M2PIM: A Dynamic Blood Pressure Estimation Model Fusing Physical and Physiological Information

M2PIM (Multimodal Physics- and Physiology-Informed Model) is an innovative multimodal blood pressure monitoring method. It combines physical constraints (e.g., hemodynamic laws) and physiological knowledge, and achieves dynamic and accurate beat-by-beat blood pressure estimation through multimodal data fusion (PPG, ECG, acceleration signals), providing new ideas for continuous non-invasive blood pressure monitoring. The original author of the project is txiang0705, the source platform is GitHub, and the release date is May 28, 2026.

Blood pressure is a core indicator of cardiovascular health. Continuous and accurate monitoring is crucial for hypertension management, intensive care, etc. Traditional cuff-based methods only measure intermittently and cannot capture dynamic changes. Among existing continuous monitoring methods, invasive ones (arterial catheterization) have high risks, while non-invasive ones (pulse wave analysis, pulse transit time) face problems such as large individual differences, difficulty in calibration, and sensitivity to motion artifacts. M2PIM proposes integrating physical constraints and physiological knowledge into deep learning models to address these challenges.

1. Introduction of Physics-Informed Neural Networks (PINN) : Incorporate hemodynamic physical constraints (mass conservation, momentum conservation, elastic tube model) as soft constraints into the loss function to reduce dependence on large-scale labeled data and improve generalization ability; 2. Physiological information modeling : Use multimodal inputs including PPG (peripheral blood flow), ECG (cardiac electrical activity), and acceleration (motion artifact compensation); 3. Multimodal fusion architecture : Hierarchical fusion—intra-modal feature extraction (dedicated encoders), inter-modal interaction (cross-attention), and physical constraint layer (rationality check).

Beat-by-beat estimation : Achieved through ECG R-peak time alignment, context information fusion (previous and next cycles), and adaptive calibration (intermittent cuff measurements); Robustness design : Motion artifact suppression (acceleration detection status), signal quality assessment (weight adjustment for low-quality data), and outlier detection (filtering anomalies via physical constraints).

Application scenarios : 24-hour ambulatory blood pressure monitoring (for hypertensive patients), intensive care assistance (replace/supplement invasive methods), exercise cardiovascular assessment (real-time monitoring), sleep apnea screening (nighttime fluctuation identification); Technical insights : The "knowledge + data" hybrid modeling (domain knowledge + deep learning) can alleviate problems such as scarce annotations, insufficient generalization, and poor interpretability in medical AI, which is worth referencing for other physiological signal analysis tasks.