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

Nabil Ibtehaz, M. Sohel Rahman+5 more|arXiv (Cornell University)|May 4, 2020

Non-Invasive Vital Sign Monitoring33 citations

TL;DR

PPG2ABP is a two-stage cascaded deep learning pipeline that converts PPG signals into continuous ABP waveforms using a Deep U-Net-based approximation stage followed by a Refinement stage, achieving strong BP metric performance from PPG alone.

ABSTRACT

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.

Motivation & Objective

• Motivation to enable non-invasive, cuff-less, continuous ABP monitoring using widely available PPG signals.
• Eliminate dependence on handcrafted features by leveraging deep learning to learn representations directly from PPG.
• Deliver continuous ABP waveform reconstruction and derive DBP, MAP, and SBP from the estimated waveform.
• Demonstrate performance on large-scale MIMIC-III-derived data and meet standard BP evaluation metrics.

Proposed method

• Preprocess PPG signals with wavelet denoising (db8, 10 levels) and normalization.
• Use a cascaded two-network pipeline: an Approximation Network (1D deeply supervised U-Net) to estimate a rough ABP waveform, followed by a Refinement Network (1D MultiResUNet) to refine it.
• Compute SBP, DBP, and MAP from the estimated ABP waveform using max, min, and mean of ABP respectively.
• Train with MAE loss for the Approximation Network and MSE loss for the Refinement Network, with Adam optimizer for 100 epochs and 10-fold cross-validation.
• Evaluate using MAE and domain metrics (BHS, AAMI), plus Pearson correlation and Bland-Altman analyses.

Experimental results

Research questions

• RQ1 Can ABP waveforms be accurately reconstructed from PPG signals alone without ECG or handcrafted features?
• RQ2 Does a two-stage deep learning pipeline improve ABP waveform reconstruction quality compared to single-stage models?
• RQ3 How well do derived BP values (DBP, MAP, SBP) from the reconstructed ABP waveform meet BHS and AAMI standards on MIMIC-III data?
• RQ4 What is the impact of network choice (U-Net vs. MultiResUNet) and loss design (MAE vs. MSE) on ABP reconstruction accuracy?

Key findings

• ABP waveform MAE: 4.604 ± 5.043 mmHg on the test set.
• DBP MAE: 3.449 ± 6.147 mmHg; MAP MAE: 2.310 ± 4.437 mmHg; SBP MAE: 5.727 ± 9.162 mmHg.
• DBP and MAP meet British Hypertension Society Grade A; SBP reaches Grade B.
• AAMI criteria met for DBP and MAP (SBP standard deviation higher but mean error near zero).
• Pearson correlations: DBP 0.894, MAP 0.966, SBP 0.936 with p < 0.000001.
• Bland-Altman indicate 95% limits of agreement roughly within −12 to 15 mmHg for DBP, −9 to 10 mmHg for MAP, and −23 to 19 mmHg for SBP.