BPNet: A multi-modal fusion neural network for blood pressure estimation using ECG and PPG

This paper aims to address the problem of estimating blood pressure (BP) based on specific physiological signals. However, several issues exist in current works. Firstly, some methods produce biased models by training them on a skewed distribution and failing to consider the assumptions underlying regression analysis. Secondly, data leakage can occur when overlapping samples are present in the dataset. Finally, some methods concatenate Electrocardiogram (ECG) and Photoplethysmogram (PPG) signals without fully capturing their relationship. To overcome these limitations, this paper applies the BoxCox transformation to correct for skewed label distributions and creates a non-overlapping dataset. Additionally, a novel end-to-end model, the blood pressure network (BPNet), is proposed, which can accurately estimate blood pressure. The Multi-parameter Intelligent Monitoring in Intensive Care database (MIMIC) is utilized to develop and verify the model, which includes ECG, PPG, and invasive BP data from patients in intensive care units (ICUs). The BPNet is evaluated on the MIMIC II (942 subjects) and MIMIC-III (833 subjects) datasets. In MIMIC-II, the estimation error for systolic blood pressure (SBP) and diastolic blood pressure (DBP) is -0.17 & PLUSMN; 4.62 mmHg and -0.24 & PLUSMN; 2.95 mmHg, respectively. In MIMIC-III, the estimation error for SBP and DBP is -0.30 & PLUSMN; 5.78 mmHg and -0.25 & PLUSMN; 3.80 mmHg, respectively. The experimental results exceed the performance of existing methods and meet the accuracy standards established by the Association for the Advancement of Medical Instrumentation (AAMI) and the British Hypertension Society (BHS). This demonstrates the effectiveness and accuracy of the proposed model, highlighting the potential for practical applications in clinical settings.