CardiacPINN: A Physics-Informed Neural Network with multi-head attention for non-invasive assessment of cardiac function

Abstract

Accurate and rapid assessment of left ventricular (LV) contractility is crucial for managing cardiovascular diseases, yet current methods are limited by invasiveness or sparse data. To address these challenges, this study introduces CardiacPINN, a Physics-Informed Neural Network (PINN) designed for fast, non-invasive, and physiologically consistent estimation of LV contractility. CardiacPINN utilizes a closed-loop lumped-parameter cardiovascular model with five compartments to capture pressure–volume interactions and integrates multi-scale Fourier features, hierarchical self-attention, and an adaptive physics-constrained loss balancing strategy to enforce hemodynamic laws and ensure numerical stability. Evaluated using synthetic datasets generated via Latin Hypercube Sampling and experimental swine data, CardiacPINN demonstrated a convergence rate up to 76% faster than a baseline PINN on swine datasets and achieved a root-mean absolute error one order of magnitude lower in key compartments like the vena cava and peripheral arteries. The model accurately reproduced pressure–volume loops in both synthetic and real swine data, the quantitative range of R${}^2$ values achieved on the swine data for both LV pressure (0.88–0.99) and LV volume (0.59–0.97), effectively capturing contractility changes induced by dobutamine. These results indicate that CardiacPINN offers robust and efficient estimation of LV contractility across various physiological conditions, paving the way for real-time cardiac monitoring and personalized healthcare.