Introduction: The rising prevalence of cardiovascular disease, especially hypertension, necessitates continuous and non-invasive blood pressure (BP) monitoring. While previous studies have investigated the potential for estimating BP from electrocardiograms (ECGs) , these claims require further verification. This study presents a preliminary investigation into the feasibility of estimating BP from ECG pulse morphology using a novel deep compositional Bayesian neural network (auto-BNN).

Methods: Our model leverages a deep learning architecture to capture the variations in ECG morphology associated with BP. It incorporates convolutional neural network (CNN) layers for ECG waveform feature extraction, a long short-term memory (LSTM) unit to capture temporal dependencies in ECG sequences, and variational layers based on auto-BNN to enable uncertainty estimation.

Results and Discussion: An initial evaluation of data from 130 individuals sourced from Physionet yielded mean errors of 3.38 mmHg (systolic) and 2.40 mmHg (diastolic) with standard deviations of 13.20 mmHg and 11.88 mmHg, respectively. These results suggest that our model could potentially capture correlations between BP variations and ECG signals, such as changes in R wave amplitude, ST-segment depression, T-wave inversions, and widened P waves associated with high BP, as well as sinus tachycardia and ST-segment/T-wave changes associated with low BP. However, it is important to note that these correlations may have captured the relationship between heart rate (HR) and BP. Further research should explore methods to exclude HR information from ECGs to ensure the validity of BP estimation findings.

Conclusions: This preliminary study demonstrates the potential feasibility of using ECG pulse morphology for BP estimation. However, further validation on larger and more diverse datasets is crucial to assess the generalizability of our approach. While our initial results are encouraging, it is important to note that achieving very high accuracy in BP estimation solely from ECGs may be inherently challenging due to the complex and multifaceted factors influencing BP.