Ascending thoracic aortic aneurysms often progress silently, but once dissection or rupture occurs, they can cause severe mortality before hospital arrival or during admission. Current clinical decisions for surgical intervention mainly rely on the maximum transverse diameter of the ascending aorta, such as approximately 5.5 cm for patients with tricuspid valves and 5.0 cm for those with bicuspid valves. However, a substantial proportion of type A dissections occur below these surgical thresholds, indicating that diameter alone cannot fully capture local wall stress, tissue strength, and individualized biomechanical risk. Finite element analysis can reconstruct patient-specific aortic anatomy from medical imaging and estimate local wall stress, providing a more detailed biomechanical assessment. Yet routine clinical use remains limited because the workflow requires manual data input, surface generation, volume meshing, material-property definition, and numerical solving, often taking several hours per case and limiting real-time clinical exploration. Recent physics-informed neural networks offer a way to embed governing equations into neural networks and rapidly approximate biomechanical solutions without the conventional mesh-and-solve loop, while multimodal large language models can integrate imaging, geometry, hemodynamics, comorbidity profiles, and guideline information into interpretable natural-language outputs. "However, existing PINNs have not yet been applied to structural wall-stress analysis in ascending thoracic aortic aneurysms, and current medical LLMs cannot directly use physics-based simulation outputs to generate biomechanical recommendations." said the author Shichao Zhu, a researcher at Fudan University, "Therefore, combining fast physics-based simulation with interpretable conversational risk assessment is an important step toward individualized biomechanical decision support for ascending thoracic aortic aneurysm management."
This study proposed BioPINN-LM, a hybrid computational framework for rapid wall-stress prediction and conversational risk interpretation in ascending thoracic aortic aneurysms. The researchers first extracted morphological information from CTA-derived ATAA geometry, encoded the luminal surface using spherical harmonics, and combined this representation with luminal pressure as input to the PINN. The PINN branch took 3D spatial coordinates, a geometry-conditioning vector, and pressure as inputs, and used a Fourier feature layer, an 8-layer multilayer perceptron, and the Holzapfel–Gasser–Ogden hyperelastic constitutive model to derive the Cauchy stress field from predicted displacement through automatic differentiation. Its training loss jointly included mechanical equilibrium residuals, boundary conditions, finite-element displacement supervision, and near-incompressibility constraints. To build the training dataset, the study generated 1,247 synthetic ATAA geometries using 7 morphological parameters, meshed them with Gmsh, solved reference finite-element simulations with FEBio under pressure loads ranging from hypotension to severe hypertension, and further incorporated 86 public FEA simulations. The high-dimensional 3D stress field was then compressed into a 42-dimensional MechToken by dividing the aorta into 8 anatomical zones, extracting 5 statistics from each zone, including mean, 95th percentile, maximum, coefficient of variation, and stress-gradient magnitude, and adding 2 global features: peak-stress-to-strength ratio and Laplace stress ratio. The multimodal LLM branch was built on LLaVA-Med: axial and sagittal maximum-intensity projection images were encoded as visual tokens, fused with MechTokens and text queries, and passed to a Vicuna-7B decoder to generate conversational reports covering peak stress, regional stress, guideline comparison, and risk stratification. The overall training followed a sequential pipeline of PINN training, MechToken generation, and LLM tuning, and the framework was evaluated using a 200-case specialist-annotated simulated vignette benchmark, baseline comparisons, ablation studies, and out-of-distribution testing.
The results showed that BioPINN-LM performed well in both ascending aortic wall-stress prediction and conversational risk interpretation. On 267 held-out test geometries, the PINN branch achieved a mean absolute error of 8.34 kPa and a relative error of 6.12% under normotensive loading, with an R² of 0.961. Under hypertensive loading, it achieved an MAE of 11.07 kPa, a relative error of 5.89%, and an R² of 0.953, outperforming baseline models including Vanilla PINN, DeepONet, PI-DeepONet, and a 3D U-Net surrogate. Compared with conventional FEBio finite element analysis, the PINN branch required only 0.83 s per geometry, whereas FEBio required an average of 38.6 min; the full BioPINN-LM pipeline, including LLM inference, still generated outputs in less than 3 s, indicating potential for interactive analysis. In conversational decision-support evaluation, BioPINN-LM achieved 87.4% agreement with specialist majority recommendations on a 200-case simulated clinical vignette benchmark, with a weighted F1 score of 0.862 and a biomechanical reasoning score of 8.3, clearly exceeding the ESC rule-based system, GPT-4V without stress data, and LLaVA-Med without MechTokens. Ablation studies further showed that HGO hyperelastic constitutive embedding, spherical-harmonic geometry conditioning, and FEA displacement supervision were critical for stress prediction. Removing MechTokens reduced concordance to 77.3%, while the diameter-only variant reached only 78.5%, indicating that regional biomechanical stress information, rather than diameter or geometry alone, was the main contributor to improved reasoning. Robustness analysis suggested that the model retained some predictive ability under out-of-distribution conditions such as severe hypertension and bicuspid-valve-associated asymmetric morphologies, although these findings remain preliminary. The highest prediction error occurred at the sinotubular junction, suggesting that future work should improve sampling and modeling strategies for local stress-concentration regions.
The significance of this work lies in proposing BioPINN-LM, a framework that combines physics-informed neural networks with multimodal large language models, offering an integrated pathway from rapid wall-stress simulation to interpretable conversational reporting for ascending thoracic aortic aneurysm risk assessment. Compared with conventional diameter-based decision-making, the framework incorporates biomechanical information such as local wall stress, regional stress distribution, and peak-stress-to-tissue-strength ratio, which may provide additional reference for patients in the 45–55 mm clinical gray zone or those with risk modifiers such as Marfan syndrome or bicuspid aortic valve. The study also shows that embedding the HGO hyperelastic constitutive model into the PINN and transmitting region-resolved stress fields to the LLM through MechTokens are key to improving stress prediction and biomechanical reasoning. However, BioPINN-LM remains a simulation-based research prototype rather than a clinically deployable decision-support tool. Its limitations include the use of population-level rather than patient-specific material parameters, expert concordance evaluation based on 200 simulated clinical vignettes rather than real longitudinal patient outcomes, reliance on static CTA geometry without 4D-flow MRI or cardiac-cycle dynamic loading, and the need to further test performance under extreme pressure conditions. Future research should validate the framework prospectively in real CTA-derived patient cohorts linked to rupture, dissection, or surgical outcomes, extend the model to dynamic flow and full-cardiac-cycle stress prediction, and combine retrieval-augmented generation with real-time guideline access to reduce LLM hallucination risk. "If these challenges are addressed, the framework may become a supplementary biomechanical tool for multidisciplinary heart-team discussions and provide a transferable paradigm for integrating mechanistic simulation with medical AI." said Shichao Zhu.
Authors of the paper include Shichao Zhu, Mieradilijiang Abudupataer, Yulin Wang, Hongqiang Zhang, Hao Lai, Kai Zhu, Yongxin Sun, and Nan Chen.
This work was supported by the National Natural Science Foundation of China (Grant No. 82200525). The authors thank the cardiovascular specialists who participated in the benchmark annotation.
The paper, "Physics-Informed Neural Networks Meet Multimodal Large Language Models: Biomechanical Simulation in Aortic Aneurysm" was published in the journal Cyborg and Bionic Systems on Sept 2, 2026, at https://doi.org/10.34133/cbsystems.0658.
