Radar-APLANC: Unsupervised Radar-based Heartbeat Sensing via Augmented Pseudo-Label and Noise Contrast

AI summary

Overview

Research area: Contactless physiological sensing — specifically radar-based heartbeat monitoring using Frequency Modulated Continuous Wave (FMCW) radar, framed here as an unsupervised deep learning problem (submitted to arXiv under cs.CV).

Technical level: Advanced. The paper assumes familiarity with radar signal processing (chirps, IF signals, range matrices, phase unwrapping), contrastive/self-supervised learning, and heart-rate regression metrics.

One-sentence scope: The paper proposes Radar-APLANC, the first unsupervised framework for radar-based heartbeat sensing, which uses a Noise-Contrastive Triplet loss plus an augmented pseudo-label generator to match supervised performance without any ground-truth physiological labels.

What This Paper Is About

Radar can measure the sub-millimeter chest wall displacements caused by a heartbeat (typically 0.1–0.5 mm), enabling contact-free heart-rate monitoring. Traditional radar methods simply extract and unwrap the phase of the received signal and degrade badly under motion artifacts, multipath interference and low signal-to-noise conditions, while newer deep learning methods are noise-robust but need costly synchronized ground-truth signals such as PPG. The goal of this work is to train a radar heartbeat extractor with no physiological labels at all, by treating the traditional radar method's output as a weak pseudo-label and explicitly exploiting the noise that contaminates radar data.

Key Contributions

1. First unsupervised framework for radar-based heartbeat sensing. Radar-APLANC removes the dependency on physiological labels while reaching performance comparable to supervised methods.
2. A Noise-Contrastive Triplet (NCT) loss. Described as the first attempt to exploit noise artifacts in the radar range matrix, it pulls the predicted heartbeat spectrum toward pseudo-labels and pushes it away from spectra extracted from non-heartbeat range bins.
3. A two-stage training strategy with an augmented pseudo-label generator that performs quality assessment and adaptive noise-aware label selection directly from radar range matrices.
4. A new benchmark dataset (RHB) collected from 80 subjects, which the authors state will be open-sourced; code and supplementary materials are available at https://github.com/RadarHRSensing/Radar-APLANC.

Main Findings

• Intra-dataset performance on Equipleth: Radar-APLANC reaches MAE 3.95, RMSE 9.72, r 0.64. Supervised methods still lead — Equipleth RF is first (MAE 2.18, RMSE 6.12, r 0.89) and VitaNet is second (MAE 3.14, RMSE 7.70, r 0.77) — but the unsupervised method is only 25.8% higher in MAE than VitaNet.
• Intra-dataset performance on RHB: Radar-APLANC reaches MAE 3.92, RMSE 7.94, r 0.77, described as near state-of-the-art and only 22.9% above the Equipleth RF method. The paper notes supervised models degrade significantly on RHB while the proposed method stays stable across the two datasets.
• Cross-dataset generalization: Trained on RHB and tested on Equipleth (RHB→Equipleth), the method shows only a +3.8% MAE increase (3.95 to 4.10, RMSE 8.51, r 0.72), versus +107.8% for the Equipleth Radar supervised method and +136.6% for VitaNet. Under Equipleth→RHB the method achieves MAE 3.52, RMSE 7.45, r 0.79, with MAE fluctuation across intra- and cross-dataset evaluations within 0.4 bpm.
• Traditional methods are far behind: FFT-based RF (Alizadeh et al. 2019) scores MAE 13.51, RMSE 21.07, r 0.24 on Equipleth, and the unsupervised method "significantly outperforms" both traditional training-free radar methods in the comparison.
• Fairness versus RGB sensing: Because radar is independent of skin tone and lighting, radar methods generally show better fairness than RGB methods. Radar-APLANC reports an MAE fairness gap of 0.91, RMSE gap 0.98, and r gap -0.06, comparable to other radar methods (e.g., VitaNet 0.30 MAE gap, Equipleth RF 0.51) and better than RGB methods such as PhysNet (2.22 MAE gap) and FusionPhys-RGB (1.23 MAE gap). Fairness is defined as the performance difference between dark and light skin tones, with values near zero indicating high fairness.
• Noise contrast is essential, and insufficient alone: In the stage-one ablation, using only the noise matrix fails to converge (MAE 34.48, RMSE 38.34, r 0.01), while pseudo-labels alone give MAE 8.94 (already well below the traditional FFT-based RF result of 13.51). Combining noise matrix and pseudo-labels drops MAE to 4.40, less than half of 8.94.
• Stage two always helps when fully enabled: The configuration with pseudo-labels and stage-two augmented pseudo-labels but no stage-one noise matrix yields MAE 7.42; enabling all components in both stages gives the best result, MAE 3.95, RMSE 9.72, r 0.64.
• The augmented pseudo-label generator depends on using all three signal sources: Using only the pretrained heartbeat signal as pseudo-labels gives MAE 4.56 (RMSE 10.09, r 0.63), which falls short of the stage-one result of 4.40. Combining traditional heartbeat signals with either the pretrained heartbeat signal (MAE 8.75) or the noise signal (MAE 14.48) performs poorly; only the combination of all three achieves the best MAE of 3.95.
• Qualitative behavior: Visualization shows stage one already produces heartbeat signals closely aligned with ground truth with minor deviations, and stage two reduces residual fluctuations for improved waveform accuracy.

Methodology in Plain English

Setting up the signal. An FMCW radar transmits a series of chirps and receives reflections. Mixing transmitted and received signals produces a complex intermediate frequency signal, and applying an FFT to each one yields a range profile. Stacking N range profiles gives a range matrix, where one axis is time (chirps) and the other is distance (range bins). Because the range resolution is on the order of centimeters, the sub-millimeter chest motion is read from the phase rather than the frequency.

The traditional baseline. The method of Alizadeh et al. is used as the source of weak labels: pick the range bin with maximum power (where the person is), take the phase angle at that bin for every chirp, unwrap the wrapped phase, bandpass filter between 0.8 Hz and 3.0 Hz, and read the heart rate from the highest spectral peak.

Stage one — noise-contrastive pretraining. Three signals are extracted from each range matrix. A windowed heartbeat matrix around the person's range bin is fed to a heartbeat extractor to produce a predicted heartbeat signal; a windowed noise matrix from a randomly chosen different range bin is fed to a noise extractor to produce a noise signal; and the traditional method produces a pseudo-label. All three are randomly time-sampled and converted to power spectral densities to create sets of K samples. The Noise-Contrastive Triplet loss is simply the positive term (mean squared distance between pseudo-label and predicted heartbeat spectra, which should be small) plus a negative term (negative mean squared distance between predicted heartbeat and noise spectra, which should be large). No ground-truth physiological signal appears anywhere in this loss.

Stage two — augmented pseudo-labels. Stage one's prediction is coarse because the traditional pseudo-labels are noisy, so a refinement stage selects better labels. Candidates come from two sources: the traditional method applied to every range bin in the heartbeat window (2Δd+1 signals), and the pretrained heartbeat extractor's output. For each candidate, two distances are computed — a noise distance to the pretrained noise signal (larger is better) and a heartbeat distance to the pretrained heartbeat signal (smaller is better). The distance metric is the mean absolute error between the two signals' heart rates. In the decision module, if the candidate with the largest noise distance is also the one with the smallest heartbeat distance, that candidate is chosen. If not, the candidate with the smallest heartbeat distance is chosen only if its noise distance exceeds the pretrained heartbeat signal's own noise distance; otherwise the pretrained heartbeat signal is used. Training then continues with these enhanced pseudo-labels.

Training and evaluation details. Both extractors are randomly initialized and trained in two stages, with the generator inheriting stage-one pretrained models. AdamW with learning rate 1e-4 is used for 200 epochs per phase, with the best epoch chosen on validation sets. Evaluation uses 10-second windows and reports MAE, RMSE, and Pearson correlation r.

Data. The public Equipleth dataset contains 550 paired facial video and FMCW radar recordings from 91 subjects, with skin tones categorized by the Fitzpatrick scale as 28 light, 49 medium, and 14 dark, each subject providing six 30-second recordings. The authors' new RHB dataset has 240 FMCW radar instances from 80 volunteers, each with three separate 30-second sessions recorded while seated 0.5 to 1 meter in front of the radar board, captured at 120 frames per second. RHB uses 4-fold cross-validation with 50 subjects for training, 10 for validation, and 20 for testing per fold.

Why This Matters

Impact on research. The paper attacks the annotation bottleneck that limits radar physiological sensing: high-quality synchronized PPG labels are expensive, and the paper shows that the noise usually treated as an obstacle can instead be used as a training signal. It also connects radar sensing to the unsupervised pseudo-label literature that has been developed largely for classification and video, providing evidence that radar has domain-specific properties (lower SNR, phase-based sensing) that require specialized designs. The released RHB dataset and code give the community a new benchmark for label-free radar heartbeat estimation.

Real-world applications.

• Continuous contactless heart-rate monitoring in hospital wards, where adhesive sensors are inconvenient or a hygiene risk.
• Telehealth and

Authors’ abstract

Frequency Modulated Continuous Wave (FMCW) radars can measure subtle chest wall oscillations to enable non-contact heartbeat sensing. However, traditional radar-based heartbeat sensing methods face performance degradation due to noise. Learning-based radar methods achieve better noise robustness but require costly labeled signals for supervised training. To overcome these limitations, we propose the first unsupervised framework for radar-based heartbeat sensing via Augmented Pseudo-Label and Noise Contrast (Radar-APLANC). We propose to use both the heartbeat range and noise range within the radar range matrix to construct the positive and negative samples, respectively, for improved noise robustness. Our Noise-Contrastive Triplet (NCT) loss only utilizes positive samples, negative samples, and pseudo-label signals generated by the traditional radar method, thereby avoiding dependence on expensive ground-truth physiological signals. We further design a pseudo-label augmentation approach featuring adaptive noise-aware label selection to improve pseudo-label signal quality. Extensive experiments on the Equipleth dataset and our collected radar dataset demonstrate that our unsupervised method achieves performance comparable to state-of-the-art supervised methods. Our code, dataset, and supplementary materials can be accessed from https://github.com/RadarHRSensing/Radar-APLANC.