Cuffless and Continuous Blood Pressure Estimation from the Heart Sound Signals

Rong-Chao Peng, Wen-Rong Yan, Ning-Ling Zhang, Wan-Hua Lin, Xiao-Lin Zhou, Yuan-Ting ZhangSensors 2015 · DOI

TL;DR

This paper presents a novel, cuffless, and continuous blood pressure estimation method using only a smartphone and a stethoscope. By analyzing the Fourier spectrum of the second heart sound (S2), the researchers developed a support vector machine model that can predict blood pressure with promising accuracy, offering a low-cost and convenient alternative to traditional bulky devices.

What Problem Does This Paper Address?

Cardiovascular diseases, particularly hypertension, are major global health concerns. Early detection and continuous monitoring of blood pressure are crucial for managing these conditions. However, traditional blood pressure measurement devices (sphygmomanometers) are often bulky, expensive, require trained personnel, and are not ideal for home healthcare or continuous monitoring. The increasing ubiquity of smartphones presents an opportunity to develop accessible, low-cost, and user-friendly solutions for physiological monitoring. While smartphones have been explored for other vital signs, cuffless and continuous blood pressure estimation remains a significant challenge.

Methods

The core of this research lies in leveraging the relationship between the second heart sound (S2) and blood pressure. The authors propose that the spectral characteristics of S2 change with blood pressure. Their methodology involves:

  • Data Acquisition: Heart sounds were recorded using a smartphone's microphone integrated into a stethoscope, placed on the chest of 32 healthy subjects. Simultaneously, continuous blood pressure was measured using a Finometer device. A cold-pressor experiment was conducted to induce variations in blood pressure.
  • Signal Processing: Heart sound signals were pre-processed with low-pass and high-pass filters, followed by decimation. The Shannon energy envelope was then calculated to identify potential peaks corresponding to heart sounds.
  • S2 Identification: Thresholding techniques were applied to the Shannon energy envelope to identify the first (S1) and second (S2) heart sounds. In cases of signal disturbance, manual adjustments were made.
  • Feature Extraction: For each identified S2, a 64 ms window was extracted, and its frequency spectrum was computed using Fast Fourier Transform (FFT). The spectrum was normalized, and 36 spectral values in the 50-400 Hz range were selected as features.
  • Blood Pressure Estimation: A Support Vector Machine (SVM) regression model was trained to map the extracted S2 features to systolic, diastolic, and mean blood pressure values. The RBF kernel was chosen for its good performance in similar applications.
  • Validation: A 10-fold cross-validation approach was used to evaluate the accuracy of the regression model. Statistical metrics like Pearson correlation coefficient (CC), mean absolute error (MAE), mean error (ME), and standard deviation (SD) were calculated.

Datasets

The study utilized a dataset collected from 32 healthy subjects (25 males, 7 females, aged 20-32 years) during a cold-pressor experiment. This experiment involved inducing blood pressure variations through cold water immersion. No pre-existing public datasets were used for training or evaluation; the data was collected specifically for this study.

Key Insights and Ideas

  • S2 as a Biomarker for BP: The study strongly suggests that the spectral characteristics of the second heart sound (S2) are significantly correlated with blood pressure. This is a key insight, as S2 is often overlooked for direct blood pressure estimation compared to pulse wave analysis.
  • Smartphone as a Medical Device: The research demonstrates the feasibility of using a standard smartphone, augmented with a simple stethoscope, as a capable tool for medical monitoring, specifically for blood pressure. This democratizes access to health monitoring.
  • Cuffless and Continuous Potential: The method's ability to estimate blood pressure continuously and without a cuff is a major advancement, addressing the limitations of current clinical practices and enabling more natural, long-term monitoring.
  • SVM's Efficacy: The successful application of SVM regression highlights its power in finding complex, non-linear relationships between physiological signals and blood pressure.

Key Results

  • Correlation Coefficients: The mean Pearson correlation coefficients (CC) between predicted and measured blood pressure values were:
  • Systolic Blood Pressure (SBP): 0.707
    • Diastolic Blood Pressure (DBP): 0.712
    • Mean Blood Pressure (MBP): 0.748
  • Mean Errors: The mean absolute errors (MAE) were:
  • SBP: 4.339 mmHg
    • DBP: 3.171 mmHg
    • MBP: 3.480 mmHg
  • Mean Errors (ME): The mean errors (ME) were very close to zero (-0.204 mmHg for SBP, -0.274 mmHg for DBP, and -0.357 mmHg for MBP), indicating minimal systematic bias.
  • Standard Deviations (SD): The mean standard deviations were:
  • SBP: 6.121 mmHg
    • DBP: 4.471 mmHg
    • MBP: 4.961 mmHg
  • These results generally met the American Association for the Advancement of Medical Instrumentation (AAMI) guidelines for ME (±5 mmHg) and SD (8 mmHg) for most parameters, suggesting sufficient accuracy.

Limitations and Future Work

  • Training Phase Complexity: The current method requires a training phase to establish the personalized model. The cold-pressor test used to induce blood pressure variation is not suitable for cardiovascular patients. Future work should focus on developing safer and more convenient methods for data collection, such as post-exercise measurements or using hand height variations. Alternatively, developing generalized models for age/gender groups or large clinical datasets could reduce or eliminate the need for individual calibration.
  • Real-time Implementation: The current SVM algorithm runs offline on a personal computer. Implementing a real-time SVM algorithm on the limited resources of a smartphone is a crucial next step for practical, continuous monitoring.
  • Environmental Noise Interference: Heart sounds are weak and susceptible to environmental noise. Future research should explore hardware improvements (e.g., specialized microphone designs) and software solutions (e.g., adaptive filters, robust segmentation algorithms) to enhance signal quality and S2 detection in noisy environments.
  • Generalizability: While the results are promising, the study was conducted on healthy subjects. Further validation on diverse populations, including those with cardiovascular conditions, is necessary.

Summary updated

BibTeX
@article{peng2015cuffless,
  author = {Rong-Chao Peng and Wen-Rong Yan and Ning-Ling Zhang and Wan-Hua Lin and Xiao-Lin Zhou and Yuan-Ting Zhang},
  title = {{Cuffless and Continuous Blood Pressure Estimation from the Heart Sound Signals}},
  journal = {Sensors},
  year = {2015},
  doi = {10.3390/s150923653},
}
Figure