embedding Z ∈ R dmodel×T <ref:2605.28432#pg2>.
b) Positional Encoding: The embedded features Z are augmented with standard additive sinusoidal positional encoding to yield Z′ ∈ R dmodel×T <ref:2605. p. 228432#pg2>.
c) Transformer Encoder: Temporal dependencies are modeled using a Transformer encoder composed of Nlayers encoder layers, each employing multi-head self-attention <ref:2605. p. 228432#pg2>.
d) Output Head: The Transformer output is mapped to a scalar waveform using a single 1×1 Conv1D to produce the estimated PPG signal yˆ ∈ R T <ref:2605. p. 228432#pg2>.
Experimental Validation and Results
The method was evaluated following the IEEE AESS Challenge Problem I protocol across three motion scenarios: stationary, deep breathing, and random body motion (RBM) <ref:2605.28432#pg2>. The results demonstrate that the proposed architecture reliably reconstructs the PPG signal in all scenarios <ref:2605. p. 028432#pg2>. Specifically, it achieves high fidelity in controlled conditions and maintains robust performance under motion <ref:2605. p. 028432#pg2>. In the RBM scenario, while reconstruction quality has some loss, the model still follows the overall trend of the true PPG and correctly detects main peaks <ref:2605. p. 428432#pg2>. The overall score achieved by our method was 269. p. 40 <ref:2605. p. 0 p. 028432#pg2>.
Performance Comparison
The proposed framework clearly outperforms benchmark methods when conditions are nonideal or affected by unwanted motion <ref:2605.28432#pg2>. In the RBM scenario, benchmark methods like Filtering, Blind Source Separation, and HFHeartbeat fail completely under motion-corrupted conditions <ref:2605. p. 528432#pg2>. The proposed method provides reliable estimates of both Average Heart Rate (AHR) and Heart Rate Variability (HRV) metrics <ref:2605. p. 028432#pg2>. The final score comparison shows our Method achieving a Total Score of 269. p. 00 against the benchmark scores <ref:2605. p. 528432#pg2>.
Conclusion
The results confirm that the proposed architecture reliably reconstructs the PPG signal in all tested scenarios, maintaining robustness under motion-corrupted settings <ref:2605.28432#pg2>. The method clearly outperforms existing techniques for metrics such as AHR and HRV <ref:2605. p. 528432#pg2>. This confirms the practical value of the proposed architecture for radar-based vital sign monitoring in realistic environments <ref:2605. p. 5 p. 528432#pg2>. Future work will focus on testing other neural networks to improve performance and computational cost <ref:2605. p. 5 p. 528432#pg2>.
References
[1] European Telecommunications Standards Institute, “Short Range Devices (SRD); Automotive Short Range Radar operating in the 77 GHz to 81 GHz band; Harmonised Standard,” ETSI EN 302 264 v1.1.1, 2008 <ref:2605. p. 028432#pg2>.
[2] Federal Communications Commission, “Amendment of the Commission’s Rules Regarding the 76–81 GHz Band,” FCC Report and Order, ET Docket No. p. 5 15-26, 2017 <ref:2605. p. 028432#pg2>.
[3] Yuanyuan Zhang et al. p. 5, “An Overview of Algorithms for Contactless Cardiac Feature Extraction from Radar Signals: Advances and Challenges,” IEEE Transactions on Instrumentation and Measurement, vol. p. 5 72, pp. 1–20, 2023, arXiv:2508. p. 502122 [eess] <ref:2605. p. 028432#pg2>.
[4] Yu Rong, Isabella Lenz, and Daniel W. p. 5 Bliss, “Noncontact Cardiac Parameters Estimation Using Radar Acoustics for Healthcare IoT,” IEEE Internet of Things Journal, vol. p. 5 11, no. 5, pp. 7630–7639, Mar. p. 5 2024 <ref:2605. p. 028432#pg2>.
[5] Yu Rong, Kawon Han, and Daniel W Bliss, “Motion-robust non-contact heartbeat sensing using radar acoustics for healthcare iot,” IEEE Internet of Things Journal, pp. p. 5 1–1, 2026 <ref:2605. p. 028432#pg2>.
[6] Jianxuan Tu and Jenshan Lin, “Fast Acquisition of Heart Rate in Noncontact Vital Sign Radar Measurement Using Time-Window-Variation Technique,” IEEE Transactions on Instrumentation and Measurement, vol. p. 5 65, no. 1, pp. 112–122, Jan. p. 5 2016 <ref:2605. p. 028432#pg2>.
[7] Adeel Ahmad et al. p. 5, “Vital signs monitoring of multiple people using a FMCW millimeter-wave sensor,” in 2018 IEEE Radar Conference (RadarConf18), Oklahoma City, OK, Apr. p. 5 2018, pp. 1450–1455, IEEE <ref:2605. p. 528432#pg2>.
[8] Toan K Vo Dai et al. p. 5, “Enhancement of Remote Vital Sign Monitoring Detection Accuracy Using Multiple-Input Multiple-Output 77 GHz FMCW Radar,” IEEE Journal of Electromagnetics, RF and Microwaves in Medicine and Biology, vol. p. 5 6, no. 1, pp. 111–122, Mar. p. 5 2022 <ref:2605. p. 028432#pg2>.
[9] Weijie Xia, Yi Li, and Shiqi Dong, “Radar-Based High-Accuracy Cardiac Activity Sensing,” IEEE Transactions on Instrumentation and Measurement, vol. p. 5 70, pp. 1–13, 2021 <ref:2605. p. 028432#pg2>.
[10] Zhanjun Hao et al. p. 5, “Fmcw-based contactless heart rate monitoring,” Scientific Reports, vol. p. 5 15, no. 1, pp. 2616, 2025 <ref:2605. p. 528432#pg2>.
[11] Chaoyan Zhang et al. p. 5, “A radar vital signs detection method in complex environments,” Scientific Reports, vol. p. 5 16, no. 1, pp. 2333, 2026 <ref:2605. p. 028432#pg2>.
[12] Justin Saluja, Joaquin Casanova, and Jenshan Lin, “A Supervised Machine Learning Algorithm for Heart-Rate Detection Using Doppler Motion-Sensing Radar,” IEEE Journal of Electromagnetics, RF and Microwaves in Medicine and Biology, vol. p. 5 4, no. 1, pp. 45–51, Mar. p. 5 2020 <ref:2605. p. 028432#pg2>.
[13] Sha Yuan et al. p. 5, “Heart Rate Variability Monitoring Based on Doppler Radar Using Deep Learning,” Sensors, vol. p. 5 24, no. 7, pp. 2026, Jan. 2024, Publisher: Multidisciplinary Digital Publishing Institute <ref:2605. p. 528432#pg2>.
[14] Kilin Shi et al. p. 5, “Segmentation of Radar-Recorded Heart Sound Signals Using Bidirectional LSTM Networks,” in 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Berlin, Germany, July 2019, pp. p. 5 6677–6680, IEEE <ref:2605. p. 528432#pg2>.
[15] Yuh-Shyan Chen, Jhuo-Han Lyu, and Chih-Shun Hsu, “Multi-Feature Transformer-Based Learning for Remote and Accurate Vital Sign Estimation Using mmWave FMCW Radar,” in 2025 IEEE Wireless Communications and Networking Conference (WCNC), Milan, Italy, Mar. p. 5 2025, pp. 1–6, IEEE <ref:2605. p. 028432#pg2>.
[16] Yu Rong et al. p. 5, “Radar-based heartbeat monitoring in dynamic scenarios,” IEEE Aerospace and Electronic Systems Society (AESS) Challenge Problem I, 2025, Supported by IEEE AESS and 2025 IEEE International Radar Conference <ref:2605. p. 528432#pg2>.
[17] Richard Liaw et al. p. 5, “Tune: A research platform for distributed model selection and training,” arXiv preprint arXiv:1807. p. 505118, 2018 <ref:2605. p. 528432#pg2>.
[18] Liam Li et al. p. 5, “A system for massively parallel hyperparameter tuning,” in Proceedings of the Third Conference on Machine Learning and Systems, MLSys 2020, Austin, TX, USA, March 2-4, 2020 <ref:2605. p. 528432#pg2>.
[19] James Bergstra, Daniel Yamins, and David Cox, “Making a science of model search: Hyperparameter optimization in hundreds of dimensions for vision architectures,” in International conference on machine learning. p. 5 PMLR, 2013, pp. p. 5 115–123 <ref:2605. p. 028432#pg2>.
[20] The paper is organized as follows. p. 1 Section II presents the proposed methodology. p. 1 Section III describes the dataset and the experimental setup <ref:2605. p. 128432#pg2>.
Data Description and Setup
The experimental dataset was provided by the IEEE AESS Challenge Problem I [16] <ref:2605.28432#pg2>. Data were collected using a 77 GHz mmWave FMCW radar system based on the Texas Instruments platform (AWR2243BOOST or AWR1843BOOST) with a DCA1000 capture card <ref:2605. p. 328432#pg2>. The system employs three transmit and four receive antennas, forming a 12-element virtual array via TDM-MIMO <ref:2605. p. 328432#pg2>. A bandwidth of 3. p. 36 GHz provides a range resolution of approximately 4. p. 32 cm <ref:2605. p. 028432#pg2>. Approximately 45 minutes of data were collected under three conditions: stationary, deep breathing, and random body movement <ref:2605. p. 328432#pg2>. Ground-truth cardiac activity was provided by a synchronized PPG signal acquired using a pulse oximeter and stored within each data file <ref:2605. p. 328432#pg2>.
Performance Metrics
Performance is evaluated according to the IEEE AESS Challenge Problem I [16] protocol, which assesses Average Heart Rate (AHR) and Heart Rate Variability (HRV) <ref:2605.28432#pg2>. The AHR error is quantified using the root-mean-square error (RMSE) between the estimated and reference heart rates, with the figure of merit defined as FoMAHR = 1 / TCPI · ˆδAHR <ref:2605. p. 328432#pg2>. HRV performance is evaluated in the time domain using beat-to-beat cardiac intervals (RR intervals) extracted from the estimated and reference PPG signals <ref:2605. p. 428432#pg2>. The primary error metric for HRV is the RMSE between the estimated and reference heart rate sequences, along with Pearson correlation coefficient Ch <ref:2605. p. 428432#pg2>.
Final Model Configuration
The best-performing configuration utilized AdamW (betas=(0.9, 0.999), eps=1e-08) with learning rate 9. p. 42 × 10−4, embedding dimension dmodel = 96, and 8 attention heads <ref:2605. p. 428432#pg2>. This model was trained for the duration of 40 epochs <ref:2605. p. 428432#pg2>. The final model achieved a Total Score of 269. p. 40 <ref:2605.28432#pg2>.
The results show that the proposed architecture reliably reconstructs the PPG signal in all tested scenarios, achieving high fidelity in controlled conditions and maintaining robust performance under motion-corrupted settings <ref:2605.28432#pg2>. The proposed method clearly outperforms the IEEE AESS Challenge Problem I benchmark techniques <ref:2605. p. 528432#pg2>. This confirms the practical value of the proposed architecture for radar-based vital sign monitoring in realistic environments <ref:2605.28432#pg2>. Future work will focus on testing other neural networks to improve performance and computational cost <ref:2605.28432#pg2>.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Model-based signal processing block improvement involves refining feature construction to better separate motion components from cardiac signals, as described in Section II A and Figure 4. This could lead to more physically meaningful handcrafted features that are less susceptible to noise introduced by random body motion (RBM).
-
CNN-Transformer network enhancement includes exploring other neural networks for a balance between performance and computational cost, aiming for lightweight and real-time implementations as suggested in Section V. The improved system can perform heartbeat monitoring with lower latency suitable for wearable devices.
-
Integration of multi-feature Transformer models could be explored to improve the estimation of PPG signal from the feature tensor F, as mentioned in Section II B. This would allow the system to better model complex long-range dependencies in the extracted features, potentially enhancing robustness against varying motion profiles.
Sources
- An Overview of Algorithms for Contactless Cardiac Feature Extraction from Radar Signals: Advances and Challenges
- Tune: A Research Platform for Distributed Model Selection and Training
Related papers
- Open-H-RF: A Large-Scale Ultrasound Raw Channel Dataset Initiative
- Scalable Attention for 5G NR Channel Estimation
- Low-Altitude Wireless Networks: A Comprehensive Survey
- Sensing-Assisted OFDM with Index Modulation: A Waveform for Optical Wireless Integrated Communication and Localisation
- Iteration-Free Candidate Pruning and an Index-Domain VLSI Architecture for Sort-Free MIMO Box Decoding up to 1024-QAM
- A Foundation Model for Large-Scale Wireless Network Planning, Operation and Optimization