PaperlayerView exact revision on arXiv

Source-linked AI summary

Cuffless Blood Pressure Estimation from Electrocardiogram and Photoplethysmogram Using Waveform Based ANN-LSTM Network

Md. Sayed Tanveer, Md. Kamrul Hasan

arXiv:1811.02214v1eess.SP

TL;DR

Morphological changes in ECG and PPG waveforms make feature-based blood-pressure estimation difficult. The proposed hierarchical ANN-LSTM model automatically extracts waveform features and models their time variation, satisfying AAMI and BHS evaluation standards with low estimation errors.

  • Problem

    Morphological contour changes in ECG and PPG signals make extracting reliable blood-pressure features difficult.

  • Method

    A hierarchical waveform-based ANN-LSTM model uses ANNs to extract ECG and PPG features and stacked LSTMs to learn their time-domain variations.

  • Results

    The model outperformed conventional feature-engineering methods, satisfied the AAMI standard, and achieved BHS grade A for both SBP and DBP estimation.

  • Takeaways & Limitations

    The model supports continuous blood-pressure estimation without feature engineering and may facilitate continuous estimation in mobile healthcare devices.

  • Takeaways & Limitations

    The authors identify the need to assess robustness using more subjects with greater subject diversity.

Abstract

from arXiv · show

Goal: Although photoplethysmogram (PPG) and electrocardiogram (ECG) signals can be used to estimate blood pressure (BP) by extracting various features, the changes in morphological contours of both PPG and ECG signals due to various diseases of circulatory system and interaction of other physiological systems make the extraction of such features very difficult. Methods: In this work, we propose a waveform-based hierarchical Artificial Neural Network - Long Short Term Memory (ANN-LSTM) model for BP estimation. The model consists of two hierarchy levels, where the lower hierarchy level uses ANNs to extract necessary morphological features from ECG and PPG waveforms and the upper hierarchy level uses LSTM layers to account for the time domain variation of the features extracted by lower hierarchy level. Results: The proposed model is evaluated on 39 subjects using the Association for the Advancement of Medical Instrumentations (AAMI) standard and the British Hypertension Society (BHS) standard. The method satisfies both the standards in the estimation of systolic blood pressure (SBP) and diastolic blood pressure (DBP). For the proposed network, the mean absolute error (MAE) and the root mean square error (RMSE) for SBP estimation are 1.10 and 1.56 mmHg, respectively, and for DBP estimation are 0.58 and 0.85 mmHg, respectively. Conclusion: The performance of the proposed hierarchical ANN-LSTM model is found to be better than the other feature engineering-based networks. It is shown that the proposed model is able to automatically extract the necessary features and their time domain variations to estimate BP reliably in a noninvasive continuous manner. Significance: The method is expected to greatly facilitate the presently available mobile health-care gadgets in continuous BP estimation.

I. INTRODUCTION

Blood pressure requires continuous, noninvasive monitoring because it varies with physiological conditions, while conventional cuff-based measurements are discontinuous and uncomfortable. The paper proposes a hierarchical ANN-LSTM model that automatically extracts ECG and PPG features and learns their temporal variations.

  • Blood pressure variability motivates continuous measurement for accurate diagnosis and treatment of hypertension and hypotension.
  • Cuff-based oscillometric and auscultatory devices measure intermittently and are reported to be uncomfortable, motivating cuffless alternatives.
  • PTT and additional PPG morphological features have been combined with machine learning to improve BP estimation over PTT-only models.
  • Disease, drugs, and external influences make ECG and PPG feature extraction difficult, while existing approaches do not adequately model feature variation over time.
  • The proposed hierarchical ANN-LSTM model uses ANNs for automatic waveform feature extraction and stacked LSTMs for temporal variation, enabling end-to-end BP estimation.

II. BACKGROUND

ECG and PPG provide physiological signals related to pulse timing, vascular properties, and blood pressure. BP estimation can therefore combine PTT with morphological PPG features, although these features depend on waveform contours.

  • ECG records cardiac electrical potential, whereas PPG measures peripheral blood-volume changes optically.
  • PTT measures the timing between the ECG R peak and a selected PPG landmark, including waveform onset, derivative peak, or waveform peak.
  • PWV relates PTT to arterial distance, while vessel elasticity and pressure provide a physiological basis for linking pulse timing to BP.
  • PPG-based predictors include pulse intervals, peak distances, crest time, dicrotic-notch distances, amplitudes, and reflection-related measures.
  • ECG morphology can vary from positive or negative R peaks to irregular bidirectional deflections, complicating reliable landmark detection.

III. PROBLEM INSIGHTS

Manual ECG and PPG feature extraction is unreliable because waveform contours vary across subjects and can be altered by disease, drugs, and other influences. ECG variation particularly complicates accurate R-peak detection for PTT measurement.

  • Automatic feature extraction is motivated by the complexity and frequent errors of manually extracting features from ECG and PPG waveforms.
  • ECG contour depends on coronary blood flow and electrode placement, producing patient-specific waveform differences that complicate R-peak detection.

B. Feature extraction from PPG

PPG contour variation determines whether morphological landmarks, especially dicrotic-notch features, can be extracted reliably. The study describes an end-to-end machine-learning workflow using data from 39 ICU patients with varied clinical conditions and drug exposure.

  • B. Feature extraction from PPG: PPG feature detectability and accuracy depend strongly on waveform contour variation between subjects.
  • B. Feature extraction from PPG: Distinct dicrotic notches permit relatively easy feature extraction, whereas absent notches and faster diastolic decay make extraction difficult or impossible.
  • The workflow covers data collection, preprocessing, and end-to-end machine-learning-based BP estimation.
  • The dataset contains simultaneous ECG, PPG, and ABP recordings from 39 ICU patients sampled at 125 Hz.
  • Patients had diverse clinical conditions, vasoactive-drug exposure, and cardiac disorders that could alter PPG contours and ECG rhythm.

B. Preprocessing

The preprocessing pipeline identifies signal characteristics in windowed ECG and PPG data, selects tunable wavelet parameters, and denoises both signals using the same method. It preserves the fundamental waveform while removing baseline wandering and high-frequency noise.

  • Wavelet denoising with soft Rigrsure thresholding removes baseline wandering and high-frequency noise from both ECG and PPG signals.
  • Fourier analysis normalizes each window and identifies a fundamental peak between 1.0 and 3.5 Hz when its prominence exceeds 0.4.
  • The pipeline determines peak boundaries from local minima, using half the fundamental frequency when no preceding minimum is found.
  • The Q value is selected from a lookup table using the fundamental peak and left-end frequencies, preserving the ECG and PPG waveform amplitude.
  • When no fundamental peak is detected, the method uses Q = 1.08, indicating that the signal may be highly corrupted and unsuitable for extracting some waveform-dependent features.

C. Proposed Method

The proposed method converts paired ECG and PPG waveforms into normalized sequential feature vectors and processes them with a hierarchical ANN-LSTM architecture. Variable-length, peak-based segmentation accommodates changing heart rates without zero padding.

  • Hierarchical Network: The lower ANN hierarchy extracts features from individual concatenated ECG-PPG waveforms, while the upper LSTM hierarchy learns temporal relations among those features.
  • Non-uniform Waveform Segmentation: The method uses variable windows spanning three consecutive ECG or PPG peaks, representing two cycles, then resamples each segment to 256 samples.
  • Non-uniform Waveform Segmentation: PPG segments and corresponding ECG segments are extracted from preprocessed signals, with ECG segments alternatively definable by three R peaks.
  • Non-uniform Waveform Segmentation: Concatenating the resampled ECG and PPG waveforms with normalized segment length produces a 513-feature waveform-based vector.
  • Non-uniform Waveform Segmentation: Consecutive peak triplets are offset by one peak to form a sequence of overlapping feature vectors for the network.
  • Hierarchical Network: The sequence output contains estimated SBP and DBP vectors, with the final sequence output used for continuous prediction.

2) Network Implementation:

The implementation trains the hierarchical ANN-LSTM model on short and long sequences using patient-level data splits and standard neural-network optimization settings. Stacked LSTMs model extracted feature variation across temporal scales.

  • Each patient’s data is split into 70% training, 10% validation, and 20% test subsets.
  • Training uses short sequences with M=10 from 16-second windows and long sequences with M=32 from 40-second windows.
  • The network uses two stacked LSTM layers with time-distributed ANNs, each ANN having one 128-neuron ReLU hidden layer.
  • The LSTM hidden-state size is 128, batch size is 128, mean squared error is the loss, and Adam optimization starts at a learning rate of 0.001.
  • The stacked LSTMs learn extracted-feature variation on different temporal scales, producing sequence outputs whose elements contain estimated SBP and DBP.

V. RESULTS AND DISCUSSION

The proposed model is evaluated with MAE and RMSE and assessed against AAMI and BHS standards. The reported evaluation also includes Bland–Altman, Pearson correlation, and box-plot analyses.

  • MAE and RMSE are the primary metrics used to evaluate the proposed model.
  • Ground-truth and estimated BP values for the final element of each time sequence are used in the metric definitions.
  • Performance is evaluated against AAMI and BHS standards, with additional Bland–Altman, Pearson correlation, and box-plot analyses.

A. Comparison with Other Works

The proposed ANN-LSTM model outperforms the other evaluated neural-network architectures, particularly by modeling temporal feature variations. SBP estimation remains less accurate than DBP estimation because of higher SBP variability.

  • The proposed ANN-LSTM model achieves the best MAE and RMSE among the compared neural-network architectures.LSTM-based models improve accuracy by modeling variations in extracted features.
  • Increasing sequence length from M = 10 to M = 32 improves the prediction capability of LSTM-based models.
  • SBP estimation is less accurate than DBP estimation because the dataset contains high-variance SBP values, including isolated systolic hypertension.

B. Performance Evaluation on AAMI Standards

The model satisfies the AAMI evaluation requirements and achieves strong performance under the BHS standard. Its errors remain within the reported agreement limits for both SBP and DBP.

  • The Bland–Altman limits of agreement are [−3.083, 3.033] mmHg for SBP and [−1.674, 1.653] mmHg for DBP.These limits correspond to the reported short-sequence evaluation.
  • The proposed model satisfies the AAMI standard, with negligible mean error and the minimum standard deviation of error among the evaluated models.Only LSTM-based models satisfy the AAMI requirement for SBP estimation.
  • More than 98% of test samples have estimation errors below 5 mmHg for both SBP and DBP under the BHS evaluation.The paper reports this as the best performance among the compared algorithms.

D. Statistical Analysis

Statistical analyses show close agreement and strong linearity between estimated and target BP values, while continuous tracking is reported to outperform a feature-engineering-based method. The conclusion also notes a need for broader robustness testing.

  • The model’s Pearson correlation coefficients are r = 0.9986 for SBP and r = 0.9975 for DBP.Both coefficients are reported to indicate high linearity between target and estimated BP.
  • The target and estimate box plots show no distinguishable difference, including for very high and very low SBP and DBP values.
  • Continuous SBP and DBP tracking is reported to be more accurate with the proposed model than with the feature-engineering-based method.The comparison includes both short and long sequences.
  • The model’s reported Bland–Altman limits of agreement remain within ±5 mmHg for both SBP and DBP.
  • Future work will investigate robustness using a larger and more diverse subject population.
Loading 1811.02214v1…