Radio-PPG: Non-Contact PPG Sensing Dataset

• The Radio-PPG dataset is a non-contact physiological sensing resource that pairs OFDM-modulated radio reflections with reference PPG signals.
• It employs advanced signal processing techniques including PCA, wavelet denoising, and cross-correlation for precise waveform synchronization and synthetic PPG reconstruction.
• The extended dataset supports digital twin PPG synthesis using deep learning models, achieving high fidelity in capturing vital-sign waveforms for potential remote health monitoring.

Searching arXiv for the cited Radio-PPG and related non-contact physiological sensing datasets. The Radio-PPG dataset denotes a line of non-contact physiological sensing datasets in which a software-defined radio (SDR) illuminates the chest with an orthogonal frequency division multiplexing (OFDM) signal and a nearby SDR records the reflected, chest movement-modulated radio signal, while a reference photoplethysmography (PPG) sensor provides synchronized ground truth. In the original formulation, this paradigm was introduced as a method for non-contact acquisition of the PPG signal from radio reflections, producing a custom dataset of 160 minutes from 16 healthy young subjects and a waveform translation pipeline from radio segments to synthetic PPG (Filho et al., 2024). A later paper extended the same general concept to a larger dataset of 300 minutes from 30 healthy subjects, adding three body-vital labels and evaluating deep neural representations for digital twin PPG synthesis under a 6G/WiFi integrated sensing and communication setting (Filho et al., 26 Sep 2025).

1. Conceptual definition and scope

Radio-PPG is distinct from camera-based remote photoplethysmography datasets because its primary observation channel is radio rather than video. The original study states that a software-defined radio exposes the chest of a subject sitting nearby to an OFDM signal with 64 sub-carriers at a center frequency 5.24 GHz, while another SDR in the close vicinity collects the modulated radio signal reflected off the chest; the resulting dataset contains both raw radio data and the reference PPG signal (Filho et al., 2024). The later study describes an analogous acquisition setting at 5.23 GHz using a wideband 6G/WiFi signal and frames the output as photoplethysmogram digital twin synthesis from near synchronous radio and PPG data, with additional labels for heart rate, SpO$_2$, and respiratory rate (Filho et al., 26 Sep 2025).

Within this literature, the term therefore refers both to a specific dataset name and to a sensing formulation: paired radio reflections and reference PPG, collected under controlled conditions, then processed as a cross-modal regression or translation problem. A common misconception is to treat Radio-PPG as a direct optical PPG measurement performed at a distance. The source papers instead describe synthetic or reconstructed PPG waveforms inferred from radio observations, with contact PPG retained as ground truth (Filho et al., 2024).

2. Original Radio-PPG-16 acquisition protocol

The original dataset, explicitly named Radio-PPG-16, comprises 16 healthy young volunteers, with 10 male, 6 female, and an age range of 25 to 32 years (Filho et al., 2024). Each subject participated in two measurement sessions, each session lasted 5 minutes, and the total duration per subject was 10 minutes. The resulting total dataset length was 160 minutes.

The acquisition protocol was deliberately controlled. The subject sits still about 35 cm from the SDR transmitter-receiver pair in order to avoid motion artifacts. During recording, radio data and reference PPG data are captured simultaneously. The radio data are reflected signals off the chest, and the reference PPG is recorded with the subject’s finger placed on the MAX86150 PPG sensor. All procedures were approved by the ethical institutional review board (EIRB) at KAUST, Saudi Arabia (Filho et al., 2024).

The radio instrumentation consisted of two USRP N210 SDRs, one transmitter and one receiver, each connected to a workstation via Ethernet and each attached to a directional horn antenna, forming a SISO setup. The transmitter generated an OFDM signal with 64 sub-carriers, QPSK per sub-carrier, and a cyclic prefix of 16 samples appended to each symbol. The transmitter antenna gain was 40 dB. The receiver removed the cyclic prefix and logged the complex-valued time-domain data. The center frequency was 5.24 GHz, the baseband sampling rate was 20,000 samples/sec (20 kHz), and the bandwidth was 20 kHz. The reference MAX86150 module sampled the PPG signal at 2.5 kHz (Filho et al., 2024).

The recorded formats reflect the cross-modal nature of the dataset. Radio data are complex-valued time sequences, recorded at 20 kHz, and PPG data are real-valued, recorded at 2.5 kHz. Each session captures raw radio data and a reference PPG waveform; the detailed dataset description specifies 64 channels/sub-carriers; data from 16 used (Filho et al., 2024).

3. Signal processing, alignment, and regression in the original dataset

The original processing pipeline is specified stepwise and combines denoising, dimensionality reduction, normalization, segmentation, and synchronization before regression (Filho et al., 2024). For the reference PPG data, the sequence is: manual artifact removal, wavelet denoising (db2 family) for baseline drift estimation and removal, a 12th order Butterworth low-pass filter (cutoff 3.4 Hz), Z-score normalization, and segmentation into 2.2-second non-overlapping segments. The normalization is given as

$x_{z} = \frac{x - \mu}{\sigma}$

where $x$ is the sample, $\mu$ is mean, and $\sigma$ is standard deviation (Filho et al., 2024).

For the radio data, the preprocessing begins with sub-carrier selection: 16 equally spaced out of 64 sub-carriers. Principal component analysis is then applied separately to the real and imaginary parts of the complex-valued radio matrix $X \in \mathbb{C}^{16 \times N}$:

$$\begin{align*} &X_R = \text{Re}(X),\quad X_I = \text{Im}(X) \ &\hat{X}_R = \text{PCA}_1(X_R),\quad \hat{X}_I = \text{PCA}_1(X_I) \end{align*}$$

The outputs are concatenated as $$\hat{X} = [\hat{X}_R; \hat{X}_I] \in \mathbb{R}^{2N}$$, followed by taking the modulus $|\hat{X}|$. Denoising is then performed via Discrete Wavelet Transform (DWT) with 10-level decomposition, zeroing coefficients corresponding to non-blood-volume changes in wavelet reconstruction, followed by Z-score normalization and 2.2-second non-overlapping segmentation (Filho et al., 2024).

Synchronization is a central element because simultaneous recording alone does not guarantee peak-to-peak alignment. The stated purpose of the cross-correlation stage is to achieve peak-to-peak alignment between radio and PPG segments, minimizing phase mismatch. The method computes

$R_{xy}(\tau) = \sum_{n} x[n]\, y[n + \tau]$

where $x_{z} = \frac{x - \mu}{\sigma}$0 is the radio segment and $x_{z} = \frac{x - \mu}{\sigma}$1 is the PPG segment. The time shift $x_{z} = \frac{x - \mu}{\sigma}$2 that maximizes $x_{z} = \frac{x - \mu}{\sigma}$3 is used for alignment. The paper describes this as ensuring that the morphological peaks, i.e., heart beats, are aligned across modalities (Filho et al., 2024).

After synchronization, the study proceeds to waveform translation. Both the radio and PPG segments are converted into the frequency domain using discrete cosine transform (DCT), and a non-linear regression is learned between them. The synthetic PPG signal is reconstructed by taking the inverse DCT of the regression output. The reported reconstruction performance is a mean absolute error of 8.1294 (Filho et al., 2024). The authors further state that the synthetic PPG waveform preserves clinically significant features, including heart rate and waveform morphology; this suggests that the dataset was designed not merely for scalar vital-sign estimation but for waveform-level inference.

4. Expanded Radio-PPG dataset and digital twin formulation

A later paper uses the same dataset name, Radio-PPG, for a larger collection designed for photoplethysmogram digital twin synthesis using deep neural representation of 6G/WiFi ISAC signals (Filho et al., 26 Sep 2025). In this version, the dataset consists of 300 minutes worth of near synchronous 64-channel radio data, PPG data, along with the labels (three body vitals) of 30 healthy subjects. The cohort comprises 30 healthy young adults, specifically 15 males aged 22–36 and 15 females aged 22–32, and each subject again contributes 2 sessions of 5 minutes each, yielding 300 minutes total. The study was approved by the Institutional Biosafety and Bioethics Committee, KAUST, Protocol 23IBEC002, and all subjects gave written consent (Filho et al., 26 Sep 2025).

The hardware remains SDR-centric. The setup uses two USRP N210 SDRs, each connected to a PC and a directional horn antenna. One horn antenna transmits a 6G/WiFi-like OFDM signal toward the subject’s chest and the other receives the reflected signal. The carrier frequency is 5.23 GHz, the bandwidth is 200 kHz, and the sample rate is 20,000 samples/sec (20 kHz). The radio modality is the raw 64-channel complex-valued OFDM channel frequency responses (CFR). Reference PPG is acquired with a MAX86150 module, specifically the red-channel photoplethysmogram, sampled at 200 Hz. The three physiological labels—heart rate, SpO$x_{z} = \frac{x - \mu}{\sigma}$4, respiratory rate—are acquired with a Massimo pulse oximeter, sampled at 1 Hz (Filho et al., 26 Sep 2025).

The later synchronization method is described as template matching optimization. The lag $x_{z} = \frac{x - \mu}{\sigma}$5 is chosen to maximize the inner product between the radio-derived waveform $x_{z} = \frac{x - \mu}{\sigma}$6 and the reference PPG $x_{z} = \frac{x - \mu}{\sigma}$7:

$x_{z} = \frac{x - \mu}{\sigma}$8

This lag is then used to time-align paired signal segments (Filho et al., 26 Sep 2025). The corresponding preprocessing pipeline differs from the original one in several details. The OFDM data are reduced from 64 to 16 non-redundant channels, each complex channel is split into real and imaginary parts to yield a 32-channel real-valued tensor, the signals are segmented into 2.5-second non-overlapping windows, and per-channel z-score normalization is applied. For robustness, Gaussian noise with $x_{z} = \frac{x - \mu}{\sigma}$9 and $x$0 is added as data augmentation. The reference PPG is detrended via wavelet decomposition (db2 family), denoised with a 12th-order lowpass Butterworth filter, 4 Hz cutoff, segmented into 2.5 s windows, normalized by Z-score, and similarly augmented (Filho et al., 26 Sep 2025).

Two model families are evaluated. The first is DCT + MLP, in which radio and PPG segments are converted to DCT type-II coefficients and a 5-layer MLP learns the mapping. The second is a custom 1D U-NET model in cascade, consisting of an Approximation Network and a Refinement Network: MultiRes U-NET. The hybrid loss combines mean absolute error, first derivative error, and second derivative error, with equal weights $x$1. The study uses both a random 80/20 split and Leave-Two-Subjects-Out (LTSO) cross-validation (Filho et al., 26 Sep 2025).

The reported main result is that the U-NET model achieves an impressive relative mean absolute error of 0.194 with a small ISAC sensing overhead of 15.62%, for DT-PPG synthesis (Filho et al., 26 Sep 2025). The paper also states that the quality assessment of the synthetic DT-PPG, performed via DT-PPG-based vitals estimation and feature extraction, is at par with that of reference PPG-based vitals estimation and feature extraction. A plausible implication is that the later Radio-PPG dataset is intended as a benchmark for waveform fidelity, downstream vital-sign estimation, and resource-aware sensing within ISAC systems rather than only for direct waveform regression.

5. Dataset variants, labels, and relation to optical non-contact datasets

The name “Radio-PPG dataset” is therefore used for at least two closely related resources with overlapping sensing logic but different scale and labeling profiles.

Variant	Acquisition profile	Status/use
Radio-PPG-16	16 subjects, 160 minutes, radio + reference PPG	Dataset name given; public sharing not explicitly mentioned
Radio-PPG	30 subjects, 300 minutes, 64-channel radio + PPG + HR/SpO$x$2/RR labels	Used for DT-PPG synthesis and downstream validation

The original Radio-PPG-16 description states that the paper does not explicitly mention public sharing, download links, or access instructions (Filho et al., 2024). The later Radio-PPG description provides detailed acquisition and evaluation protocols but, in the supplied record, does not specify a public release mechanism (Filho et al., 26 Sep 2025). As a result, access conditions are less explicit than in several contemporary video-based datasets.

Relative to optical datasets, Radio-PPG occupies a different point in the design space. DLCN is a large-scale, publicly available remote photoplethysmography dataset with approximately 13 hours of video data from 98 participants under four representative nighttime lighting scenarios, together with synchronized PPG, HR, and SpO$x$3 (Li et al., 6 Jul 2025). LADH provides 240 synchronized RGB and infrared facial videos from 21 participants across five common personal care scenarios, with ground-truth PPG, respiration, and blood oxygen signals, and is explicitly open with code (Ma et al., 11 Jun 2025). VideoPulse is a neonatal facial-video dataset with 157 recordings totaling 2.6 hours from 52 neonates, containing synchronized reference PPG, HR, and SpO$x$4, but it is not publicly downloadable due to privacy concerns (Dewagiri et al., 27 Feb 2026). VicarPPG 2 and CleanerPPG were introduced for HR and HRV evaluation, combining webcam video with synchronized ECG and PPG, plus hand-cleaned beat annotations (Gudi et al., 2020).

These comparisons clarify that Radio-PPG is not primarily a variant of facial-video rPPG benchmarking. It is a radio-reflection dataset for cross-modal biosignal synthesis. This distinction matters methodologically: optical datasets emphasize illumination, facial ROI localization, and video-domain augmentation, whereas Radio-PPG emphasizes OFDM channel selection, complex-valued signal handling, synchronization between radio and PPG, and waveform reconstruction from radio observations.

6. Research significance, limitations, and interpretive cautions

The original paper attributes great clinical significance to the synthetic PPG waveform because it could be used for non-contact performance assessment of cardiovascular and respiratory systems of patients suffering from infectious diseases, for example covid19 (Filho et al., 2024). The later paper similarly frames the dataset as a foundational step toward non-contact screening tools for covid-19, cardiovascular diseases, and well-being assessment of people with special needs (Filho et al., 26 Sep 2025). Across both formulations, the proposed benefits include continuous, seamless, and in-situ health monitoring, support for remote healthcare solutions, and a modality that, according to the authors, favors privacy over camera-based systems (Filho et al., 2024).

At the same time, the dataset family has explicit constraints. Both main versions were collected from healthy subjects, and the acquisition protocols instructed participants to sit still to avoid motion artifacts (Filho et al., 2024, Filho et al., 26 Sep 2025). The original dataset is limited to 16 healthy young subjects, while the later version expands to 30 healthy subjects but remains within a young-adult cohort. This suggests that cross-population robustness, pathological variability, and performance under unconstrained motion remain open questions rather than established properties of the dataset. The later paper itself identifies future prospects in larger/more diverse cohorts, pathological conditions, and exploration of sex, distance, and environmental variables (Filho et al., 26 Sep 2025).

Another interpretive caution concerns the meaning of “PPG acquisition.” In these papers, the non-contact signal is not a directly sensed optical PPG but a radio-derived synthetic waveform trained or aligned against a contact reference PPG. That distinction is central for evaluation: the relevant target is morphological and physiological agreement with the reference waveform, not merely detection of mean heart rate. The use of DCT-domain regression in the original work and U-NET-based time-domain synthesis with derivative-aware loss in the later work underscores that Radio-PPG is best understood as a paired cross-modal waveform dataset for biosignal translation, digital twinning, and downstream vitals inference rather than as a conventional scalar-label vital-sign corpus (Filho et al., 2024, Filho et al., 26 Sep 2025).