Source-linked AI summary
Estimating Blood Pressure from Photoplethysmogram Signal and Demographic Features using Machine Learning Techniques
Moajjem Hossain Chowdhury, Md Nazmul Islam Shuzan, Muhammad E. H. Chowdhury, Zaid B Mahbub, M. Monir Uddin, Amith Khandakar, Mamun Bin Ibne Reaz
TL;DR
The paper addresses the need for continuous, non-invasive BP monitoring because cuff-based measurements are discrete and uncomfortable. It combines PPG and demographic features with preprocessing, feature selection, and machine-learning regression, with optimized models achieving strong SBP and DBP estimation performance.
Problem
Cuff-based BP measurements are discrete and uncomfortable, while continuous invasive monitoring exposes patients to infection risk.
Method
The study preprocesses PPG signals, extracts multi-domain features and demographic characteristics, applies feature selection, and trains machine-learning regression models for SBP and DBP.
Results
After optimization, ReliefF with GPR achieved R scores of 0.95 for SBP and 0.96 for DBP estimation.
Takeaways & Limitations
The study demonstrates non-invasive BP estimation from PPG features without cuff-based pressure measurement.
Takeaways & Limitations
PPG measurement faces challenges including noise reduction and creation of multiple photodetectors.
Abstract
from arXiv · showhide
Hypertension is a potentially unsafe health ailment, which can be indicated directly from the Blood pressure (BP). Hypertension always leads to other health complications. Continuous monitoring of BP is very important; however, cuff-based BP measurements are discrete and uncomfortable to the user. To address this need, a cuff-less, continuous and a non-invasive BP measurement system is proposed using Photoplethysmogram (PPG) signal and demographic features using machine learning (ML) algorithms. PPG signals were acquired from 219 subjects, which undergo pre-processing and feature extraction steps. Time, frequency and time-frequency domain features were extracted from the PPG and their derivative signals. Feature selection techniques were used to reduce the computational complexity and to decrease the chance of over-fitting the ML algorithms. The features were then used to train and evaluate ML algorithms. The best regression models were selected for Systolic BP (SBP) and Diastolic BP (DBP) estimation individually. Gaussian Process Regression (GPR) along with ReliefF feature selection algorithm outperforms other algorithms in estimating SBP and DBP with a root-mean-square error (RMSE) of 6.74 and 3.59 respectively. This ML model can be implemented in hardware systems to continuously monitor BP and avoid any critical health conditions due to sudden changes.
1. Introduction
The paper motivates cuff-less, continuous BP monitoring because cuff measurements are discrete and uncomfortable, while invasive monitoring carries infection risk. It proposes combining PPG-derived features, demographic characteristics, and machine learning to improve BP estimation.
- Monitoring need: Continuous BP monitoring is useful because abnormal nighttime dipping may indicate elevated cardiovascular risk.The paper contrasts overnight monitoring needs with routine once- or twice-daily measurements.
- Limitations of current methods: Cuff-based measurements are intermittent and uncomfortable, whereas invasive arterial monitoring exposes patients to infection risk.The cuff method also requires the arm to remain steady and can disturb sleep during inflation and deflation.
- PPG basis: PPG is a low-cost optical technology whose waveform typically contains systolic peak, diastolic peak, and a notch.PPG records light absorbed or reflected by blood vessels and can support cardiovascular measurements including BP estimation.
- Research gap: Prior work used time-, frequency-, time-frequency-, and statistical-domain PPG features, but the paper identifies limited prior combination of these domains for BP estimation.The study also incorporates demographic characteristics such as height, weight, and age.
- Proposed direction: The proposed analysis combines PPG signal features with demographic characteristics to estimate SBP and DBP using machine learning.The paper presents this combination as an approach to reduce BP-estimation error.
2. Materials and Methods
The method uses quality-screened PPG data, preprocessing, feature extraction, feature selection, and machine-learning regression to estimate SBP and DBP. Data are split into training/validation and testing sets before model evaluation.
- Method overview: The study describes its methodology as including the dataset, preprocessing, extracted features, feature-selection techniques, and trained machine-learning models for SBP and DBP estimation.
- Processing pipeline: PPG signals were assessed for quality, randomly split into 85% training and validation and 15% testing, then preprocessed before feature extraction.Feature selection was applied before machine-learning training to reduce computational complexity and over-fitting risk.
- Dataset: The dataset contained 657 PPG samples from 219 subjects, recorded at 1000 Hz with 2100 data points per 2.1-second signal.Age, gender, height, weight, SBP, DBP, and heart rate were also recorded.
- Dataset: Signal-quality screening retained 222 signals from 126 subjects for feature extraction.Signals lacking prominent waveform features, clear diastolic portions, or sufficient length were excluded.
2.2 Preprocessing Signals
PPG signals undergo normalization, filtration, and baseline correction before feature extraction. These steps limit amplitude, remove high-frequency noise, and reduce baseline wandering while preserving waveform information.
- Pipeline: The preprocessing pipeline prepares raw PPG signals for feature extraction.The stages are introduced as preprocessing steps preceding feature extraction.
- Normalization: Z-score normalization produces amplitude-limited PPG signals and facilitates subsequent preprocessing.The study illustrates signals before and after normalization.
- Signal Filtration: Low-pass filtering removes high-frequency noise from the database signals.Moving-average, FIR, and Butterworth IIR zero-phase filters were tested, with the Butterworth filter selected.
- Baseline Correction: Polynomial detrending corrects baseline wandering by estimating the signal trend and subtracting it.The waveform’s baseline wandering is associated with respiration at 0.15–0.5 Hz.
2.3 Feature Extraction
The study extracts time-, frequency-, statistical-, derivative-, and demographic features from PPG signals. It identifies waveform landmarks such as the dicrotic notch and foot to build a broad feature set.
- Feature Extraction: PPG feature extraction uses waveform landmarks, derivatives, thresholds, and frequency-domain representations.The extracted landmarks include systole, diastole, notch, pulse width, and peak-to-peak interval.
- Dicrotic Notch: The dicrotic notch is detected by subtracting a line between systolic and diastolic peaks, then refining the minimum within a 50ms window.The refinement uses a Fix index to improve robustness across different situations and age groups.
- Foot Detection: The PPG foot is located from the highest APG point inside a zone where moving-average APG exceeds an adaptive threshold.APG denotes the second derivative of the PPG waveform, and the method handles prominent and flat feet.
- Feature Set: The study derives 107 features for each PPG signal: seventy-five time-domain, sixteen frequency-domain, ten statistical, and six demographic features.Derivative-based features and demographic variables are included in the extracted feature set.
- Demographic Features: The demographic features include height, weight, gender, age, BMI, and heart rate.These six variables are listed as demographic features.
2.4 Feature Selection
Feature selection reduces the feature set to lower computational complexity and the risk of over-fitting. The study evaluates correlation-based selection, ReliefF, and FSCMRMR methods and retains features producing the best results.
- Selection Methods: Three feature-selection methods are evaluated: CFS, ReliefF, and FSCMRMR.ReliefF updates feature weights using randomly selected instances and nearest neighbors.
- Selection Methods: CFS evaluates class correlation while FSCMRMR seeks mutually dissimilar features that remain relevant to the response variable.FSCMRMR minimizes feature-set inconsistency and maximizes relevance.
- Selected Features: The selected features are those reported to produce the best results.They are listed in Table 9.
- Selected Features: The listed selected features include demographic, time-domain, frequency-domain, and statistical variables.Examples include weight, age, heart rate, spectral entropy, skewness, kurtosis, and width-based measures.
2.5 Machine Learning (ML) Algorithms
The study trains multiple regression algorithms to estimate blood pressure from the extracted feature matrix. Gaussian Process Regression and Ensemble Trees are retained as the two best-performing algorithm families for further testing.
- Model Training: Nineteen algorithm variants spanning linear regression, regression trees, SVR, GPR, and Ensemble Trees are trained using 10-fold cross-validation.The models are trained through MATLAB 2019b’s Regression Learner App.
- Model Selection: Gaussian Process Regression and Ensemble Trees are selected as the two best-performing algorithms for testing.The selection follows evaluation of the broader set of trained algorithms.
- Gaussian Process Regression: GPR is described as a non-parametric Bayesian regression approach that can work well on small datasets and provide prediction uncertainty measures.The Bayesian approach infers a probability distribution over possible function values.
- Ensemble Trees: An Ensemble Tree combines multiple regression trees through a weighted combination to form a stronger predictive model.Its stated core idea is combining weak learners.
2.6 Hyper‐parameters Optimization of the Best Performing Algorithm
The study optimized machine-learning hyper-parameters to improve algorithm performance beyond default settings.
- Hyper-parameter optimization was applied because default parameters may not provide the best algorithm performance.The optimization used MATLAB 2019b Regression Learner App.
2.7 Evaluation Criteria
The study evaluated blood-pressure estimation using four criteria that quantify prediction error and agreement with ground truth. RMSE was selected as the primary criterion.
- Four criteria were used to evaluate machine-learning algorithms for blood-pressure estimation.The criteria were MAE, MSE, RMSE, and correlation coefficient R.
- Mean Absolute Error is the mean of all absolute prediction errors.
- Mean Squared Error calculates the squared errors and reflects both estimator variance and bias.
- Root Mean Squared Error measures the spread of residual prediction errors around the regression line.
- Correlation coefficient R measures the relationship between predictors and predictions and their proximity to the trendline.
- RMSE was chosen as the main evaluation criterion among the automatically calculated measures.
3. Results and Discussion
The results identify GPR, particularly with ReliefF feature selection and optimization, as the strongest approach for SBP and DBP estimation. The reported performance was also assessed against AAMI and BHS standards, with a limitation in SBP standard deviation.
- Algorithm performance: GPR and Ensemble Trees outperformed other algorithms across SBP and DBP estimation cases.Features from Table 5, together with demographic features, contributed significantly to estimation.
- Feature selection: ReliefF with GPR performed best for SBP, while CFS with GPR performed best for DBP before optimization.The corresponding R scores were 0.74 for SBP and 0.68 for DBP.
- Hyper-parameter optimization: 30 iterations tuned the model hyper-parameters through Bayesian optimization by updating the lowest observed MSE.
- Optimized results: 0.95/0.96 was the post-optimization R score for SBP and DBP estimation using ReliefF with GPR.Table 11 reports this combination as outperforming the other algorithms after optimization.
- Comparison with related work: The proposed results are compared with related work across dataset, methodology, MAE, MSE, RMSE, and R.The supplied table references include dataset and methodology comparisons, but no row-level values are provided here.
- Comparison with related work: The study reports low error using time-domain, frequency-domain, and statistical features in an optimized feature-reduced regression model.The authors state that extracting all these features with classical machine learning was not previously reported to their knowledge.
- Standards-based evaluation: The AAMI criterion accepts estimated blood pressure when it is less than 10 mmHg from actual values.The GPR results were completely accepted for DBP, while SBP standard deviation exceeded the maximum permissible range despite an acceptable mean.
- Standards-based evaluation: The GPR model achieved BHS grade B for both SBP and DBP estimation.BHS grades use cumulative percentages of readings within 5, 10, and 15 mmHg of the mercury standard.
4. Conclusions
The study demonstrates non-invasive SBP and DBP estimation from PPG features and machine learning, using feature reduction and separately trained models. ReliefF with GPR produced the best results, while the DBP estimator met the AAMI requirement and the SBP estimator narrowly missed its standard-deviation requirement.
- ReliefF feature selection combined with Gaussian Process Regression produced the best SBP and DBP estimation results.The study also evaluated Ensemble Trees as a promising alternative before further hyper-parameter optimization.
- The system extracted 107 time-domain, frequency-domain, statistical, and demographic features after PPG preprocessing and feature reduction.The processed waveform had a high signal-to-noise ratio and was free from baseline wandering.
- SBP and DBP models were trained separately because their important features often differed.
- Hyper-parameter optimization produced R scores of 0.95 for SBP and 0.96 for DBP.
- The DBP estimator fulfilled the AAMI standard, whereas the SBP estimator met the mean requirement but narrowly missed the standard-deviation requirement.