Unsupervised machine learning cluster analysis to identification EVAR patients clinical phenotypes based on radiomics
Yonggang Wang1, Min Zhou2, Yong Ding2
1Department of Vascular Surgery, The First Affiliated Hospital of Naval Medical University, Shanghai, China.
Vascular
|June 17, 2024
Summary
Unsupervised machine learning identified distinct clinical phenotypes in patients undergoing endovascular aortic repair (EVAR) for abdominal aortic aneurysms (AAA). These radiomics-based clusters predict significantly different rates of severe adverse events after EVAR.
Area of Science:
- Radiomics and Machine Learning in Medical Imaging
- Vascular Surgery Outcomes Research
- Quantitative Medical Imaging Analysis
Background:
- Endovascular aortic repair (EVAR) is a common treatment for abdominal aortic aneurysms (AAA).
- Predicting patient outcomes after EVAR remains a challenge.
- Radiomics offers a novel approach to extract quantitative imaging features for outcome prediction.
Purpose of the Study:
- To explore clinical phenotypes of EVAR patients using unsupervised machine learning (UML) cluster analysis based on radiomics.
- To identify distinct patient groups with varying risks of severe adverse events (SAEs) post-EVAR.
- To validate the identified phenotypes in independent training and test sets.
Main Methods:
- Retrospective review of 1785 patients with infra-renal AAA undergoing elective EVAR (2010-2020).
- Radiomics feature extraction using Pyradiomics, followed by statistical analysis to identify features related to SAEs.
- UML cluster analysis on selected radiomics features for phenotype identification, validated in a test set. Kaplan-Meier analysis for freedom from SAEs.
Main Results:
- 1180 patients analyzed, with 353 experiencing EVAR-related SAEs.
- 23 radiomics features identified from 1223 extracted features to define phenotypes.
- Three distinct clusters identified with similar clinical/morphological features but varied radiomics; validated in test set.
- Kaplan-Meier analysis revealed significant differences in freedom from SAEs rates between clusters (p=.0216 training, p=.0253 test).
Conclusions:
- Unsupervised machine learning cluster analysis, utilizing radiomics, can effectively identify distinct clinical phenotypes in EVAR patients.
- These radiomics-defined phenotypes are associated with significantly different long-term outcomes regarding severe adverse events.
- This approach holds promise for personalized risk stratification and improved management of AAA patients undergoing EVAR.
