Summary: | This paper presents a neural network model to estimate arterial blood pressure (ABP) waveforms using electrocardiogram (ECG) and photoplethysmography (PPG) signals and its first two order mathematical derivatives (PPG<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><msup><mrow></mrow><mo>′</mo></msup></semantics></math></inline-formula>, PPG<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><msup><mrow></mrow><mrow><mo>″</mo></mrow></msup></semantics></math></inline-formula>). In order to achieve this objective, a lightweight and optimized neural network architecture has been proposed, made of Conv1D and BiLSTM layers. To train the network, the UCI Database “Cuff-Less Blood Pressure Estimation Data Set” has been used, which contains ECG and PPG signals together with the corresponding ABP waveform data; then the first two PPG derivatives have been computed. Four different configurations and parameter sets have been tested to choose the best structure and set of parameters. Additionally, various batch sizes, numbers of BiLSTM layers, and the presence of a maximum pooling layer have been tested. The best performing model achieves a mean absolute error of around 2.97, which is comparable to the state-of-the-art methods. Results prove deep learning techniques can be effectively used for non-invasive cuffless arterial blood pressure estimation. The lightweight and optimized model can be effectively used for continuous monitoring of blood pressure, which has significant clinical implications. Further research can focus on integrating the proposed model with wearable devices for real-time blood pressure monitoring in daily life.
|