P222. Multicenter Study Design for Development of a Predictive Model for Ascending Aortic Aneurysm Growth Using Artificial Intelligence via Federated Learning

Sheraton Times Square 

Objective: Thoracic aortic aneurysms (TAAs) are associated with an increased risk of aortic rupture or dissection. However, the optimal timing for preemptive surgical intervention remains uncertain. Current societal guidelines rely on maximum aneurysm diameter to determine when to intervene, but this one-size-fits-all approach has limitations. Modern artificial intelligence (AI) enabled models have significant potential for characterizing disease patterns, but they require large datasets to achieve clinically useful performance. TAAs and their related complications are relatively uncommon within the general population, making it difficult for any single institution to achieve sufficient cohort size alone. Multi-institutional studies are traditionally time consuming, logistically challenging, and expensive to implement to ensure patient data security. Federated learning is an approach that allows for the training of a single AI prognostic model across multiple institutions without the need for sharing of protected patient data between the centers. We present a framework for a multi-institutional study to train an AI model capable of predicting a TAA patient-specific risk of aortic complication from computed tomography (CT) scans taken at multiple timepoints.

Methods: The data pre-processing can be distilled into the following steps: (1) patient cohort creation, (2) inclusion filtering, (3) study acquisition, (4) image series selection, (5) image resizing, and (6) aorta segmentation. The data pre-processing must be robust enough to remain highly accurate across many different CT acquisition protocols. These pre-processing steps are automated and can be run locally at each institution. The AI model can then be trained from the processed CT scans on a primary outcome of aortic annualized growth rate. Federated learning will be utilized in order to keep all patient sensitive data at each institution, decreasing costs and improving data security. Open-source segmentation models are utilized and all code is written in Python (Version 3.1).

Results: We performed retrospective review of our tertiary academic center patient population to identify patients with TAA via billing documentation. We next developed natural language processing (NLP) methods of identifying patients that have no history of prior aortic intervention and would be eligible for the study. Cross sectional imaging for eligible patients when available were collected in an automated fashion, and open-source algorithms were utilized for series selection, image pre-processing, and aortic segmentation. Aortic size was calculated from the raw images and cross referenced to the radiologist reported measurements. Data integrity and quality checks were incorporated throughout the process. The final dataset was found to be in the appropriate format and data structure to undergo subsequent AI model development.

Conclusions: This proof-of-concept study demonstrates the feasibility of the proposed study design to create a patient specific TAA prognostic AI model. The end result is an automated, repeatable and scalable process for the creation of an institution's TAA imaging dataset that is ready for collaborative multi-institutional AI model development via federated learning.

Authors
Brian Ayers (1), Aaron Aguirre (1), Michael Lu (1), Thoralf Sundt (1)
Institutions
(1) Massachusetts General Hospital, Boston, MA

PODS will be on display in the exhibit hall for the duration of the meeting during exhibit hall hours. PODS will also be available for viewing on the meeting website. There is no formal presentation associated with your POD, but we encourage you to visit the PODS area during breaks to connect with those viewing.