Unraveling blood pressure estimation with a deep learning approach using multiple embeddings

We introduce a calibration-free machine learning framework for BP estimation using pulse arrival time (PAT), computed from the electrocardiogram's R-peak and photoplethysmography P-peak. To enhance pattern recognition and unveil hidden patterns within the data samples, we introduce the use of similarity-based features based on Euclidean and Manhattan distance matrices, which are then processed by an attention-guided convolutional neural network. The model was successfully evaluated across three datasets: Cabrini Hospital, PTT PPG, and MIMIC-II, where our framework achieved a R² values of 0.89, 0.95, and 0.92 for systolic BP (SBP) and 0.89, 0.94, and 0.91 for diastolic BP (DBP), respectively, along with mean absolute errors of 6.45, 1.31, and 2.12 mmHg for SBP and 2.92, 0.98, and 1.14 mmHg for DBP. The framework meets the Advancement of Medical Instrumentation standard on all datasets and achieves British Hypertension Society Grade 'A' for both BP types on the PTT PPG and MIMIC-II, and Grade 'A' and 'B' for DBP and SBP on the Cabrini, respectively. With strong generalizability, real-time compatibility, and no requirement for subject-specific calibration, the proposed framework demonstrates strong correlation, low prediction errors, and clinical applicability across diverse populations, offering a promising solution for continuous, comfortable, and reliable BP monitoring.