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

Thrombo- and microembolic complications following abdominal aortic aneurysm (AAA) repair are hypothesized to be associated with wall thrombus burden. Fully automatic volume segmentation (FAVS) of imaging enables extraction of morphological features from which thrombogenic phenotypes may be identified.

This was a multi-centre retrospective cohort study using FAVS to examine pre-operative imaging for elective AAA repairs (2013-23). Radiological data were matched with National Vascular Registry thromboembolic outcomes data (cerebral, bowel, renal or limb ischaemia). Principal component analysis was used for dimensionality reduction, followed by unsupervised machine learning with k-nearest neighbours clustering, with number of clusters determined using silhouette scores. Clusters were compared using multivariate logistic regression, adjusting for aortic size index, cardiovascular risk parameters, and repair-type. Of 1655 patients, 1455 had sufficient quality imaging for FAVS (145 women and 1310 men). k-nearest neighbours clustering identified two morphological subtypes (n = 878 and n = 577), with sex imbalance (13.8 vs. 4.1% women, P < 0.001). The clusters differed in wall thrombus burden in visceral vessels, infra-renal aorta, aneurysmal neck, and common iliac arteries (P < 0.001). On adjusted multivariate regression, there was no significant differences in thromboembolic events between clusters, although event rate was low (n = 31, 2.1%) (odds ratio 1.56, 95% confidence interval 0.71-3.43, P = 0.23).

Unsupervised machine learning can identify distinct aneurysm morphological phenotypes with significant thrombus burden difference, which exhibit sex imbalance. While thromboembolic events were infrequent and did not differ significantly between clusters, these anatomical phenotypes may provide a framework for future studies investigating embolic risk and sex-specific disease mechanisms.