Recover from Horcrux: A Spectrogram Augmentation Method for Cardiac Feature Monitoring from Radar Signal Components

Yuanyuan Zhang, Sijie Xiong, Rui Yan, Eng Gee Lim, Yutao YueOpenReview 2025 · DOI · arXiv · OpenReview

TL;DR

This paper introduces "Horcrux," a novel data augmentation technique for radar-based cardiac monitoring. Horcrux addresses the critical issue of data scarcity in deep learning for vital sign extraction by creating diverse yet faithful spectrograms, significantly improving performance on both classification (heartbeat detection) and regression (ECG reconstruction) tasks.

What Problem Does This Paper Address?

Contactless vital sign monitoring using radar is a promising technology for unobtrusive health tracking. However, the widespread adoption of deep learning in this field is hampered by a lack of sufficient training data. Traditional data augmentation methods, while effective for classification tasks, often fail for regression tasks like reconstructing an ECG signal from radar data. This is because simply altering the input (e.g., stretching, masking) can distort the ground truth vital sign, making the augmented data unfaithful and potentially leading to incorrect learning. The paper aims to bridge this gap by proposing an augmentation method that generates diverse training samples while preserving the integrity of cardiac features for both classification and regression.

Methods

The core of Horcrux lies in its two-pronged approach:

  1. Harmonic and Percussive (H&P) Components Decomposition: The radar signal's spectrogram is decomposed into stable "harmonic" components (related to heart rate) and transient "percussive" components (representing individual heartbeats). This decomposition helps isolate and preserve the essential features of cardiac activity.
  2. Dynamic Template Matching (DTM) and Zero Masking:
    • The DTM module identifies the precise time and frequency characteristics of prominent cardiac vibrations within a given segment.
  3. Using these identified "templates," Horcrux selectively applies a "zero mask" to specific regions of the harmonic spectrogram. This masking is designed to be informative, forcing the deep learning model to focus on subtle, crucial cardiac features rather than being distracted by noise.
  4. Crucially, the percussive component is preserved and added back to the masked harmonic spectrogram. This ensures that time-consistency is maintained and the augmented spectrogram remains a faithful representation of the original signal, preventing the loss of vital information.

Datasets

The study utilizes a public dataset collected using a 77GHz TI AWR-1843 radar. The dataset consists of 91 trails, each containing 3 minutes of radar and ground truth ECG signals sampled at 200Hz. The data was pre-processed to amplify cardiac features and reduce noise. The dataset was split into training, validation, and testing sets with an 80%:10%:10% ratio.

Key Insights and Ideas

  • Augmentation for Regression is Tricky: The paper highlights that standard augmentation techniques designed for classification (where labels are discrete classes) are not directly applicable to regression tasks (like reconstructing a continuous signal) without potentially corrupting the ground truth.
  • Preserving Time-Consistency is Key: For radar-based cardiac monitoring, maintaining the temporal relationships within the signal is paramount. Horcrux's strategy of masking only the harmonic component while preserving the percussive one is a clever way to achieve this.
  • Focusing on Subtle Features: By strategically masking parts of the spectrogram, Horcrux encourages deep learning models to pay closer attention to the subtle, yet critical, cardiac features, leading to improved performance, especially with limited data.
  • H&P Decomposition as a Foundation: Decomposing the signal into harmonic and percussive components provides a robust basis for targeted augmentation, allowing for more intelligent masking.

Key Results

Horcrux demonstrated significant improvements in cardiac monitoring:

  • Overall Improvement: Achieved an overall improvement of 16.20% in cardiac monitoring performance.
  • Best Performance: Outperformed existing augmentation methods when applied to 20% of the dataset, achieving a 16.20% improvement.
  • Ablation Study:
    • H&P decomposition alone improved performance by 8.57%.
  • Random masking on the original spectrogram yielded an 8.42% improvement.
  • DTM-based masking, combined with H&P decomposition, led to the largest improvement of 16.20%.
  • Masking the time domain (13.18%) was more effective than masking the frequency domain (10.69%) when H&P decomposition was used.
  • Dataset Scale: Horcrux showed a more robust performance across different dataset scales, delaying the "tipping point" of performance degradation to 70% of the data compared to 80% for unaugmented data.

Limitations and Future Work

  • Overuse of Horcrux: The paper notes that excessive masking (e.g., 25% and 30% of the dataset) can lead to degradation in RMSE and PCC, suggesting that the learned data representation might drift.
  • Future Work: The authors suggest that Horcrux has the potential to be extended to other spectrogram-based tasks. They also propose further evaluation for downstream tasks such as radar-based cardiovascular disease diagnosis.

Summary updated

Abstract

Radar-based wellness monitoring is becoming an effective measurement to provide accurate vital signs in a contactless manner, but data scarcity retards the related research on deep-learning-based methods. Data augmentation is commonly used to enrich the dataset by modifying the existing data, but most augmentation techniques can only couple with classification tasks. To enable the augmentation for regression tasks, this research proposes a spectrogram augmentation method, Horcrux, for radar-based cardiac feature monitoring (e.g., heartbeat detection, electrocardiogram reconstruction) with both classification and regression tasks involved. The proposed method is designed to increase the diversity of input samples while the augmented spectrogram is still faithful to the original ground truth vital sign. In addition, Horcrux proposes to inject zero values in specific areas to enhance the awareness of the deep learning model on subtle cardiac features, improving the performance for the limited dataset. Experimental result shows that Horcrux achieves an overall improvement of 16.20% in cardiac monitoring and has the potential to be extended to other spectrogram-based tasks. The code will be released upon publication.
BibTeX
@misc{zhang2025recover,
  author = {Yuanyuan Zhang and Sijie Xiong and Rui Yan and Eng Gee Lim and Yutao Yue},
  title = {{Recover from Horcrux: A Spectrogram Augmentation Method for Cardiac Feature Monitoring from Radar Signal Components}},
  year = {2025},
  eprint = {2503.19649},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.19649},
  howpublished = {OpenReview},
}
Figure