Identification of High-Risk Left Ventricular Hypertrophy on Calcium Scoring Cardiac Computed Tomography Scans: Validation in the DHS

Abstract

Background: Coronary artery calcium scoring only represents a small fraction of all information available in noncontrast cardiac computed tomography (CAC-CT). We hypothesized that an automated pipeline using radiomics and machine learning could identify phenotypic information about high-risk left ventricular hypertrophy (LVH) embedded in CAC-CT.

Methods: This was a retrospective analysis of 1982 participants from the DHS (Dallas Heart Study) who underwent CAC-CT and cardiac magnetic resonance. Two hundred twenty-four participants with high-risk LVH were identified by cardiac magnetic resonance. We developed an automated adaptive atlas algorithm to segment the left ventricle on CAC-CT, extracting 107 radiomics features from the volume of interest. Four logistic regression models using different feature selection methods were built to predict high-risk LVH based on CAC-CT radiomics, sex, height, and body surface area in a random training subset of 1587 participants.

Results: The respective areas under the receiver operating characteristics curves for the cluster-based model, the logistic regression model after exclusion of highly correlated features, and the penalized logistic regression models using least absolute shrinkage and selection operators with minimum or one SE λ values were 0.74 (95% CI, 0.67-0.82), 0.74 (95% CI, 0.67-0.81), 0.76 (95% CI, 0.69-0.83), and 0.73 (95% CI, 0.66-0.80) for detecting high-risk LVH in a distinct validation subset of 395 participants.

Conclusions: Ventricular segmentation, radiomics features extraction, and machine learning can be used in a pipeline to automatically detect high-risk phenotypes of LVH in participants undergoing CAC-CT, without the need for additional imaging or radiation exposure. Registration: URL http://www.clinicaltrials.gov. Unique identifier: NCT00344903.

Keywords: cardiac-gated imaging techniques; heart failure; hypertrophy; phenotype; tomography.

Figures

Figure 1.

Schematic illustration of the study pipeline. (a) Segmentation: the automated multiatlas-based algorithm performed semantic segmentation of the cardiac chambers; color overlay representing the segmentation results (LV, left ventricle; RV, right ventricle; LA, left atrium; RA, right atrium). (b) Radiomics extraction: the LV segmentation was used to extract radiomics features from the CT data. (c) The final cohort was randomly split into training (80%) and validation (20%) datasets; radiomics features and correspondent cardiac magnetic resonance phenotypes were used to build the predictive models on the training dataset; posteriorly, the models were tested against the validation set.

Figure 2.

Left ventricle (LV) measurements: Coronary artery calcium scoring computed tomography (CAC-CT) versus cardiac magnetic resonance (CMR). A, Scatter plot of LV volumes measured on CAC-CT versus diastolic CMR (R=0.57 [95% CI, 0.54 – 0.60], P

Figure 2.

Left ventricle (LV) measurements: Coronary artery calcium scoring computed tomography (CAC-CT) versus cardiac magnetic resonance (CMR). A, Scatter plot of LV volumes measured on CAC-CT versus diastolic CMR (R=0.57 [95% CI, 0.54 – 0.60], P

Figure 2.

Left ventricle (LV) measurements: Coronary artery calcium scoring computed tomography (CAC-CT) versus cardiac magnetic resonance (CMR). A, Scatter plot of LV volumes measured on CAC-CT versus diastolic CMR (R=0.57 [95% CI, 0.54 – 0.60], P

Figure 2.

Left ventricle (LV) measurements: Coronary artery calcium scoring computed tomography (CAC-CT) versus cardiac magnetic resonance (CMR). A, Scatter plot of LV volumes measured on CAC-CT versus diastolic CMR (R=0.57 [95% CI, 0.54 – 0.60], P

Figure 3.

Heat map using unsupervised clustering, participants in the horizontal axis (n = 1,982) and features along the vertical axis (n = 110). Normalized values are color-coded between zero and one, as shown by the color key. Features are grouped by color along the first left column, according to the classes detailed in the top left legend.

Figure 4.

Receiver operating characteristic curve analysis of the predictive models against the validation set. LASSO, least absolute shrinkage and selection operator