Abstract 2011: Development Of Machine Learning Model To Combine Clinical And Imaging Data For Prediction Of Endoleaks: Proof-of-concept Study

Abstract

Listen

Translate article

Objective: To develop a foundational ensemble machine learning (ML) classifier to train on computed tomography angiography (CTA) scans and clinical data to predict the occurrence of endoleaks after endovascular aneurysm repair (EVAR) of the abdominal aorta. Methods: An ensemble supervised ML model was developed by using the Convolutional Neural Network (CNN) classifier to transform imaging data into a vector that can be analyzed by a supervised ML model. To test this model, we applied CTA scans from 14 patients with an endoleak after infra-renal EVAR and 14 propensity-matched patients without an endoleak. Sixty DICOM images in batches of 3 were selected from each patient for a total of 20 batches per patient. These batches were assessed by the CNN model to output a singular representative vector for the respective patient (n=28). A supervised ML model was then trained on these cases using Leave-One-Out Cross Validation. To account for variability based on the random state, a for-loop was used to iterate through random states 0 to 49 and the average accuracies and AUC were taken from these 50 iterations. The clinical data was then standardized using a MinMaxScaler and appended to the representative vectors so that both the clinical and imaging data could be used as input for the supervised ML model. The ensemble model was assessed by testing accuracy and the area under the receiver operating characteristic curve (AUROC) and compared to the best ML model alone and imaging alone predictions for validation. Results: Using the clinical data alone, ML accuracy and AUROC performed with and accuracy of 57% and AUROC curve of 0.57. Using the imaging data alone, the final RFC model reported an accuracy of 59% and an AUROC curve of 0.59. Combining the imaging and clinical data, the final RFC model reported an accuracy of 60% and an area under the receiver operating curve of 0.60. Conclusions: Preliminary data from these models have shown proof-of-concept that a deep learning model such as a CNN can be combined with a supervised machine learning model to combine CTA imaging and clinical data to predict complications from an endovascular intervention. These models can serve as a proof of concept for future studies analyzing endovascular cases with larger datasets.