Unsupervised machine learning for identifying morphological phenotypes in abdominal aortic aneurysms using fully...

Area of Science:

• Vascular Surgery
• Medical Imaging Analysis
• Machine Learning in Healthcare

Background:

• Thromboembolic complications after abdominal aortic aneurysm (AAA) repair are linked to thrombus burden.
• Fully automatic volume segmentation (FAVS) can extract morphological features to identify thrombogenic phenotypes.

Purpose of the Study:

• To use FAVS and machine learning to identify distinct morphological phenotypes of AAA.
• To investigate the association between these phenotypes and thromboembolic outcomes.
• To explore potential sex-specific differences in AAA morphology and outcomes.

Main Methods:

• Retrospective multi-center cohort study of elective AAA repairs (2013-2023).
• Utilized FAVS on pre-operative imaging to analyze thrombus burden.
• Applied principal component analysis and k-nearest neighbors clustering to identify morphological subtypes.
• Matched imaging data with National Vascular Registry thromboembolic outcomes.

Main Results:

• Two distinct AAA morphological subtypes were identified in 1455 patients, with a significant sex imbalance (13.8% vs. 4.1% women).
• Clusters differed significantly in thrombus burden across various aortic regions (P < 0.001).
• No significant difference in thromboembolic events was observed between clusters (2.1% event rate).

Conclusions:

• Unsupervised machine learning effectively identifies distinct AAA morphological phenotypes with varying thrombus burden and sex distribution.
• While thromboembolic events were infrequent and not significantly different between identified clusters, these phenotypes offer a framework for future embolic risk and sex-specific research.