Clinical prediction of aortic arch anatomy to guide stroke CT angiography protocols: A machine learning study.
Authors
Affiliations (4)
Affiliations (4)
- Departamento de Neurología. Hospital General de Zona No 2, Instituto Mexicano Del Seguro Social, Aguascalientes, México.
- Unidad de Investigación Epidemiológica y Servicios de Salud. Centro Médico Nacional Siglo XXI, Instituto Mexicano Del Seguro Social, Ciudad de México, México.
- Departamento de Terapia Endovascular, Centro Médico ABC, Ciudad de México, México.
- Departamento de Estadística, Centro de Ciencias Básicas, Universidad Autónoma de Aguascalientes, Aguascalientes, México.
Abstract
BackgroundThe aortic arch (AA) influences catheter navigation during endovascular treatment (EVT) for acute ischemic stroke. Whether routine inclusion of the AA in stroke CT angiography (CTA) protocols is necessary remains debated, particularly in settings aiming to streamline imaging workflows.PurposeTo evaluate whether machine learning algorithms (MLAs) can use routinely available clinical and laboratory data to estimate AA configuration in patients with stroke risk factors.Materials and MethodsIn this retrospective study, 108 patients who underwent CTA, including the AA, were analyzed. Eight MLAs were trained to classify AA anatomy as favorable or unfavorable using clinical and laboratory variables available at presentation. Class imbalance was addressed using the Synthetic Minority Oversampling Technique. Model performance was evaluated using five-fold cross-validation and summarized with seven performance metrics.ResultsAfter class balancing, mean accuracy across classifiers ranged from 0.52 to 0.77. Gaussian copula and Naive Bayes classifiers showed the highest accuracy, while k-nearest neighbors and support vector machines demonstrated higher sensitivity. Specificity was highest for regularized discriminant analysis and the Gaussian copula. AUROC values ranged from 0.49 to 0.80 across models.ConclusionRoutinely available clinical and laboratory data contain a limited yet detectable signal, enabling several machine learning algorithms to estimate AA configuration with moderate accuracy. These findings are exploratory and do not support modification of CTA protocols. However, they suggest that pre-imaging clinical data may provide supportive information for future workflow-oriented decision-support tools. Further validation in larger, externally tested cohorts is required before clinical integration can be considered.