Gyrocardiography (GCG) in Cardiac Monitoring

• Gyrocardiography (GCG) is a non-invasive modality that measures the heart’s rotational dynamics via acoustic sensing and deep learning-driven reconstruction.
• The method repurposes consumer hearables by leveraging speaker-as-microphone techniques and rigorous signal conditioning to capture fine-grained cardiac events.
• Innovative attention mechanisms and frequency-domain normalization enable robust, continuous cardiac monitoring without the need for specialized calibration.

Gyrocardiography (GCG) is a modality for non-invasive assessment of the heart’s rotational dynamics via micro-mechanical motion monitoring, historically requiring specialized inertial sensors positioned on the chest. Recent advances employ acoustic sensing through consumer hearable devices, repurposing standard hardware such as speakers and microphones for cardiac activity capture. Modern methodologies map easily accessible heart sounds (“lub-dub,” i.e., S1 and S2) to fine-grained GCG signals using deep learning-driven, cross-modal reconstruction pipelines that allow unobtrusive, continuous cardiac monitoring. This article details the signal capture, processing, reconstruction, normalization, performance metrics, innovations relative to traditional methods, and future implications of GCG as instantiated in the LubDubDecoder system (Zhang et al., 13 Sep 2025).

1. Signal Acquisition via Hearable Devices

Gyrocardiography historically relied on chest-mounted inertial measurement units (IMUs) to directly capture the heart’s 3D rotational movements. The LubDubDecoder paradigm instead utilizes hearable devices’ built-in transducers:

• Speaker-As-Microphone Sensing: The acoustic reciprocity principle is leveraged so that standard speakers (in over-ear headphones or bone-conduction devices) function as microphones, passively detecting low-frequency heart sounds (S1/S2).
• Ear-Anatomy Advantage: Fixed placement in the ear canal or on the ear provides repeatable signal acquisition, contrasting the position-sensitive variability of chest-worn IMUs.

Shared temporal and spectral features between heart sounds and underlying mechanical events enable ear-captured audio to serve as a proxy input for mechanical signal inference.

2. Signal Conditioning and Cardiac Cycle Segmentation

Raw signals from hearable sensors undergo rigorous conditioning:

• Resampling: All streams are resampled to 500 Hz, matching the requirements of high-fidelity cardiac event capture.
• Bandpass Filtering: A 4th-order Butterworth bandpass filter (5–45 Hz) suppresses respiration and environmental noise, isolating cardiac frequencies.
• Cycle Segmentation: Peak-finding algorithms identify the primary S1 peaks for ear-acoustic data and aortic opening (AO) for SCG/GCG, extracting 800 ms windows (200 ms pre-anchor and 600 ms post-anchor) per cycle.

This segmentation aligns input and ground-truth signals for supervised cross-modal reconstruction.

3. Deep Learning-Based Cross-Modal Reconstruction

LubDubDecoder employs an autoencoder architecture optimized for temporal signal transformation:

• Encoder Structure:
  • Local Branch: 1D convolutions with small, dilated kernels extract fine-scale features (micro-mechanical detail).
  • Global Branch: 1D convolutions with larger receptive fields capture periodic and contextual cardiac cycle structure.
• Attention Mechanism: Outputs from both branches are concatenated and processed via a temporal self-attention module, adaptively weighting time segments according to feature relevance.
• Decoder: Attention-enhanced convolutions and upsampling reconstruct the output waveform, mapping coarse heart sounds to detailed GCG and SCG traces.

The network learns to translate time-domain acoustic signals into the time-domain rotational cardiac motion underlying GCG.

4. Frequency-Domain Device Normalization and Cross-Device Adaptation

Device-specific hardware variations in hearables introduce spectral distortions. LubDubDecoder incorporates a zero-effort normalization scheme to align signals across devices:

• Frequency-Domain Mapping:
  • For reference device mean cycle $x_\text{ref}(t)$ and target device mean cycle $x_\text{tgt}(t)$:

  $$X_\text{ref}(f) = \mathcal{F}\{x_\text{ref}(t)\}, \quad X_\text{tgt}(f) = \mathcal{F}\{x_\text{tgt}(t)\}$$

  $$H(f) = \frac{X_\text{ref}(f) X_\text{tgt}^*(f)}{|X_\text{tgt}(f)|^2 + \epsilon}$$ - For each new cycle from the target device: 1. $X_\text{tgt}^{(i)}(f)$ is computed. 2. Apply mapping: $$\hat{X}_\text{tgt→ref}^{(i)}(f) = H(f) \cdot X_\text{tgt}^{(i)}(f)$$. 3. Invert: $$\hat{x}_\text{tgt→ref}^{(i)}(t) = \mathcal{F}^{-1}\{\hat{X}_\text{tgt→ref}^{(i)}(f)\}$$. - Energy normalization: $$\alpha = \frac{||x_\text{ref}(t)||_2}{||\hat{x}_\text{tgt→ref}^{(i)}(t)||_2}$$, final output $$\hat{x}_\text{norm}^{(i)}(t) = \alpha \hat{x}_\text{tgt→ref}^{(i)}(t)$$.

This procedure ensures that the input features across devices are spectrally and energetically standardized, enabling robust cross-device generalization without user calibration.

5. Performance Metrics and Evaluation

The LubDubDecoder system’s efficacy was evaluated in an IRB-approved feasibility paper with 18 users:

Evaluation Condition	GCG Waveform Correlation	Calibration Required
Within-user	$0.95 \pm 0.04$	None
Cross-user (brief calibration)	$0.89 \pm 0.05$	$\approx 4$ seconds
Cross-device (zero-effort adaptation)	$0.91 \pm 0.04$	None

• Timing Accuracy: Fiducial event detection (e.g., AO) median errors: 0–4 ms, with 95th percentile errors up to 4–30 ms; these errors are typically less than 1% of an 800 ms cardiac cycle.
• Robustness: The method is resilient across remounting sessions and during music playback.

This suggests near ground-truth fidelity for GCG waveform recovery via hearables, even in heterogeneous hardware settings.

6. Comparison with Traditional and Alternative Methods

Traditional GCG monitoring requires dedicated IMUs mounted on the chest, with inconsistent placement leading to signal variability. The LubDubDecoder approach:

• Uses repeatable ear placement for more consistent recordings.
• Employs the acoustic reciprocity principle to repurpose commercial hearable hardware for sensing, minimizing the need for specialized equipment.
• Addresses the low SNR challenge of ear-acoustic measurements through attention-based deep learning and frequency-domain normalization.
• Enables device-agnostic deployment and rapid adaptation without user calibration, accelerating scalability in both research and clinical settings.

A plausible implication is that such a methodology allows GCG and other cardiac signals to be unobtrusively monitored with everyday consumer electronics.

7. Prospects and Applications in Cardiac Monitoring

GCG reconstruction from ear-based acoustics via LubDubDecoder enables multiple healthcare applications:

• Continuous Cardiac Health Tracking: Users may monitor cardiac function unobtrusively during daily living or sleep with standard earbuds or hearing aids.
• Early Pathology Detection: Micro-mechanical metrics accessible with GCG may facilitate earlier detection of heart failure, ischemia, or arrhythmias, with potential sensitivity advantages over traditional ECG in specific contexts.
• Scalable Screening: Zero-effort device adaptation supports population-scale, economic cardiac screening in diverse environments—including clinics, homes, and remote settings.
• Emergency Alerts: Integration with consumer electronics may provide early warning systems for acute cardiac events through real-time mechanical signature tracking.

Overall, mapping the “lub-dub” acoustic signature to gyrocardiographic motion via hearables establishes an extensible platform for personalized, continuous cardiac monitoring, promising substantial contributions to preventive healthcare and chronic disease management.