PaperlayerView exact revision on arXiv

Source-linked AI summary

PPG2ABP: Translating Photoplethysmogram (PPG) Signals to Arterial Blood Pressure (ABP) Waveforms using Fully Convolutional Neural Networks

Nabil Ibtehaz, Sakib Mahmud, Muhammad E. H. Chowdhury, Amith Khandakar, Mohamed Arselene Ayari, Anas Tahir, M. Sohel Rahman

arXiv:2005.01669v2eess.SPcs.LG

TL;DR

Continuous blood-pressure monitoring is important, but established measurement methods can be invasive, and prior signal-based approaches may rely on ECG or handcrafted features. PPG2ABP uses cascaded deep-learning networks to estimate continuous ABP waveforms from PPG alone. The authors report that its derived DBP, MAP, and SBP values outperform most existing works, despite not being explicitly trained for those values.

  • Problem

    Continuous blood-pressure monitoring is important, but catheter-based measurement is invasive, cuff-based methods are unsuitable for continuous monitoring, and many prior signal-based approaches require ECG or handcrafted features.

    Source: §1 · pp. 1–2

  • Method

    PPG2ABP uses a cascaded deep-learning pipeline in which an Approximation network estimates ABP waveforms from PPG and a Refinement network refines them.

    Source: §1 · p. 2, §2.2.3 · p. 5

  • Results

    PPG2ABP outperforms most existing works on derived DBP, MAP, and SBP values despite not being explicitly trained for those values.

    Source: §1 · p. 2, §5 · p. 12

  • Takeaways & Limitations

    The approach estimates a complete ABP waveform from PPG alone, rather than limiting output to blood-pressure values such as DBP, SBP, and MAP.

    Source: §1 · p. 2, §5 · p. 12

Abstract

from arXiv · show

Cardiovascular diseases are one of the most severe causes of mortality, taking a heavy toll of lives annually throughout the world. The continuous monitoring of blood pressure seems to be the most viable option, but this demands an invasive process, bringing about several layers of complexities. This motivates us to develop a method to predict the continuous arterial blood pressure (ABP) waveform through a non-invasive approach using photoplethysmogram (PPG) signals. In addition we explore the advantage of deep learning as it would free us from sticking to ideally shaped PPG signals only, by making handcrafted feature computation irrelevant, which is a shortcoming of the existing approaches. Thus, we present, PPG2ABP, a deep learning based method, that manages to predict the continuous ABP waveform from the input PPG signal, with a mean absolute error of 4.604 mmHg, preserving the shape, magnitude and phase in unison. However, the more astounding success of PPG2ABP turns out to be that the computed values of DBP, MAP and SBP from the predicted ABP waveform outperforms the existing works under several metrics, despite that PPG2ABP is not explicitly trained to do so.

1. Introduction

Continuous blood-pressure monitoring is important, but accurate catheter-based methods are invasive and cuff-based methods are unsuitable for continuous monitoring. PPG2ABP uses PPG alone in a cascaded deep-learning pipeline to estimate ABP waveforms, addressing prior reliance on ECG signals and handcrafted features.

1. Introduction

  • Catheter-based measurements are invasive, and cuff-based methods are unsuitable for continuous monitoring, motivating noninvasive, cuffless approaches such as PPG.The paper describes PPG as popular in wearables because it is simple and inexpensive.Source: §1 · p. 1
  • Prior PPG-based blood-pressure methods often require ECG signals or handcrafted features that depend on near-ideal signals, while earlier waveform-reconstruction work used varied deep-learning techniques.The introduction also notes that many prior methods estimate DBP, SBP, and MAP rather than reconstructing the complete ABP waveform.Source: §1 · p. 2
  • PPG2ABP uses cascaded deep-learning 1D segmentation models to estimate ABP waveforms from PPG alone.The Approximation network estimates the waveform, and the Refinement network refines that output.Source: §1 · p. 2
  • The authors report that PPG2ABP outperforms most existing works on DBP, SBP, and MAP despite not being explicitly trained for those values.The method also avoids requiring signals to maintain a standard shape and does not require parallel ECG probes.Source: §1 · p. 2

2. Materials and Methods

PPG2ABP estimates continuous ABP waveforms from PPG segments through a preprocessing stage followed by approximation and refinement networks. The resulting waveform is also used to calculate SBP, DBP, and MAP.

2.1 Dataset

  • The dataset uses simultaneous PPG and ABP signals from MIMIC-III, with DBP as low as 50 mmHg and SBP as high as 200 mmHg.The study selected 127,260 episodes, used 100,000 for training and 27,260 for independent testing, and retained sub-ideal-quality signals.Source: §2.1 · p. 3

2.2 Proposed Methodology

  • PPG2ABP filters PPG segments and passes them through an Approximation Network followed by a Refinement Network to produce ABP waveforms.The pipeline also computes SBP, DBP, and MAP from the estimated waveform.Source: §2.2 · p. 3, Figure 1 · p. 4

2.2.1 Preprocessing

  • Preprocessing applies wavelet denoising, segments and filters PPG signals, and globally min-max normalizes them before training.The wavelet transform uses 10 decomposition levels with Daubechies 8; reconstructed ABP waveforms are denormalized using the earlier global min-max values.Source: §2.2.1 · p. 4

2.2.2 Approximation Network

  • A deeply supervised 1D U-Net approximates ABP from filtered PPG, using skip connections to preserve feature maps and auxiliary losses before decoder upsampling operations.The final convolution uses a linear activation to produce a regression output.Source: §2.2.2 · pp. 4–5

2.2.3 Refinement Network

  • A 1D MultiResUNet refines the approximate waveform, using MultiRes blocks and Res paths to address differences between encoder and decoder feature maps.The MultiRes blocks use factorized convolutions for multiresolution analysis; the Res paths add convolutions along shortcut connections.Source: §2.2.3 · p. 5, Figure 3 · p. 5

2.2.4 BP Parameters Calculation

  • SBP, DBP, and MAP are calculated from the refined ABP waveform using its maximum, minimum, and mean, respectively.The ABP waveform used for these calculations is the Refinement Network estimate.Source: §2.2.4 · p. 5

3. Experiments

The experiments examine model and training choices for the PPG2ABP pipeline and evaluate waveform and blood-pressure estimates using regression and domain-specific metrics.

3. Experiments

  • The authors implemented the experiments in Python using TensorFlow 2.0 with a Keras frontend and made the code publicly available.The experiments include setup choices and ablations intended to select architecture parameters for estimating ABP waveforms from PPG signals.Source: §3 · p. 5

3.1 Selection of Models

  • The selected pairing uses U-Net for approximation and MultiResUNet for refinement; U-Net refinement did not match MultiResUNet, while changing the approximation model left overall performance quite similar.The authors chose U-Net for approximation because it is computationally lighter than MultiResUNet.Source: §3.1 · pp. 5–6

3.2 Selection of Loss Functions

  • MAE in the Approximation Network significantly improved performance over MSE, while the authors used MSE for refinement to diminish larger errors.Their rationale is that equal error weighting suits a rough initial estimate, whereas penalizing larger errors is useful once an approximation is available.Source: §3.2 · p. 6

3.3 Effect of Number of Convolutional Filters

  • Models with more convolutional filters performed better, but their increasing computational cost led the authors to limit U-Net filters to [64, 128, 256, 512, 1024] and MultiResUNet alpha to 2.5.The selected convolution kernel size and stride were 3 and 1, respectively.Source: §3.3 · p. 6

3.4 Effect of Deep Supervision

  • Deep supervision substantially improved U-Net but minimally improved MultiResUNet, so it was used only in the Approximation Network to balance accuracy and computation.Auxiliary loss weights were [1, 0.9, 0.8, 0.7, 0.6], with the largest weight on the final output.Source: §3.4 · p. 6

3.5 Training Methodology

  • The authors used Adam with its default parameters and trained each model for 100 epochs with a patience of 20.MAE and MSE were assigned to the Approximation and Refinement networks, respectively.Source: §3.5 · pp. 6–7

3.6 K-Fold Cross Validation

  • Ten-fold cross-validation on the training data produced 10 models, and the best-performing model was evaluated on the independent test set.Each fold used 90% of the training data for training and 10% for validation.Source: §3.6 · p. 7

3.7 Evaluation Metrics

  • MAE was the primary evaluation metric, supplemented by BHS and AAMI standards, Pearson correlation, and Bland–Altman plots.The authors describe MAE as a commonly used regression loss and evaluation measure.Source: §3.2 · p. 6, §3.7 · p. 7

4. Results

On the independent test set, PPG2ABP reconstructed ABP waveforms from PPG, with a reported waveform MAE of 4.604 ± 5.043 mmHg. Its results met BHS grade A for DBP and MAP and satisfied AAMI requirements for both, while SBP showed greater variability.

4.1 Estimating ABP Waveform

  • 4.604 ± 5.043 mmHg was the overall waveform MAE on the independent test set; the final output preserved shape, magnitude, and phase.For one example, refinement reduced reconstruction error from 9.52 mmHg to 2.37 mmHg.Source: §4.1 · p. 7, Figure 4 · p. 8

4.2 BHS Standard

4.3 AAMI Standard

  • DBP and MAP satisfied the AAMI requirements, but SBP’s standard deviation was higher than the criterion.The AAMI criterion requires mean error below 5 mmHg and standard deviation below 8 mmHg.Source: §4.3 · p. 9, Figure 6 · p. 9, Table 3 · p. 9

4.4 Statistical Analysis

  • For 95% of samples, Bland–Altman errors spanned DBP [-11.825:15.064], MAP [-9.095:10.357], and SBP [-22.531:19.367] mmHg; outliers were especially common for SBP.The authors also report that most errors fell within the 5-mmHg range.Source: Figure 7 · p. 10, §4.4 · p. 10
  • PCC was 0.894 for DBP, 0.966 for MAP, and 0.936 for SBP; each correlation had p < .000001 on 27,260 samples.The authors report that these results reject the null hypothesis of no correlation between predictions and ground truth.Source: Figure 8 · p. 10, §4.4 · p. 10

4.5 Comparison with the Existing Methods

  • The comparison covered methods evaluated on MIMIC-III with sufficiently large patient counts, using DBP, MAP, and SBP MAE plus BHS and AAMI measures.The supplied passages describe the comparison’s scope and metrics but do not provide its comparative outcomes.Source: Table 4 · p. 11, §4.5 · p. 11

5. Conclusion

PPG2ABP infers continuous blood-pressure waveforms from PPG alone, with waveforms that closely resemble the actual signal and derived DBP, SBP, and MAP values that outperform methods tailored to those measures.

5. Conclusion

  • PPG2ABP waveforms retain the actual signal’s shape, amplitude, and phase, while derived DBP, SBP, and MAP values outperform methods explicitly trained to predict them.The paper reports that the method was not explicitly trained to predict these values.Source: §5 · p. 12
  • Earlier approaches often estimated only DBP, SBP, and MAP, sometimes required ECG, and could rely on handcrafted features that required ideal, clean signals.Source: §5 · p. 12
  • Deep learning models adaptively compute high-level features from data, alleviating the need for manual features that may impose signal criteria.Source: §5 · p. 12
  • The complete pressure profile could support continuous patient monitoring, while mapping blood-pressure trends to behavior and activities may yield insights.Source: §5 · p. 12
  • Because smartwatches and fitness bands have PPG sensors, the authors say PPG2ABP applications could reach the mass market without additional expensive sensors such as ECG.Source: §5 · p. 12
  • Future work includes replacing the two-network pipeline with one optimized model, developing wearable applications and conducting clinical studies, and exploring personalized calibration.Source: §5 · p. 12

Related papers

Loading 2005.01669v2…