Digital Twin and Ml-basEd MOdel of TEVAR Interventions (MEMO)

In recent years, Thoracic Endovascular Aortic Repair (TEVAR) has become increasingly utilized for the treatment of thoracic aortic pathologies. Over the past two decades, the adoption of TEVAR has grown significantly, progressively replacing open surgery as the preferred treatment approach in many cases. Initially designed for interventions involving the descending thoracic aorta, TEVAR is now being extended to more complex anatomies, including the aortic arch and even regions closer to the aortic root.

Successful TEVAR procedures rely on accurate preoperative planning and detailed clinical assessment to optimize patient outcomes. Although TEVAR offers several advantages over open surgery, including reduced procedural risk, shorter recovery time, and lower morbidity, it is not without limitations. Major complications include endoleaks, stent-induced new entry tears, vessel obstruction, and stent migration, all of which may significantly affect patient prognosis. Despite existing manufacturer guidelines and deployment strategies, these complications remain difficult to predict.

Previous studies have reported endoleak rates ranging from 4% to 15%, stent migration rates between 1.0% and 2.8%, and device-related complications occurring in up to 38% of cases. Recent advances in computational modeling have demonstrated considerable potential for improving TEVAR planning and risk prediction. Finite element analysis (FEA) and fluid-structure interaction (FSI) simulations have proven valuable for assessing stent behavior within patient-specific anatomies. Through in silico simulations, different stent types and diameter configurations can be virtually tested, providing surgeons with critical insights for clinical decision-making.

However, despite their high accuracy, these techniques are computationally intensive and require large datasets as well as specialized expertise, limiting their accessibility for routine clinical practice. To address these challenges, numerical models (e.g., finite element simulations) and machine learning (ML) approaches represent promising alternatives for real-time, data-driven perioperative decision support. By integrating finite element simulations with clinical imaging data, ML algorithms can be trained to predict procedural outcomes, optimize prosthesis selection, and estimate post-interventional risks. This approach not only enhances pre-procedural planning but also facilitates postoperative risk assessment, ultimately contributing to improved patient management.

A critical challenge in developing robust ML models for TEVAR planning is the limited accessibility of high-quality annotated datasets and their integration into clinical workflows. To overcome this limitation, the study proposes a comprehensive methodology aimed at:

I) collecting clinical and imaging data relevant to TEVAR procedures; II) augmenting patient-specific anatomical data using statistical shape modeling (SSM) to generate a diverse training dataset; III) developing high-fidelity digital twins that provide personalized virtual replicas of individual TEVAR cases; and IV) training ML models on these augmented datasets to predict procedural outcomes based on patient-specific characteristics.

Using these techniques, the study aims to develop a clinically viable framework capable of predicting surgical outcomes and increasing the information available for surgeons during preoperative decision-making, thereby improving patient outcomes in TEVAR interventions.