Physics-Informed Neural Networks Meet Multimodal Large Language Models: Biomechanical Simulation in Aortic Aneurysm

Axial dissections of the thoracic artery are common causes of death in people diagnosed with aortic dissection; however, decisions to intervene on ascending thoracic aortic patients are determined by the size of the ascending thoracic aorta based on its diameter. Diameter-based criteria fail to take into consideration the biomechanical properties of the aorta as well as other characteristics of the patient, and finite element analysis (in determining aortic wall stresses) would ideally address the issues associated with today’s diameter-based criteria. A novel computational solution, termed BioPINN-LM, integrates 2 computational methods for the rapid real-time prediction of aortic wall stress: a physics-informed neural network (PINN) trained with mechanical simulation data to predict wall stress and a multimodal large language model that uses output data from the PINN along with image-based geometry descriptors to provide interpretable risk assessments from both ends of the aorta. Comparison of the PINN to the reference database of finite element method results demonstrates that the PINN produces an average wall stress prediction error of 8.34 kPa (6.12% relative error) for a blood pressure of 120/80 mmHg, with an average prediction time of 0.83 s per geometry compared to 38.6 min for the traditional finite element method. The conversational component of BioPINN-LM achieves 87.4% agreement with board-certified specialist recommendations on a benchmark of 200 simulated clinical scenario vignettes. As a simulation-based proof of concept, these results suggest that integrating mechanistic simulations with language-based inquiry may complement biomechanical decision-making for ascending thoracic aortic aneurysms; clinical validation on real patient cohorts with longitudinal outcomes is deferred to future work.