Deep-Learning Based ECG-free Heart-Carotid Pulse Transit Time Estimation from Laser Doppler Vibrometry

Technical Summary: Deep-Learning Based ECG-free Heart-Carotid Pulse Transit Time Estimation from Laser Doppler Vibrometry

Problem Statement
Arterial stiffness is a critical biomarker for cardiovascular disease (CVD) risk, often assessed via Pulse Wave Velocity (PWV). PWV is calculated by measuring the Pulse Transit Time (PTT)—the time delay of a pressure wave between two arterial sites. While Laser Doppler Vibrometry (LDV) offers a non-invasive method to measure skin displacement and acceleration at arterial sites (specifically the heart and carotid arteries) to estimate PTT, current automated approaches face significant hurdles. Traditionally, accurate PTT estimation requires simultaneous Electrocardiography (ECG) to identify heartbeat onsets (R-peaks) for signal segmentation. However, integrating ECG electrodes complicates the measurement procedure, increases setup time, and reduces ease-of-use, making it unsuitable for future commercial LDV devices. Existing ECG-free alternatives, such as template matching (TM), rely on rigid analytical models and predefined templates, which are sensitive to signal quality and often fail to process a significant portion of subjects due to strict quality thresholds.

Methodology
The study proposes a Deep Neural Network (DNN) approach to generate an "ECG-proxy" signal directly from LDV data, enabling fully automated PTT estimation without parallel ECG acquisition.

• Data Sources: Two databases were utilized. The CARDIS database (100 hypertensive subjects) provided 20-second recordings used for training, validation, and testing the DNN. The INSIDE database (100 subjects) served as a completely unseen dataset for benchmarking the final PTT estimation performance against a manual reference.
• Signal Processing & Target Generation:
  • LDV signals (skin acceleration) were recorded at the carotid and heart sites using a prototype with six laser channels per handpiece.
  • To train the DNN, raw ECG signals were transformed into a target signal ($E'(n)$). This involved converting R-peaks into a pulse train using Kronecker delta functions and convolving them with a Gaussian filter ($\sigma=75$ samples). This transformation simplified the learning task for Mean Squared Error (MSE) loss while preserving cardiac timing information.
• DNN Architecture: A modified Wave-U-Net architecture was employed, featuring four convolutional layers in both encoder and decoder blocks, batch normalization, ReLU activations, and a transformer layer in the bottleneck for context-aware self-attention. The model maps 4-second LDV segments to the corresponding ECG-proxy signal.
• PTT Estimation Workflow:
  1. Segmentation: The DNN predicts the ECG-proxy ($\hat{E}'(n)$) from the carotid LDV signal. Peak detection is applied to this proxy to identify heartbeat onsets.
  2. Quality Control: A Signal-to-Noise Ratio (SNR) metric ($\zeta$) is calculated based on the energy of the predicted peaks to filter out poor-quality channels. Only channels with high $\zeta$ values are used for segmentation.
  3. Fiducial Point Detection: Once segmented into epochs, fiducial points are identified within physiologically defined windows: the carotid dicrotic notch (DN) and the first perturbation before the minimum of the second heart sound (S2) in the heart signal.
  4. Calculation: PTT is calculated as the time difference between these fiducial points. The final PTT is the median of all valid channel combinations, provided a minimum number of estimates (e.g., 50 over 20 seconds) are met.

Key Contributions

• ECG-Free Segmentation: The primary contribution is the development of a DNN that successfully maps LDV signals to an ECG-proxy, eliminating the need for parallel ECG recordings in clinical LDV setups.
• Robustness to Signal Quality: The study introduces a novel SNR-based metric derived from the DNN's own output to filter channels, allowing the system to handle variable signal quality more effectively than rigid template-matching algorithms.
• Benchmarking: The work provides a comprehensive comparison between three methods: (1) a manual reference using ECG, (2) an automated ECG-based method, (3) a state-of-the-art template-matching (TM) method, and (4) the proposed DNN-based method.

Results

• Heartbeat Detection: The DNN correctly predicted the onset of 68% of heartbeats in the test set and 77% in the validation set. The mean temporal distance between predicted and reference peaks was approximately 4–6 samples (4–6 ms at 1 kHz).
• PTT Correlation: When applied to the unseen INSIDE database (100 subjects):
  • The DNN-based method yielded valid PTT estimates for 64 subjects, achieving a correlation with the manual reference of $\rho = 0.71$ ($p < 0.001$).
  • The Template Matching (TM) method yielded valid estimates for only 46 subjects, with a correlation of $\rho = 0.69$ ($p < 0.001$).
  • The ECG-based automated method (using ground truth ECG for segmentation) achieved $\rho = 0.69$ for 90 subjects.
• Statistical Significance: Wilcoxon signed-rank tests indicated no statistically significant difference in the PTT values produced by the DNN, TM, and ECG-based methods. However, the DNN method demonstrated superior inclusivity, successfully processing a significantly larger fraction of subjects compared to the TM approach.

Significance and Claims
The authors claim that the DNN-based approach provides a viable "ECG-proxy" that enables robust, automated heartbeat segmentation in LDV signals without the logistical burden of ECG electrodes. While the correlation coefficients of the DNN method are comparable to both the template-matching and ECG-based methods, its primary advantage lies in inclusivity: it successfully generates reliable PTT estimates for a larger number of subjects (64 vs. 46 for TM) in the presence of variable signal quality. The study concludes that this technique demonstrates potential for broader biomedical LDV applications beyond heart-carotid PTT, such as carotid-femoral PWV estimation or arrhythmia detection, provided the DNN is retrained on the relevant signal types. The work emphasizes that this advancement is a necessary step toward the clinical deployment of user-friendly, non-invasive arterial stiffness monitoring devices.