PaperlayerView exact revision on arXiv

Source-linked AI summary

An Overview of Algorithms for Contactless Cardiac Feature Extraction from Radar Signals: Advances and Challenges

Yuanyuan Zhang, Rui Yang, Yutao Yue, Eng Gee Lim, Zidong Wang

arXiv:2508.02122v1eess.SP

TL;DR

Contactless cardiac monitoring could reduce reliance on uncomfortable or unsuitable contact-based measurements, but radar cardiac-feature extraction has not been systematically reviewed. This paper surveys extraction algorithms, introduces a taxonomy, evaluates public datasets, and discusses challenges and future directions. It serves as a guide for understanding algorithm principles, application suitability, and recent developments.

  • Problem

    Radar architecture design has been reviewed more thoroughly than algorithms for extracting cardiac features from received radar signals.

  • Method

    The review proposes a taxonomy of radar cardiac-feature extraction algorithms, evaluates their pros and cons, and assesses public datasets containing synchronized radar and cardiac-feature signals.

  • Results

    The review identifies algorithm principles and application suitability, evaluates five public datasets, and discusses unsolved challenges and possible future directions.

  • Takeaways & Limitations

    Researchers and practitioners can use the review to understand radar cardiac-feature extraction methods and select suitable public datasets for future work.

Abstract

from arXiv · show

Contactless cardiac monitoring has vast potential to replace contact-based monitoring in various future scenarios such as smart home and in-cabin monitoring. Various contactless sensors can be potentially implemented for cardiac monitoring, such as cameras, acoustic sensors, Wi-Fi routers and radars. Among all these sensors, radar could achieve unobtrusive monitoring with high accuracy and robustness at the same time. The research about radar-based cardiac monitoring can be generally divided into the radar architecture design and signal-processing parts, where the former has been thoroughly reviewed in the literature but not the latter. To the best of the author knowledge, this is the first review paper that focuses on elaborating the algorithms for extracting cardiac features from the received radar signal. In addition, a new taxonomy is proposed to reveal the core feature of each algorithm, with the pros and cons evaluated in detail. Furthermore, the public datasets containing the received radar signal and ground-truth cardiac feature signal are listed with detailed configurations, and the corresponding evaluations may help the researchers select the suitable dataset. At last, several unsolved challenges and future directions are suggested and discussed in detail to encourage future research on solving the main obstacles in this field. In summary, this review can be served as a guide for researchers and practitioners to quickly understand the research trend and recent development of the cardiac feature extraction algorithms, and it is worth further investigating the relative area based on the proposed challenges and future directions.

I. INTRODUCTION

Contactless cardiac monitoring addresses the discomfort and limited long-term suitability of contact-based measurements, with radar offering a promising sensing approach for unobtrusive monitoring. This review focuses on cardiac feature-extraction algorithms, proposes a taxonomy, evaluates datasets, and identifies future challenges.

  • Motivation: Contact-based cardiac measurements can be uncomfortable for some patients and unsuitable for long-term monitoring, motivating contactless alternatives for smart homes, driver monitoring, and postdisaster search.Cardiac features include HR, BBI, HRV, ECG, and SCG.
  • Motivation: Cameras, acoustic sensors, Wi-Fi routers, and radars support contactless monitoring, but alternative sensors may face privacy, accuracy, or environmental-vulnerability concerns.Radar receives reflected signals containing heart vibration, chest-wall displacement, and ambient noise.
  • Radar Monitoring Pipeline: Radar cardiac monitoring separates cardiac-feature extraction from radar architecture design, with received signals requiring isolation from body movement, respiration, multiperson, and multipath interference.The monitoring process includes sensing, single-person signal isolation, and cardiac-feature extraction.
  • Research Gap: Earlier studies emphasized radar hardware and signal processing, while cardiac extraction commonly relied on FFT and BPF with limited accuracy and noise mitigation.The review identifies the extraction stage as a distinct area requiring systematic attention.
  • Contributions: The review proposes a taxonomy based on algorithmic core principles and details algorithmic advantages, disadvantages, and future improvements.The taxonomy is intended to help researchers and practitioners quickly understand methods used in radar-based cardiac feature extraction.
  • Contributions: The review evaluates public datasets containing synchronized radar signals and cardiac features, providing guidance on dataset usage and future design.It also discusses current challenges and future research directions.

A. Types, Working Principles and Operating Frequencies

Radar-based cardiac monitoring uses distinct waveform architectures and operating frequencies, each with different sensing capabilities and trade-offs. The review organizes extraction algorithms through a taxonomy intended to clarify their core features and applications.

  • Radar types: CW, FMCW, and IR-UWB radars use fixed-frequency waves, linearly increasing-frequency waves, and short wideband pulses, respectively.These waveform distinctions determine the available radar architecture and sensing behavior.
  • Radar types: CW radar has simple architecture but lacks range information, whereas FMCW radar improves SNR and enables chest-region range detection.FMCW uses frequency modulation to isolate reflections from the chest region.
  • Operating frequencies: Higher operating frequency narrows antenna beamwidth and can increase phase difference from heart vibration, improving monitoring sensitivity.The frequency-beamwidth relationship assumes antennas with the same diameter.
  • Usage trends: Since 2015, CW was most popular before 2020, FMCW interest has grown with commercial platforms, and IR-UWB has remained less popular because of architectural complexity.The trend is reported from the papers reviewed after 2015.
  • Algorithm taxonomy: The review introduces a taxonomy because domain-based or scenario-specific classifications are too coarse or narrow to reveal algorithmic core features.The taxonomy supports a structured presentation of extraction algorithms and their applications.

1) Fast Fourier Transform with Improved Filter:

FFT-based extraction estimates HR from spectral peaks, while improved filtering, differentiation, and time-frequency processing address noise, leakage, changing HR, and limited observation duration. These refinements balance resolution, real-time operation, and noise robustness.

  • Improved filtering: Fixed-passband filters are unsuitable when HR varies widely or noise components, such as RR harmonics and engine vibration, fall within the HR passband.Adaptive or multiple-passband strategies are introduced to address these conditions.
  • Differentiation: Differentiation suppresses RR harmonics and enhances HR components because heartbeat displacement acceleration exceeds respiration acceleration.A third-order differentiator achieved the best reported SNR and HR estimation among orders 0th to 3rd, while high orders can amplify high-frequency noise.
  • STFT: FFT spectra require suitable signal truncation and at least 15 sec observation for high frequency resolution, motivating STFT for shorter windows and time-related spectra.Reported sliding-window examples achieved 0.05 Hz resolution with 12 sec data and HR monitoring with an 8 sec window.
  • Spectral refinements: Short windows improve real-time monitoring but widen spectral lobes, so DCT and phase-alignment methods are used to improve estimation accuracy.The trade-off arises from overlapping main and side lobes in short-window spectra.
  • Wavelet transform: WT decomposes mixed signals across time and frequency using adjustable scale and shift parameters, supporting transient analysis beyond STFT.The scale parameter dilates or compresses the signal, while the shift parameter moves the wavelet along time.

5) Wavelet Transform:

Wavelet-transform and periodicity-based methods extract cardiac features by separating time-frequency components, detecting heartbeat-consistent peaks, or recovering repeated patterns. Their implementations include denoising, adaptive peak selection, autocorrelation, and template or probabilistic models.

  • Wavelet Transform: WT-based methods select scales, mother wavelets, or synchronized signal lengths to mitigate RBM, identify RR harmonics, and preserve HR peaks.Synchrosqueezing reallocates WT energy to reduce smearing and improve frequency resolution.
  • Wavelet Transform: WT also supports predenoising by locating time-frequency artifacts, applying moving-average attenuation, or using adaptive soft thresholds.These procedures target artifacts or noise while retaining signal continuity at threshold boundaries.
  • Peak detection: Heartbeat peaks are periodic whereas noise peaks are random, so derivative-based detection requires additional selection to distinguish heartbeats from other peaks.Detected peaks and heartbeat selections are illustrated with separate markers and lines.
  • Peak detection: Empirical BBI guides amplitude and temporal thresholds, adaptive passbands, and multi-peak windows for selecting heartbeat peaks and stabilizing HR estimation.The strategies reject closely spaced candidates and can use prior BBI information to constrain selection.
  • Autocorrelation: Autocorrelation estimates HR by locating repeated patterns in filtered time-domain signals, while self-similarity heatmaps can support CNN-based BBI estimation.After RR filtering, the dominant autocorrelation frequency represents the HR frequency.

2) Auto-correlation:

The review describes several periodicity- and morphology-based approaches for extracting cardiac events from noisy radar signals. These include autocorrelation, template matching, HMM-based state segmentation, and blind source separation.

  • Auto-correlation: Autocorrelation recovers periodic heartbeat patterns from filtered time-domain signals and estimates HR from dominant periodicity or peak intervals.WT preprocessing can enhance periodicity before autocorrelation, while phase regression can improve frequency resolution.
  • Auto-correlation: Self-similarity matrices create heatmaps of repeated heartbeat patterns that can be processed by a 2D CNN to estimate BBI.The matrix is calculated from two copies of the phase signal.
  • Cross-correlation with Template: Template-based methods use cross-correlation to match radar segments against cardiac morphology templates or sequential feature templates.The approach relies on consecutive heartbeats having similar morphology with possible stretching or compression.
  • Hidden Markov model: HMM-based extraction models four cardiac states—S1, systole, S2, and diastole—and derives cardiac features from intervals between identified states.The model is trained with radar signals whose states are labelled; hidden semi-Markov models additionally represent uneven state durations.
  • Blind source separation: Blind source separation decomposes radar mixtures containing respiration, RBM, cardiac vibration, and multipath interference using assumptions such as orthogonal spaces, independence, time-scale structure, or sparsity.The reviewed methods include MUSIC, ICA, EMD, VMD, and sparse signal reconstruction.

1) Multiple Signal Classification:

MUSIC decomposes mixed radar signals into signal and noise subspaces to isolate cardiac spectral components, but its performance depends strongly on selecting the source count P. ICA instead separates mixed observations into statistically independent components and requires sufficient or synthesized observations.

  • MUSIC: MUSIC applies covariance-matrix eigenvalue decomposition to separate cardiac signal components from noise subspaces.The P largest eigenvalues define the signal subspace, while remaining eigenvectors form the noise subspace.
  • MUSIC: The source-count parameter P must match ambient sources: large values create spurious peaks, whereas small values lose spectral components.Careful P selection enables high frequency resolution with short windows.
  • MUSIC: MUSIC can reveal heart-rate harmonics despite motion artifacts or car-vibration noise, with some studies reselecting P before each measurement.One reported search over P ∈[60, 200] selected P=140 after error evaluation.
  • ICA: ICA separates mixed signals into statistically independent components by estimating a decomposition matrix W from observed signals x.The method is illustrated as a general radar cardiac-feature extraction procedure.
  • ICA: Single-radar ICA requires at least as many observed signals as source signals, so whitened pseudo-observations must be generated.Prior work used wavelet-generated pseudo-signals, singular-value decomposition, or EMD-based pre-whitening before fast-ICA.

3) Empirical Mode Decomposition:

EMD-family methods decompose radar signals into intrinsic modes to identify cardiac components, while VMD and sparse signal reconstruction impose alternative structural assumptions. Their effectiveness depends on decomposition or noise parameters and on the quality of the underlying signal model.

  • Empirical Mode Decomposition: EMD iteratively decomposes mixed radar signals into intrinsic mode functions and a residual, allowing nonlinear modulated signals to be analyzed.Only some IMFs contain cardiac information, and conventional EMD can suffer mode mixing.
  • Empirical Mode Decomposition: EEMD variants select cardiac IMFs using signal power, frequency, or statistical differences to suppress radar interference and random body movement.Two-layer EEMD further separates background-scatter suppression from slow-time heart-rate estimation.
  • Empirical Mode Decomposition: EEMD performance depends strongly on white Gaussian noise amplitude: high amplitude distorts low-frequency extrema, while low amplitude leaves high-frequency mode mixing.Adaptive-noise and complete-ensemble variants were proposed to mitigate these effects and reduce residual noise or iterations.
  • Variational Mode Decomposition: VMD non-recursively decomposes signals into K narrow-band IMFs by minimizing total bandwidth while preserving their sum, helping avoid mode mixing.Its performance still relies on selecting the decomposition number K and penalty value α.
  • Sparse Signal Reconstruction: Sparse signal reconstruction models the heartbeat as a time-domain sparse signal and seeks the sparsest solution from mixed observations using a known basis matrix Φ.The reconstructed spectrum can distinguish heartbeat frequency components from noise.

D. Deep Learning Methods

Deep-learning methods learn nonlinear mappings from radar signals to cardiac features using CNNs, recurrent networks, and self-supervised contrastive learning. Architectures increasingly combine spatial, temporal, and fine-grained feature extraction, while DCL removes the need for ground-truth signals during representation learning.

  • Convolutional Neural Network: CNN-based cardiac extraction applies multiple convolutional layers to learn high-level features and outputs signals such as SCG or ECG.Training commonly minimizes a loss such as the ℓ2-norm until convergence.
  • Convolutional Neural Network: CNN architectures have been adapted with filter-like layers, encoder-decoder structures, and transformer blocks for fine-grained cardiac reconstruction.These designs combine frequency-component extraction, consecutive-window cascading, or temporal-spatial encoding.
  • Recurrent Neural Networks: LSTM networks use gates to retain or forget information across time, making them suitable for translating radar sequences into cardiac-feature sequences.CNN-LSTM and Bi-LSTM variants combine spectral, spatial, temporal, or bidirectional dependencies for cardiac monitoring.
  • Deep Contrastive Learning: Deep contrastive learning learns nonlinear features without ground-truth signals by grouping embeddings from positive samples and separating different samples.Positive samples may be augmented observations, while negative samples are randomly drawn observations.
  • Deep Contrastive Learning: For strong body movement, DCL constructs positive samples from original and time-shifted radar signals and negative samples from randomized versions before feature extraction.The resulting representation is used to obtain the heart-rate signal.

IV. REVIEW OF PUBLIC DATASETS

The review catalogs public radar datasets with synchronized cardiac ground truth and evaluates their suitability for different scenarios and algorithms. Datasets range from short, mild-noise records for early testing to long, varied records supporting realistic algorithm development, though broader coverage remains needed.

  • Dataset review: Public datasets are important because collecting synchronized radar and cardiac recordings is labor-intensive and technically complex.The review lists datasets with collection configurations and evaluates their potential use across scenarios and algorithms.
  • Dataset coverage: Dataset-specific configurations broaden research targets, including heart-sound detection, arbitrary radar positions, long-distance monitoring, vigorous movement, and smart-bed sensing.Examples include synchronized ECG and PCG, multiple radar orientations, whole-body FMCW monitoring, and two under-bed radars.
  • Dataset suitability: Short datasets from Yoo et al. and Edanami et al. support initial algorithm testing under single, mild-noise scenarios.They are suitable for testing real-time spectrum methods or validating early-stage algorithm designs.
  • Dataset suitability: Longer or more varied datasets from Chen et al. and Schellenberger et al. support deep-learning training, parameter selection, and evaluation under realistic cardiac variation.Schellenberger et al. provides 24 hours of raw radar and ground-truth signals with procedures inducing abrupt HRV.
  • Future dataset needs: The review concludes that existing public datasets remain insufficient for all research interests and calls for high-quality data collected across ample scenarios.Public releases also reduce duplicated collection efforts and improve comparisons between algorithms.

V. CHALLENGES AND FUTURE WORK

Radar cardiac-feature extraction algorithms offer distinct trade-offs in real-time performance, noise robustness, prior-knowledge requirements, and deployment complexity. Key challenges include parameter selection, calibration, dataset coverage, and implementation on compact devices.

  • Spectrum-based methods: Spectrum-based methods support good real-time performance on hardware with limited computational resources.Several commercial data-capture boards already embed these algorithms on FPGA.
  • Spectrum-based methods: Spectrum-based methods estimate only one HR value per segment and can be shifted by window truncation or abrupt HRV.Noise such as RR harmonics and slight RBM further limits their reliability.
  • Periodicity-based methods: Periodicity-based methods identify individual heartbeats, support fine cardiac-event segmentation, and resist nonperiodic noise.They use template matching or probabilistic prediction based on cardiac periodicity.
  • Periodicity-based methods: Periodicity-based methods require prior cardiac knowledge and may need calibration for new participants.Templates optimized on datasets may not fit diverse individuals or balance rare-event detection with accurate HR estimation.
  • Deep learning methods: Complex deep learning models remain difficult to deploy on compact real-time devices because they require substantial computation and memory.This creates a practical deployment boundary despite their algorithmic capabilities.
  • BSS methods: BSS methods avoid prior cardiac-event knowledge but require empirically selected decomposition parameters.Poor parameter choices can cause mode mixing or over-decomposition, preventing HR extraction.

4) Deep Learning Methods:

Deep learning methods can model complex cardiac signals and potentially handle difficult monitoring conditions, but their robustness and deployment depend on representative datasets and adequate computational resources. Across algorithm categories, practical use requires balancing accuracy, robustness, complexity, and hardware constraints.

  • Capabilities: Deep learning methods model complex nonlinear projections and can produce fine-grained cardiac signals.Specialized networks such as LSTMs can also memorize long-term dependencies for stable monitoring.
  • Capabilities: Deep learning may resist large RBM, long-distance monitoring, fine-grained features, and irregular cardiac events.These capabilities are described as potential advantages rather than universally established performance.
  • Limitations: Deep learning performance relies on large, full-featured datasets and may fail on out-of-distribution data.Examples include training on normal participants or on one noise type while evaluating another.
  • Comparative deployment: Spectrum-based methods are already embedded in commercial radar platforms and offer real-time operation with limited resources.Their robustness to real-world noise remains limited, especially under slight RBM.
  • Comparative deployment: Periodicity-based methods resist nonperiodic noise but sacrifice robustness to periodic noise and detection of rare cardiac events.Their behavior reflects the prior cardiac structure built into the method.
  • Comparative deployment: BSS methods can mitigate RR noise and slight RBM and may address mutual-radar interference and multipath propagation.More experiments are needed to establish their robustness outside laboratory settings.
  • Future directions: Future platforms must balance algorithm complexity, computational resources, and desired monitoring accuracy or robustness.Neural-network processing modules may be needed to support complex algorithms on commercial devices.
  • Open challenges: Strong or consistent body movement can significantly distort radar signals and affect cardiac-feature extraction.Current methods generally tolerate only slight or abrupt movement, while vigorous-motion mitigation remains mode-specific.

2) Complex Electromagnetic Environment:

Real-world cardiac monitoring remains constrained by complex electromagnetic environments, long distances, interruptions, orientation, platform motion, abrupt HRV, and limited HR passbands. The review identifies experimental and algorithmic directions for extending monitoring beyond controlled laboratory conditions.

  • Electromagnetic environment: Cardiac-monitoring studies rarely test mutual-radar interference and multipath propagation in genuinely complex electromagnetic environments.Existing experiments commonly use mild interference or simple rooms with prominent human-body reflections.
  • Long-range monitoring: Long-range monitoring weakens body reflections and increases background clutter, reducing cardiac-monitoring accuracy.Compact systems typically operate at 0.3−2 meters, while longer ranges may require complex hardware or digital beamforming.
  • Long-range monitoring: Long-distance monitoring can be interrupted by hand movements or people entering the radar's view.Current systems detect abrupt signal-power increases and substitute previous estimates, but prolonged interruptions terminate monitoring.
  • Orientation: Smart-home monitoring may require orientation-independent sensing because optimal SNR is obtained when the radar faces the chest.Distributed radar networks are proposed because a single radar's orientation cannot be guaranteed.
  • Platform stability: Radar self-movement adds platform-vibration noise that is difficult to remove and is rarely evaluated experimentally.Existing work identified this noise in some settings, but the reported method could reconstruct only respiration and could not handle large RSM.
  • HRV: Large or abrupt HRV remains difficult to detect because many algorithms rely on intrinsic cardiac-event patterns.This limitation matters for monitoring patients with conditions such as arrhythmia.
  • Heart-rate range: Narrow passbands such as 1−1.6 Hz may exclude heart rates after mild exercise and require stronger noise robustness when widened.Future techniques should cover a wider HR range while resisting in-band noise.

1) Deep Learning for Spatiotemporal Cardiac Feature Extraction:

Recent directions combine richer learning strategies with better signal quality and data resources to improve cardiac feature extraction. Key priorities include spatiotemporal modeling, abnormal-event data, transfer learning, evaluation metrics, and SNR improvement.

  • Spatiotemporal feature extraction: Combining deep learning blocks in series or parallel could extract cardiac characteristics in both spatial and temporal domains.The passage contrasts this approach with research using a single model focused on either spatial or temporal features.
  • Abnormal cardiac features: Most current algorithms cannot reveal abnormal cardiac features such as arrhythmia because abnormal cardiac data are scarce.The review identifies data scarcity as restricting further development of deep learning algorithms for abnormal-feature detection.
  • Abnormal cardiac features: Advanced data augmentation is proposed to generate more valuable training data for abnormal cardiac events than simple overlapping, stretching, noising, and random polynomials.The suggested direction is to base augmentation on real-collected datasets.
  • Transfer learning: Transfer learning could address inaccurate extraction caused by individual differences and insufficient data by pre-training on related signal tasks before fine-tuning with limited radar-ECG data.The review notes that current transfer learning mainly uses pure ECG data for diagnosis, whereas related-task pre-training is proposed for radar-based extraction.
  • Evaluation: Mean squared error can flatten diagnostically important cardiac peaks and valleys, motivating evaluation metrics that preserve signal features.The review calls for metrics beyond global average distance between ground-truth and estimated signals.
  • SNR improvement: Improving received-signal SNR before extraction can reduce downstream denoising or decomposition because coarse heartbeat signals may be directly measured from high-SNR inputs.Suggested approaches include antenna design, digital beamforming, and programmable wireless environments.
Loading 2508.02122v1…