Non-Invasive Blood Pressure Estimation from ECG Using Machine Learning Techniques

Monika Simjanoska, Martin Gjoreski, Matjaz Gams, Ana Madevska Bogdanova

TL;DR

This paper presents a novel method for estimating blood pressure (BP) using only electrocardiogram (ECG) signals, eliminating the need for traditional cuff-based devices or additional sensors like PPG. By employing complexity analysis for feature extraction and a stacking machine learning approach, the system achieves promising accuracy, especially when combined with a subject-specific calibration technique.

What Problem Does This Paper Address?

Accurate and continuous blood pressure monitoring is crucial for managing cardiovascular health, but current methods often rely on cumbersome cuff-based devices or require multiple sensors. The proliferation of wearable ECG monitors presents an opportunity to leverage this readily available data for BP estimation. However, directly deriving BP from ECG signals is challenging due to the indirect relationship and the influence of various physiological factors. Existing research often combines ECG with other signals (like PPG) or focuses on morphological ECG features, which can be unreliable for BP prediction. This paper aims to bridge this gap by developing a non-invasive BP estimation method using only ECG signals, focusing on signal complexity rather than morphology.

Methods

The core of the proposed method involves several key steps:

  • Signal Preprocessing: Raw ECG signals are segmented into 30-second intervals and filtered using a band-pass Butterworth filter (0.3 Hz to 50 Hz) to remove baseline wander and preserve relevant ECG information.
  • Feature Extraction via Complexity Analysis: Instead of traditional morphological features, the authors extract seven features based on signal complexity. These include:
  • Signal Mobility and Complexity: Derived using the Hjorth parameters method, these quantify signal variation and complexity.
  • Fractal Dimension: Calculated using the Higuchi algorithm, this measures signal self-similarity.
  • Entropy: Quantifies the randomness of the signal.
  • Autocorrelation: Measures the similarity of the signal with its shifted version.
  • Age: The participant's age is included as a feature, acknowledging its influence on arterial properties.
  • Machine Learning Model: A two-stage machine learning approach is employed:
  • Stacking Classification: Seven different classifiers (KNN, J48, Naive Bayes, SVM, Random Forest, Bagging, Boosting) are used to predict the BP class (normal, prehypertension, hypertension). The outputs of these classifiers are then fed into a meta-classifier (Random Forest) to produce prediction probabilities.
  • Regression: The output of the classification module, along with the extracted complexity features and the predicted BP class, are used as input for three distinct Random Forest regression models to predict Systolic BP (SBP), Diastolic BP (DBP), and Mean Arterial Pressure (MAP).
  • Probability Distribution-Based Calibration: A crucial step to improve accuracy involves calibrating the predictions. This is achieved by analyzing the probability distributions of the errors from the validation set and using these distributions to adjust the predicted BP values for individual subjects.

Datasets

The study utilized data from four distinct sources:

  1. Cooking hacks 3-lead ECG sensor: Recordings from 16 participants (ages 16-72).
  2. 180 eMotion FAROS 3-lead ECG sensor: Recordings from 3 participants (ages 25-27).
  3. Zephyr Bioharness single-lead ECG biosensor: Recordings from 14 participants (ages 20-73), plus 11 additional patients from a hospital.
  4. Charis Physionet database: ECG and arterial BP recordings from 7 patients (ages 20-74) with traumatic brain injuries.

In total, 3129 30-second ECG segments were obtained from 51 participants. The data was randomly split into training (60%), validation (10%), and testing (30%) sets.

Key Insights and Ideas

  • Complexity over Morphology: The paper's central hypothesis is that signal complexity, rather than specific waveform shapes, holds valuable information about BP. This is a significant departure from many existing ECG-based BP estimation methods.
  • ECG-Only Approach: Demonstrating that reliable BP estimation can be achieved using only ECG signals is a major step towards simplifying wearable health monitoring.
  • Stacking for Robustness: Employing a stacking ensemble of classifiers for BP class prediction before regression likely improves the robustness and accuracy of the overall system by leveraging diverse learning algorithms.
  • Calibration is Key: The significant improvement in accuracy after applying probability distribution-based calibration highlights its importance for achieving clinically relevant results, especially for personalized BP estimation.

Key Results

  • Without Calibration: The proposed method achieved a Mean Absolute Error (MAE) of 8.64 mmHg for SBP, 18.20 mmHg for DBP, and 13.52 mmHg for MAP on the testing set.
  • With Calibration: After applying probability distribution-based calibration, the MAE significantly decreased to 7.72 mmHg for SBP, 9.45 mmHg for DBP, and 8.13 mmHg for MAP.
  • Comparison to Medical Devices: The calibrated results are noted to be close to those of certified medical devices for BP estimation, which typically aim for ±5 mmHg error and SD within 8 mmHg.
  • Feature Importance: Analysis suggests that features like Mobility, Complexity, and Entropy, derived from complexity analysis, show discriminatory power across different BP classes.

Limitations and Future Work

  • Sedentary Conditions: The current method is primarily validated for sedentary conditions. For activity monitoring, an activity recognition module would be necessary.
  • Larger Study Needed: While the study involved 51 participants, the authors suggest that a larger study with hundreds of diverse participants would be beneficial for more robust testing.
  • Context-Aware Estimation: Future work could explore context-based BP estimation, potentially integrating information from other sensors or environmental factors.
  • Real-time Performance: The prediction time is very fast (0.0001 s), making it suitable for real-time applications once the model is built. The complexity analysis and model building times are also relatively short.

Summary updated

Abstract

Background: Blood pressure (BP) measurements have been used widely in clinical and private environments. Recently, the use of ECG monitors has proliferated; however, they are not enabled with BP estimation. We have developed a method for BP estimation using only electrocardiogram (ECG) signals. Methods: Raw ECG data are filtered and segmented, and, following this, a complexity analysis is performed for feature extraction. Then, a machine-learning method is applied, combining a stacking-based classification module and a regression module for building systolic BP (SBP), diastolic BP (DBP), and mean arterial pressure (MAP) predictive models. In addition, the method allows a probability distribution-based calibration to adapt the models to a particular user. Results: Using ECG recordings from 51 different subjects, 3129 30-s ECG segments are constructed, and seven features are extracted. Using a train-validation-test evaluation, the method achieves a mean absolute error (MAE) of 8.64 mmHg for SBP, 18.20 mmHg for DBP, and 13.52 mmHg for the MAP prediction. When models are calibrated, the MAE decreases to 7.72 mmHg for SBP, 9.45 mmHg for DBP and 8.13 mmHg for MAP. Conclusion: The experimental results indicate that, when a probability distribution-based calibration is used, the proposed method can achieve results close to those of a certified medical device for BP estimation.
BibTeX
@inproceedings{simjanoska2018non,
  author = {Monika Simjanoska and Martin Gjoreski and Matjaz Gams and Ana Madevska Bogdanova},
  title = {{Non-Invasive Blood Pressure Estimation from ECG Using Machine Learning Techniques}},
  booktitle = {Sensors},
  year = {2018},
  url = {https://openreview.net/forum?id=mhfVSyeZYH},
}
Figure