ParticipantsNam-Kyu Kang, MD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Young Joo Suh, MD, Seoul, Korea, Republic Of (Presenter) Nothing to Disclose
Kyunghwa Han, PhD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Jong Seub Jeon, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Young Jin Kim, MD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Byoung Wook Choi, MD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Sang A Lee, Bucheon-si, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
rkdskarb77@yuhs.ac
PURPOSEWe aimed to develop a model that can predict severe aortic stenosis (AS) using the computed tomography (CT) radiomic features of aortic valve calcium (AVC) and machine learning algorithm.
METHOD AND MATERIALSWe retrospectively enrolled 408 patients who underwent cardiac CT scan from March 2010 to August 2017 and had echocardiographic exams (231 patients with severe AS on echocardiography [severe AS group] and 177 patients without severe AS [non-severe AS group]). Datasets were divided into training sets (312 patients) and validation sets (96 patients) with a reference of specific time point. On the non-contrast cardiac CT scan, volume of AVC was calculated and a total of 128 radiomic features of AVC were extracted. Three feature selection methods (least absolute shrinkage and selection operator [LASSO] using 5-cross validation, Random Forest [RF], and XGBoost) were assessed for their performance in the diagnosis of severe AS, using c-index. The performances of the radiomics models were compared with the prediction model based on AVC volume.
RESULTSThe radiomic score derived from LASSO was significantly different between severe AS group and non-severe AS group (median 1.37 vs. 0.19, p<0.001). Radiomics prediction model based on features selection by RF showed the highest c-index 0.8632 (95% confidence interval [CI] 0.7822-0.9443) in the validation set, followed by model based on XGBoost (c-index 0.8655, 95% CI 0.7893-0.9417) and LASSO (c-index 0.857, 95% CI 0.7762-0.9377). Radiomics models based on the feature selection methods of RF and XGBoost showed higher predicted probability of severe AS compared with a model based on AVC volume only (c-index 0.8618, 95% CI 0.7762-0.9376 in the validation set), although it was not statistically significant (P>0.05 for all).
CONCLUSIONRadiomic feature of AVC performs better than AVC volume for prediction of severe AS.
CLINICAL RELEVANCE/APPLICATIONBy applying radiomics and machine learning to the aortic valve calcium score, it may allow to distinguish severe AS better in patients with limited assessment from echocardiography, such as patients with low-flow, low-gradient AS.