All research

Physiological signal processing · 2022–2023

Cuffless blood pressure and arterial waveform estimation from PPG and ECG

Deep learning methods that estimate systolic and diastolic pressure and reconstruct continuous arterial blood pressure waveforms from PPG and ECG signals.

Qatar National Research Fund NPRP12S-0227-190164 · Qatar University IRCC-2021-001 and QUST-2-CENG-2022-711

Flowchart. One to four input channels, drawn from PPG, its derivatives and ECG, train a 1D segmentation network that predicts a normalised arterial pressure waveform. Systolic and diastolic values from a separate prediction pipeline rescale the waveform before evaluation.
Figure 1 from Mahmud et al. (2023). The combined pipeline uses a segmentation network for the waveform shape and predicted systolic and diastolic values to restore its scale.View full-size figureSource article© 2022 Elsevier Ltd · author reuse

Approach

Can non-invasive PPG and ECG signals provide accurate blood pressure values and a continuous arterial pressure waveform?

Cuffs provide intermittent readings, while a continuous arterial pressure waveform usually requires an arterial catheter. This project used PPG and ECG to estimate blood pressure without either limitation. The work first predicted systolic and diastolic values, then reconstructed the waveform from PPG alone, and finally combined pressure prediction with waveform reconstruction.

  1. Prepare PPG and ECG
  2. Predict pressure values
  3. Reconstruct the waveform

Predicting systolic and diastolic pressure with a shallow U-Net

Sensors · 2022

The study used the UCI cuffless blood pressure dataset, with 12,000 records from 942 intensive-care patients containing simultaneous PPG, ECG and arterial pressure. Signals were split into 8.192-second segments and distorted segments were removed. Three parts of the dataset formed the training data, and the fourth part provided 53,043 independent test segments.

A U-Net learned to map PPG, its two derivatives and ECG to the arterial pressure waveform. Its bottleneck supplied 1,024 learned features to a multilayer perceptron, which predicted systolic and diastolic pressure. Performance fell as the encoder became deeper, so the final model used one level and about 0.55 million parameters.

Using all four inputs gave the lowest error. PPG with its two derivatives performed close to PPG with ECG, suggesting a simpler setup without ECG electrodes. The model met the British Hypertension Society (BHS) Grade A and AAMI criteria on the test set, and it also kept low errors on an external dataset of 40 people.

Systolic mean absolute error
2.333 mmHg
Diastolic mean absolute error
0.713 mmHg
BHS grade for both pressures
Grade A
Mean absolute error in mmHg, from Table 6 and the external validation in Mahmud et al. (2022).
SystolicDiastolic
PPG only4.9711.361
PPG and ECG2.5130.825
PPG and two derivatives2.7390.960
All four inputs2.3330.713
External dataset of 40 people2.7281.166
Pipeline diagram. One to four input channels train a shallow U-Net autoencoder. Features from its dense bottleneck train a machine learning regressor that predicts systolic and diastolic pressure for evaluation.
Figure 4 from Mahmud et al. (2022). Features learned by a shallow U-Net are passed to a regressor that predicts blood pressure.View full-size figureSource articleCC BY 4.0
Two regression plots of estimated against target pressure. Diastolic and systolic points both lie in a narrow band along the identity line.
Figure 7a from Mahmud et al. (2022). Predicted against measured diastolic (left) and systolic (right) pressure.View full-size figureSource articleCC BY 4.0

Translating PPG alone into arterial pressure waveforms

Bioengineering · 2022

PPG2ABP reconstructs the continuous arterial pressure waveform from PPG without ECG. A deeply supervised 1D U-Net first approximates the waveform, and a 1D MultiResUNet then refines it. Systolic, diastolic and mean pressure are read directly from the reconstructed signal, so the network is not trained on those values separately.

The model was evaluated on 127,260 segments from the UCI data, with an independent test part and lower-quality PPG kept in the data. Diastolic and mean pressure met BHS Grade A and the AAMI criteria. Systolic pressure reached Grade B, and its error spread was above the AAMI limit.

Waveform mean absolute error
4.604 mmHg
Mean arterial pressure error
2.310 mmHg
Input signal
PPG only
Pressure values derived from the reconstructed waveform, from Tables 2 and 3 of Ibtehaz et al. (2022).
Mean absolute errorWithin 5 / 10 / 15 mmHgBHS gradeAAMI
Diastolic3.449 mmHg82.8% / 92.2% / 95.7%AMet
Mean arterial2.310 mmHg87.4% / 95.2% / 97.7%AMet
Systolic5.727 mmHg70.8% / 85.3% / 90.9%BNot met
Four-step diagram: input PPG signal, preprocessed PPG signal, approximate arterial pressure waveform and refined arterial pressure waveform.
Figure 1 from Ibtehaz et al. (2022). The two-stage PPG2ABP pipeline.View full-size figureSource articleCC BY 4.0
Five stacked plots over about eight seconds: input PPG, the approximate network output, the refinement network output, the measured arterial pressure, and an overlay showing the refined waveform closer to the measured signal.
Figure 4 from Ibtehaz et al. (2022). An example of the input PPG, both network outputs and the measured arterial pressure.View full-size figureSource articleCC BY 4.0

Combining pressure prediction with waveform reconstruction

Biomedical Signal Processing and Control · 2023

Segmentation networks learn waveform shape best from normalised signals, but normalisation removes absolute pressure. This study separated the two tasks. The shallow U-Net pipeline predicted systolic and diastolic pressure, a new network estimated the normalised waveform, and a linear transformation restored its scale.

NABNet extends UNet++ with attention-guided bidirectional ConvLSTM blocks in its skip connections, which help it handle timing shifts between PPG, ECG and arterial pressure. It was evaluated with four-fold cross-validation on 191,198 segments, or about 435 hours of signals, against ten other segmentation networks.

The final waveforms kept the prediction accuracy of the first stage, with about 92% of systolic and 99% of diastolic values within 5 mmHg. Compared with PPG2ABP, the pipeline reduced systolic and diastolic error and gave a similar mean pressure error, although it also used ECG.

Systolic mean absolute error
2.63 mmHg
Diastolic mean absolute error
1.09 mmHg
Mean arterial pressure error
2.37 mmHg
Selected segmentation networks from Table 3 of Mahmud et al. (2023). Lower waveform error and a higher share within 5 mmHg are better.
Final waveform errorMean pressure within 5 mmHg
NABNet4.77887.58%
Attention UNet4.78486.46%
BCDUNet4.80486.83%
UNet++4.82685.71%
UNet4.93085.01%
Nested network diagram with encoder, middle and decoder nodes linked by attention-guided LSTM blocks, direct skip connections and deep supervision outputs.
Figure 2 from Mahmud et al. (2023). The NABNet architecture.View full-size figureSource article© 2022 Elsevier Ltd · author reuse
Three stacked signals over about 1,000 samples: a normalised PPG, a normalised ECG, and arterial blood pressure between about 75 and 115 mmHg. The estimated pressure waveform closely follows the measured one.
Figure 7a from Mahmud et al. (2023), cropped to one example. The estimated arterial pressure waveform (red) and measured waveform (blue), with the PPG and ECG inputs above.View full-size figureSource article© 2022 Elsevier Ltd · author reuse
Left: scatter plot of estimated against measured mean arterial pressure from about 60 to 140 mmHg, clustered along the identity line. Right: Bland–Altman plot with most differences in a narrow band around zero.
Figure 5 from Mahmud et al. (2023). (a) Regression and (b) Bland–Altman plots for mean arterial pressure.View full-size figureSource article© 2022 Elsevier Ltd · author reuse

My contribution

For the shallow U-Net study, I led conceptualisation, methodology, software, formal analysis, investigation, data curation, validation, visualisation and manuscript preparation. For PPG2ABP, I contributed to data curation, formal analysis, methodology, investigation, validation and manuscript drafting and revision, and worked with the first author to rework the study for publication. For the combined pipeline, I led development and evaluation of NABNet, including data preparation, implementation, experiments, analysis, visualisation and manuscript preparation.

Study scope

All three studies were developed on the same public intensive-care dataset of clinical finger PPG, and distorted or extreme signals were excluded in two of them. Only the shallow U-Net study tested an external dataset, with 40 people. Motion artefacts from wearable sensors were not resolved, and PPG2ABP is not intended for wrist PPG. Prospective validation is needed before clinical use.