Seismocardiography: Cardiac Mechanics via Chest Vibrations

• Seismocardiography (SCG) is a non-invasive technique that captures chest wall vibrations from cardiac activity using accelerometers, enabling insights into heart function and timing events.
• Advanced signal processing methods such as FFT demodulation, time-frequency analysis (STFT, PCT, SPWVD), and hybrid multimodal frameworks enhance SCG's ability to accurately detect cardiac events.
• Recent developments in machine learning, sensor innovations (contactless, earable), and open-source toolchains are expanding SCG applications in diagnostics, wearable monitoring, and emotion recognition.

Seismocardiography (SCG) is the measurement and analysis of the mechanical vibrations of the chest wall induced by cardiac activity, typically acquired using an accelerometer placed on the sternum or chest surface. SCG provides a direct mechanical surrogate of heart function, capturing events such as valve openings/closures, myocardial contractions, and blood flow-related thrusts. Over the last two decades, SCG has progressed from a niche research modality to a platform with demonstrated use in cardiac timing interval estimation, non-invasive hemodynamic assessment, wearable health monitoring, emotion recognition, and contactless or earable implementations.

1. Signal Characteristics, Amplitude Modulation, and Frequency-Domain Analysis

The SCG signal consists of a time series reflecting the superposition of complex, time-varying, and nonlinear cardiac events. Importantly, SCG signals exhibit pronounced amplitude modulation (AM) effects. The mechanical cardiac cycle can be modeled as an amplitude-modulated carrier:

$y(t) = [A + M \cos(\omega_m t)] \sin(\omega_c t)$

where $A$ is the carrier amplitude, $M$ is the modulation amplitude, $\omega_c = 2\pi f_c$ is the carrier angular frequency (arising from high-frequency mechanical oscillations), and $\omega_m = 2\pi f_m$ is the modulation frequency reflecting underlying cardiac periodicity. This structure leads to the modulation information (heartbeat and respiratory rates) being embedded in spectral sidebands around the carrier.

Direct application of a fast Fourier transform (FFT) to the raw SCG often results in the temporal modulation information (e.g., ~1 Hz for the heartbeat) being nearly "invisible" in the power spectrum, due to sideband masking by the carrier (Peters et al., 2010). Demodulation is achieved through preprocessing such as:

• Rectification (half- or full-wave): transforms the modulated signal such that the modulation frequency $f_m$ manifests as a distinct spectral line.
• Teager–Kaiser (T–K) operator: For a discrete signal $x_n$, the instantaneous energy estimate is $E_n = x_n^2 - x_{n-1} x_{n+1}$. This operator enhances AM components, making the key cardiac/respiratory frequencies distinct in the spectrum.

Both approaches significantly clarify the frequency-domain representation, improving the detection of heart and respiratory rates, and can be essential in studies of heart rate variability or detection of abnormal rhythms.

2. Time-Frequency, Morphological, and Multimodal Processing

Time-frequency analysis is necessary for characterizing SCG signals whose components are nonstationary and exhibit time-varying spectral properties. Approaches include:

• Short-Time Fourier Transform (STFT): Windowed Fourier analysis; simple but exhibits fundamental time-frequency resolution trade-off.
• Polynomial Chirplet Transform (PCT): Extends chirplet transforms with polynomial kernels, adapting to nonlinear time-frequency trajectories intrinsic to SCG waveforms.
• Smoothed Pseudo Wigner-Ville Distribution (SPWVD): Applies joint time and frequency smoothing to reduce cross-term artifacts from the classical Wigner-Ville distribution, yielding tight, interpretable spectral localization.

PCT and SPWVD have been shown to deliver superior instantaneous frequency (IF) estimation accuracy compared to STFT and WVD, with lower normalized root-mean-square errors particularly in signals with time-varying frequency content (Taebi et al., 2017).

For fiducial point extraction (e.g., AO, AC, IM, IC, pAC, MO), hybrid multimodal delineation frameworks integrate wavelet-based scalographic processing of photoplethysmogram (PPG) signals with amplitude histogram-based rules applied to SCG. PPG is used to precisely align diastolic features; masking and envelope detection followed by amplitude-based rules enable robust extraction of all clinically significant SCG points (Choudhary et al., 2020). These frameworks have demonstrated over 95% sensitivity and accuracy in fiducial detection, further supporting their use in diagnostic and monitoring applications.

3. Machine Learning, Deep Learning, and Signal Reconstruction

Recent advances leverage both traditional classifiers and deep neural models for automated SCG signal characterization:

• Adaptive Feature Extraction: By dividing events into variable-length bins adaptively scaled to signal variability, salient, compact feature vectors are constructed. When used with SVMs, F1-scores as high as $0.91 \pm 0.05$ can be achieved for tasks such as classifying events by lung volume (Taebi et al., 2018).
• Stacked Autoencoders: For breathing-state recognition from SCG, 15 morphological features (including time/frequency domain descriptors, energy/entropy, kurtosis, and spectral centroids) are extracted and classified using a multi-stage stacked autoencoder, achieving $>$91% accuracy for discrimination among breathless, normal, and labored breathing (Choudhary et al., 2020).
• End-to-End Deep Learning: SCG-based detection of cardiac R-peaks has been achieved via fully convolutional U-net–like networks, which regress from noisy SCG input to a surrogate (distance transform) ECG target. Reported sensitivity and positive predictive value reach 0.98 each, even on noisy signals (Suresh et al., 2020).
• SCG Super-Resolution and Modality Translation: In unconventional form factors (e.g., earables), deep denoising-autoencoders and Transformer-based models are used to reconstruct fine-grained SCG from noisy, low-rate in-ear ballistocardiogram (BCG) signals. Pipeline robustness is improved by multi-axis fusion, channel attention, and explicit modeling of packet loss and motion artifacts (Fu et al., 12 Jan 2025).

For SCG-based assessment in uncontrolled conditions or across devices, adaptation and personalization steps—such as frequency-domain normalization to account for inter-device differences—enable consistent high-fidelity SCG reconstruction from nonstandard acoustic, video, or radar-derived inputs (Zhang et al., 13 Sep 2025).

4. Sensing Modalities, Signal Acquisition, and Device Innovations

SCG signals are conventionally acquired via tri-axial MEMS accelerometers placed on the lower sternum, xiphoid process, or adjacent costal notches. Signal fidelity and the accuracy of cardiac time interval (CTI) measurement are affected by sensor location, with studies highlighting systematic differences in CTI and morphology across top, middle, and bottom sternal placements. While heart rate estimation via SCG shows high correlation with ECG across sites ($R^2 > 0.98$ for top/middle), absolute CTIs may differ, underscoring the need for standardized protocols (Mann et al., 2023).

Recent innovations have greatly expanded SCG acquisition paradigms:

• Contactless SCG: Leveraging Gunnar-Farneback optical flow on smartphone video, chest wall vibration signatures are extracted from skin pixels. Differences in RoI placement modulate signal similarity to reference accelerometers (MSE range: 0.2–1.5 in head-to-foot axis), but HR estimation remains within 0.8 bpm of ECG gold standard (Rahman et al., 18 Aug 2024).
• In-ear/Earable SCG: BCG signals measured via commercial earbud IMUs are processed to reconstruct SCG waveforms. Stationary wavelet transforms, multi-channel fusion, neural denoising, and Transformer-based super-resolution restore the needed temporal fidelity for HR, HRV, and arrhythmia detection (Fu et al., 12 Jan 2025).
• Acoustic Microphone-based Monitoring: Hearables capturing “lub-dub” sounds with microphones (or reversed speakers) can reconstruct SCG/gyrocardiogram (GCG) using convolutional autoencoders and frequency-domain normalization. Within- and cross-device correlation coefficients to chest reference reach 0.88–0.95, with <1% micro-event timing error (Zhang et al., 13 Sep 2025).
• Noncontact mmWave Radar: mmWave MIMO radar records chest vibrations from multiple locations. Signal phase is processed to infer precise displacement, with cross-correlation to multi-point SCG reference signals reaching $0.84–0.88$ (Ren et al., 14 Nov 2024). This approach allows multi-point, noncontact, and synchronous vital sign monitoring.

5. Clinical, Physiological, and Affective Applications

SCG enables non-invasive measurement of parameters including:

• Cardiac Time Intervals (CTIs): Including pre-ejection period (PEP), left ventricular ejection time (LVET), electromechanical systole (QS2), and heartbeat intervals, with applications in longitudinal monitoring and disease stratification.
• Cuffless Blood Pressure: SCG-based LVET’ calculation (interval between AO and pAC peaks) is used in a subject-calibrated log-linear model for BP estimation, achieving mean errors of –0.19 ± 3.3 mmHg (SBP) and –1.29 ± 2.6 mmHg (DBP), meeting IEEE standards for cuffless BP monitoring (Das et al., 2020).
• Cardiac Valve and Hemodynamic Assessment: Deep CNN models trained on SCG scalograms can regress cardiovascular metrics such as peak aortic Vmax (R = 0.76 with 4D flow MRI) and classify valve pathologies (ROC-AUC 0.83–0.95), supporting screening and pre-imaging triage (Khani et al., 2023). ECG-free template-bank–based detection of AO/MC is also feasible, with F1-scores >95% in template-matched subjects (Rahman et al., 18 Aug 2024).
• Emotion Recognition and Affective Computing: SCG (AO peak-based HR/HRV) paired with accelerometry-derived respiration (ADR) provides features for valence/arousal recognition, with accuracy and F1 scores comparable to ECG and BVP, enabling single-sensor emotion recognition frameworks (Rahmani et al., 30 Nov 2024).

6. Algorithmic Tooling, Open-Source Frameworks, and Methodological Considerations

• Toolchains: The PulsatioMech open-source MATLAB toolbox provides modular support for multifractal analysis via Holder exponent estimation, wavelet leader approaches, MODPWT-based Shannon entropy, and AR feature extraction using Burg’s method. It enables standardized SCG processing, ensembling, feature extraction, and spectral analysis, supporting reproducibility and interoperability across studies (Zavanelli, 10 Jan 2024).
• Personalization and Domain Adaptation: Model performance is sensitive to domain shift between controlled and real-world settings or among device platforms. Cross-dataset analyses underline that personalization (fine-tuning on per-subject/data segments) and multi-channel input strategies (accelerometer + gyroscope) are vital for robustness, often boosting segmentation/detection F1-scores by up to 5% (Craighero et al., 8 Aug 2024).
• Standardization and Validation: Given sensor location sensitivity and population variability in SCG morphology, standardized acquisition, calibration with gold-standard methods (ECG, PPG, PCG, echocardiography), and protocol harmonization are required for rigorous, generalizable SCG research and clinical adoption (Mann et al., 2023).

---

Seismocardiography encompasses a set of tightly coupled physical, signal processing, and computational challenges. Progress in demodulation theory, time-frequency analysis, adaptive feature construction, and modality translation (contact, ear, video, radar) underpins newly demonstrated applications in non-invasive diagnostics, lightweight wearable monitoring, and even affective computing. Anchored by advances in both theory and robust open-source implementations, the field continues to expand the reach and clinical relevance of mechanical cardiac sensing.