Prediction of persistent form of atrial fibrillation using left atrial morphology on preprocedural computed tomography: application of radiomics

Abstract

Abstract Background Radiomics is a comprehensive analysis methodology of medical image and involves the extraction of numerous features from standard imaging. Its usefulness has been reported mainly in the field of cancer for diagnosis and prediction of prognosis. In the territory of cardiac imaging, several reports have investigated the utility of radiomics for classifying the risk of prognosis in coronary artery disease, and few practical applications have been reported for patients with atrial fibrillation (AF) who underwent catheter ablation (CA). Purpose The objective of this study was to evaluate the utility of radiomics analysis applying to the preprocedural cardiac computerized tomography (CT) in AF patients. Methods We analyzed 525 consecutive three-dimensional CT in patients with AF who underwent CA. After marking the region of interest on left atrium (LA) (including the root of pulmonary veins) semiautomatically, 107 radiomics feature values were obtained by Python program. We calculated the amount of representative statistics for each radiomics feature for prediction of persistent AF (PeAF) (Wald statistic in logistic regression analysis) and LA diameter (LAD) (coefficient correlation), respectively. To compare the distribution of the two statistics, the relative importance (calculated as the ratio of statistic to the maximum statistics among 107 radiomics features [%]) was calculated for each statistic. Further, we compared the area under the curve (AUC) in receiver operation characteristic (ROC) curve analysis for predicting PeAF between radiomics features (multivariate model) and LAD (single parameter). Results In 525 study patients (age 63±10 years old and male 80%), 253 (48%) were PeAF and remaining were paroxysmal AF (PAF). LAD was 43±6 mm and 38±6 mm in patients with PeAF and PAF, respectively. The relative importance of the two statistics (Wald statistic for PeAF and coefficient correlation for LAD) of 107 radiomics features are displayed in Figure 1, which shows similar distribution of two statistics. It means the close relationship between LA morphology and the form of PeAF in AF patients and the radiomics features possibly well explain the relationship. In Figure 2, the predictive capability for PeAF was compared between radiomics feature values and LAD, where the AUC was 0.85 (95% confidence interval [CI], 0.82–0.88) and 0.73 (95% CI, 0.69–0.78) for radiomics feature values and LAD, respectively (Delong test, P<0.001). Conclusion We applied the radiomics features for the evaluation of LA morphology. The predictive capability for PeAF in the prediction model with the radiomics feature values was much better than that with LAD alone. Since radiomics feature analysis yields a huge number of numerical values representing the LA morphology in a reproducible manner, it would provide a new direction to construct a good prediction model using machine learning including artificial intelligence out of a routine cardiac CT scan. Funding Acknowledgement Type of funding sources: None.