Generalization Capability of a Neural Network for Blood Pressure Estimation from Photoplethysmography

Clémentine Aguet1, Jérôme Van Zaen1, Martin Proença1, Guillaume Bonnier1, Pascal Frossard2, Mathieu Lemay1
1Swiss Center for Electronics and Microtechnology (CSEM), 2Ecole Polytechnique Fédérale de Lausanne (EPFL)


Introduction: Hypertension is a serious condition that greatly increases the risk of developing cardiovascular diseases. Continuous blood pressure (BP) monitoring allows early detection and prevents potential complications. In the effort of developing non-invasive, continuous and cuffless BP monitoring devices, photoplethysmography (PPG) has recently gained increasing interest. This simple and low-cost optical technology detects blood volume variations, which are related to cardiovascular parameters. Recent works highlighted the potential of feature learning approaches for BP estimation from PPG signals. However, such data-driven models raise the question of their generalization capability, which describes the ability of a model to adapt to unseen data. Although crucial to the success of the model, this aspect is, most of the time, not addressed in published studies.

Methods: This work proposes to evaluate the generalization capability of a PPG-based BP estimation model. We use PPG signals with associated BP reading from two datasets. Both were collected during non-cardiac surgical interventions but using different PPG sensors (finger-clip in transmission mode vs. smartphone in reflectance mode) and including different target populations (Asian vs. Caucasian). The proposed model based on convolutional layers combines the feature learning task with the regression task in a single architecture. It extracts representative information from an ensemble average pulse computed over PPG windows and estimates systolic (SBP) and diastolic BP (DBP) accordingly. The model is trained on one of the datasets and evaluated on the other.

Results and conclusion: On unseen subjects from the training dataset, the model achieved mean and standard deviation errors of -0.77±10.53 mmHg for SBP and -0.85±5.96 mmHg for DBP. Whereas errors on the other dataset were of 1.13±12.86 mmHg for SBP and -0.44±6.99 mmHg for DBP. Taken together, these results show that a feature learning model can extract feature representations that are generalizable over different populations and different PPG sensors.