Abstract
Purpose
To extract radiomic features from coronary artery calcium (CAC) on CT images and to determine whether this approach could improve the ability to identify individuals at risk for a composite endpoint of clinical events.
Materials and Methods
Participants from the Offspring and Third Generation cohorts of the community-based Framingham Heart Study underwent noncontrast cardiac CT (2002–2005) and were followed for more than a median of 9.1 years for composite major events. A total of 624 participants with CAC Agatston score (AS) of greater than 0 and good or excellent CT image quality were included for manual CAC segmentation and extraction of a predefined set of radiomic features reflecting intensity, shape, and texture. In a discovery cohort (n = 318), machine learning was used to select the 20 most informative and nonredundant CAC radiomic features, classify features predicting events, and define a radiomic-based score (RS). Performance of the RS was tested independently for the prediction of events in a validation cohort (n = 306).
Results
The RS had a median value of 0.08 (interquartile range, 0.007–0.71) and a weak and modest correlation with Framingham risk score (FRS) (r = 0.2) and AS (r = 0.39), respectively. The continuous RS unadjusted, adjusted for age and sex, FRS, AS, and FRS plus AS were significantly associated with events (hazard ratio [HR] = 2.2, P < .001; HR = 1.8, P = .002; HR = 2.0, P < .001; HR = 1.7, P = .02; and HR = 1.8, P = .01, respectively). In participants with AS of less than 300, RS association with events remained significant when unadjusted and adjusted for age and sex, FRS, AS, and FRS plus AS (HR = 2.4, 2.8, 2.8, 2.3, and 2.6; P < .001, respectively). In the same subgroup of participants, adding the RS to AS resulted in a significant improvement in the discriminatory ability for events as compared with the AS (area under the receiver operating curve: 0.80 vs 0.68, respectively; P = .03).
Conclusion
A radiomic-based score, including the complex properties of CAC, may constitute an imaging biomarker to be further developed to identify individuals at risk for major adverse cardiovascular events in a community-based cohort.
Supplemental material is available for this article.
© RSNA, 2020
Summary
A radiomics-based score that incorporates complex properties of coronary artery calcium may constitute an imaging biomarker to identify individuals at risk for composite endpoint of major adverse clinical events.
Key Points
■ A radiomic-based score (RS) was derived, trained, and validated from noncontrast cardiac CT scans obtained in 624 patients in the Framingham Heart Study.
■ The RS predicted composite endpoint of clinical events during a median follow-up of 9.1 years, independent of other cardiovascular risk factors and coronary artery calcium (CAC).
■ The RS that incorporates complex properties of CAC may constitute as an imaging biomarker to further develop and characterize individuals at risk for composite endpoint of clinical events.
Introduction
Coronary artery calcium (CAC), as assessed by cardiac CT, is a well-established and robust risk predictor for major adverse cardiovascular events, independent of traditional cardiovascular risk factors (1–3). Agatston score (AS) is the most commonly used method to quantify CAC in clinical practice and is a simple technique defined as the product of CAC volume and weighted calcium peak CT attenuation factor (4). In addition, other measures of CAC such as volume, mass (5), number of segments and arteries with CAC (6,7), and calcium density (8,9) were discovered as further predictors of cardiovascular events.
Radiomics is a field in biomedical imaging that aims to extract high-dimensional quantitative features from digital images. Several radiomic features or biomarkers have shown predictive value in patients with cancer (10–12). In radiomics studies, it is hypothesized that image voxels contain information that can be converted into meaningful phenotypic characteristics of diseased tissues via computer vision (10,13). Upon image acquisition, the workflow of radiomic analysis includes image segmentation, feature extraction, and informatics and/or machine learning analysis. These features may carry additional information about intensity, shape, and texture and are extracted based on the segmented and masked regions of interest. Some features may be related to routinely used, radiologist-defined semantic characteristics, whereas others may be generally higher order and filtered metrics of different textural characteristics (10,13,14).
Thus, we hypothesized that using radiomics may uncover informative multidimensional properties of CAC that are incremental in predictive value of major cardiac events. Specifically, we determined whether a machine learning–based radiomic-based score (RS) applied to cardiac CT could improve risk stratification for composite major adverse clinical events as compared with Framingham risk score (FRS) and traditional AS in the Framingham Heart Study (FHS), an asymptomatic community cohort.
Materials and Methods
Study Population
In this retrospective cohort study, a subanalysis of the FHS multidetector CT imaging substudy was performed. The selection criteria and study design have been described previously in detail (2,15). Briefly, the CT substudy included participants from the FHS Offspring (16) and Third-Generation (17) cohorts who underwent noncontrast coronary calcium CT imaging from 2002 to 2005. Inclusion criteria for men and women were participant age of greater than or equal to 35 or 40 years, respectively, and participant weight of less than 450 lbs. For the radiomics analysis, we included participants with CAC and no prior known cardiovascular events (Fig 1). In addition, we included participants without CAC as a reference group. Participants with missing CT or clinical data were excluded. This study was reviewed and approved by the institutional review boards of Boston University Medical Center and Massachusetts General Hospital; all participants provided written consent to undergo cardiac CT.
Figure 1:
Patient selection criteria for the Framingham Heart Study (FHS) community-based cohort who had undergone cardiac CT imaging between 2002 and 2005; a final total of 624 FHS participants with CT scans were randomly split into a discovery cohort (n = 318) and validation cohort (n = 306). Ca = calcium, CAC = coronary artery calcium, CVD = cardiovascular disease, MDCT = multidetector CT.
Clinical Characteristics and Cardiovascular Outcomes
Participants were followed for a median follow-up length of 9.1 years (interquartile range [IQR], 7.8–10.1 years). The primary outcome was a composite of all-cause mortality, nonfatal ischemic stroke, or myocardial infarction. All study participants were under continuous surveillance for the development of cardiovascular events and death. The methods for cardiovascular event adjudication by an endpoint committee physician have been previously described (15,17).
CT Imaging
The cardiac CT examinations were performed using an eight-slice multidetector-row CT scanner (LightSpeed Ultra; General Electric, Milwaukee, Wis). The protocol included a noncontrast prospectively electrocardiographically triggered CT scan acquired with 120 kVp and 320 or 400 mAs for participants weighing less than 220 or greater than 220 lbs, respectively (18).
Image Segmentation and Traditional Agatston Score
For this study, image quality was scored on a scale of 1 to 4 as excellent, good, moderate, or poor, respectively, based on the presence and extent of motion artifact, presence of beam hardening, visibility of coronary calcium, and noise level. Only participants who had good or excellent image quality were included to avoid the influence of exogenous artifacts (eg, motion) in the radiomics analysis (Fig 1). Coronary arteries were manually segmented into eight segments, and CAC was labeled by three independent readers (P.E., B.F., J.E.S.) using an open source software (3DSlicer, v.4.7.0 [19]). All readers were blinded to the clinical endpoints. The four major coronary vessels segments were: (a) left main, proximal, and mid to distal left anterior descending, (b) proximal and mid to distal circumflex artery, (c) ramus intermedius, (d) proximal and mid to distal right coronary artery, using a simplified model derived from the American Heart Association and Society of Cardiovascular Computed Tomography 18-segment model guidelines (20). In the event of the continuation of calcified plaques to neighboring segments, plaques were labeled as the proximal segment. Similar to previous studies (18,21), CAC was identified as greater than or equal to 3 connected voxels with attenuation of greater than or equal to 130 HU. AS was calculated according to standard methods described by Agatston et al (4) using the volume and maximum pixel CT number. AS calculated per patient was categorized as low (1–100), intermediate (101–300), and high (> 300) (22). The interobserver reliability for AS was evaluated by comparing the results in a random subset of 30 patients. The intraobserver reliability for AS was evaluated by measuring AS twice for 30 cases (P.E. and B.F.).
Extraction of Radiomic Features
Overall, 2120 quantitative radiomic features were extracted (for both discovery and validation cohort) and normalized by the mean of each feature from the segmented CAC using an open source Python-based tool—Pyradiomics (23)—as described in detail in Appendix E1 (supplement).
Radiomics Modeling
All included FHS participants with CAC AS of greater than 0 were split into a discovery cohort (318 with odd participant identifiers) and a validation cohort (306 with even participant identifiers). Feature dimensionality reduction and machine learning modeling was performed (Fig 2) and is found in Appendix E2 (supplement). On the basis of these models, we derived an RS in the discovery cohort and then validated the performance of this score in the validation cohort, independently. To align our results with other studies on the prognostic value of CAC (6,18), we also included individuals with CAC equal to 0 (n = 1986) as a reference group in the final supplemental analysis, resulting in a final supplementary analytic validation cohort of 2292 participants. Therefore, all the main results reported herein are performed on the validation cohort of 306 participants, and the supplemental analysis is based on the validation cohort of 2292 participants. For steps involved in data integration and dimensionality reduction and machine learning modeling, please refer to the Appendix E2 (supplement).
Figure 2:
Methodology workflow. A, Coronary artery calcium (CAC) segmentation on coronary CT images. B, Radiomic phenotyping and extraction of features per plaque. C, Dimensional reduction of data, going from multiple plaques per patient for each feature, resulted in six summary statistics for each patient within each feature. Top 20 minimum redundancy maximum relevancy (MRMR) features were elected for each of the six descriptive statistics feature sets. D, Machine learning modeling was performed where the radiomic features combined with clinical outcomes in the discovery cohort were used to create a predictive model. The radiomic predictive model (radiomic-based score) was then validated using the validation cohort. LM = left main, LAD = proximal left anterior descending, LCX = proximal left circumflex, RCA = proximal right coronary artery, Std = standard.
Statistical Analysis
All statistical analysis was performed using R version 3.3.3 statistical package. Continuous data were presented as median and IQR, and categorical and ordinal variables were presented as frequencies and percentages. The differences in demographics and clinical risk factors between the discovery and validation cohort, as well as tertiles of RS, were assessed using the Kruskal-Wallis test and Fisher exact test for continuous and ordinal variables, respectively.
First, we used Cox proportional hazard regression models to investigate the association between RS and time to an event. Associations were expressed as hazard ratios (HRs), and 95% confidence intervals were reported. HR was first calculated treating RS as a continuous variable. To avoid skewed distribution, RS values were transformed using a base 2 logarithmic transformation. Cox regression models were initially adjusted for age and sex, Framingham global cardiovascular disease risk score (or FRS) (15), AS, followed by FRS and AS combined together. For all Cox models we tested whether the proportional hazards assumptions were met on the basis of Schoenfeld residuals. Furthermore, we assessed the overall goodness-of-fit of our Cox models by using the Grønnesby and Borgan (24) test which extends the Hosmer and Lemeshow goodness-of-fit test to survival data. The overall tests were all nonsignificant with P values of at least .15, suggesting that there is not sufficient evidence for gross model violations.
Due to the lack of established thresholds for RS, we repeated the Cox regression using RS tertiles, as this enables a relevant comparison between ordinal RS and the clinically accepted ordinal use of the AS. We further repeated subanalysis including all patients with zero CAC.
Second, we assessed the discriminatory ability of the RS using the area under the receiver operating characteristic curve (AUC) where the incremental value to AS and FRS was compared using a two-sided Venkatraman test as implemented in R package “pROC” (25). To include time-to-clinical event into the discriminatory analysis, we used the “survivalROC” (26) package in R. P value of < .05 (for two-sided) was considered as significant.
Results
Study Population
Of 3496 participants who underwent cardiac CT and had a complete risk profile in the FHS community study, 1312 of 3496 (37.5%) participants had interpretable images with CAC of greater than 0 and 1986 of 3496 participants with CAC equal to 0. After exclusion of 688 of 1312 (52.4%) participants with CAC of greater than 0 and moderate or poor image quality, the final study cohort consisted of 624 participants with CAC of greater than 0 and good and excellent image quality (Fig 1). The inter- and intraobserver reliability were excellent between the readers (intraclass correlation coefficient = 96.7% and 98.0%, respectively).
The 624 participants (average age: 58.70 years ± 11.20 [standard deviation], women: 226 of 624 [36.2%]) were asymptomatic, white, and had low-to-intermediate FRS (median: 0.08; IQR, 0.05–0.13). During a median follow-up of 9.1 years (range, 7.8–10.1 years), 59 participants experienced an event (myocardial infarction, ischemic, stroke, or all-cause mortality). There were 30 of 318 (9.4%) events in the discovery and 29 of 306 events (9.5%) in the validation cohort. There were no clinically meaningful differences in baseline demographics or cardiovascular disease risk factors between the discovery and the validation cohorts (Table E2 [supplement]). Similarly, AS did not differ significantly between the discovery and validation cohorts (median AS, 63.09 [IQR, 15.67–220.80] and 72.43 [IQR, 18.69–224.01]; respectively, P = .45).
Radiomic Features and RS
In the top 20 selected features, the textural differences were sensitive to plaque radiographic heterogeneity, such as neighboring gray tone difference matrix (NGTDM) features, emphasizing the coarseness of the plaque (Figure E1 [supplement]). A more detailed explanation of what NGTDM feature represents can be found in Appendix E1 (supplement). The final resulting RS included features quantifying CAC textural differences (eg, coarseness, busyness, complexity) as well as first-order intensity-statistics features. Coarseness is a measure of the average difference between the center voxel and its neighborhood in plaque and is an indicator for the spatial rate of change. For more information, please see Appendix E1 (supplement). The RS for participants who did not have calcium was set to be RS = 0, similar to the values assigned by AS = 0, as these participants did not present with calcium to be analyzed and scored. When discriminatory analysis was performed considering time to the composite endpoint of clinical events for the cutoff timepoint of 9 years, the AUC values remained the same.
RS and Traditional Risk Factors
In the validation cohort (n = 306, 38.6% women, mean FRS, 0.09 ± 0.07), RS had a median value of 0.08 (IQR, 0.007–0.71). RS had a nonsignificant weak correlation of r = 0.2 with FRS as a continuous variable. Similar to the AS, increasing ordinal RS (tertiles) was associated with older age, higher blood pressure, and higher FRS values (Table 1).
Table 1:
Baseline Characteristics and Outcomes in the Validation Cohort Stratified by Tertiles of the Radiomic-based Score
Association of RS versus Agatston Score
There was a modest correlation between RS and the AS (Pearson correlation coefficient: r = 0.39) with higher discordance at lower values (Fig 3). This was also observed for ordinal AS and RS.
Figure 3a:
(a) Correlation of Agatston score (AS) with the radiomic-based score (RS) and event distribution. (b) Comparison of performance between AS, RS, and the combined (AS plus RS) scores in the entire validation cohort. (c) Subanalysis in participants with AS of less than 300. AUC = area under the receiver operating characteristic curve.
Association of RS versus Other Calcium Measures
CAC volume and mean attenuation increased with RS tertiles significantly (P < .001). In addition, number of segments with CAC—previously reported to have a significant association with clinical events in the FHS cohort—significantly increased going from lower to higher tertiles in RS (P < .001) (Table 1).
Association between RS and Composite Endpoint of Clinical Events
There was a significant association between the RS and number of participants with composite endpoint of clinical events (X2= 20.7; P < .001). Composite endpoint of clinical events increased significantly from one of 306 (0.3%) in the lower RS tertile to nine of 306 (3.9%) in the middle tertile and 19 of 306 (6.2%) in the highest RS tertile (P < .001) (Table 1).
In multivariate analyses, the RS as a continuous imaging biomarker was significantly associated with composite endpoint of clinical events unadjusted, adjusted for age and sex, FRS, AS, and FRS plus AS (HR = 2.2, P < .001; HR = 1.8, P = .002; HR = 2.0, P < .001; HR = 1.7, P = .02; and HR = 1.8, P = .01; respectively). When this analysis was repeated for a subgroup of participants with AS of less than 300, RS association with events remained significant when unadjusted and adjusted for age and sex, FRS, AS, and FRS plus AS (HR = 2.4, 2.8, 2.8, 2.3, and 2.6; P < .001, respectively). Consequently, when divided into tertiles, individuals with an RS in the second and third tertiles had a significant association with events as compared with individuals with RS in the lowest tertile (used as reference). The magnitude and significance of these associations remained even after adjustment for both AS and FRS in the middle and higher tertiles (HR = 8.8 and 14.8, 95% confidence interval: 1.1, 69.8 and 1.9, 114.9; respectively) (Table 2). Similar results were found when the associations with events were performed including participants with CAC = 0 (Table E4 [supplement]).
Table 2:
Association of Radiomics Score as a Continuous and Categorical Variables
Discriminatory Value of RS
To evaluate the discriminatory ability of the RS for predicting composite endpoint of clinical events, we calculated the AUC. As shown in Figure 3b, the AUC for AS, RS, and the combined AS plus RS were 0.73, 0.76, and 0.79, respectively (P = .05 for AS vs AS plus RS). When repeated in participants with AS of less than 300, the RS improved discriminatory ability compared with AS (AUC = 0.68 for AS vs 0.79 for the RS; P = .07). In addition, the RS plus AS significantly improved discriminatory ability compared with AS alone (AUC = 0.80; P = .03) (Fig 3c). In a further supplementary analysis, the FRS+AS+RS multivariate model improved discriminatory ability of predicting events compared with FRS alone (AUC = 0.69 for FRS vs AUC = 0.81 for FRS+AS+RS, P < .001) (Figure E3 [supplement]).
Figure 3b:
(a) Correlation of Agatston score (AS) with the radiomic-based score (RS) and event distribution. (b) Comparison of performance between AS, RS, and the combined (AS plus RS) scores in the entire validation cohort. (c) Subanalysis in participants with AS of less than 300. AUC = area under the receiver operating characteristic curve.
Figure 3c:
(a) Correlation of Agatston score (AS) with the radiomic-based score (RS) and event distribution. (b) Comparison of performance between AS, RS, and the combined (AS plus RS) scores in the entire validation cohort. (c) Subanalysis in participants with AS of less than 300. AUC = area under the receiver operating characteristic curve.
Discussion
The primary objective of our study was to determine whether a machine learning–based assessment of CAC combined with radiomics in cardiac CT improves the ability to identify individuals at highest risk for composite endpoint of clinical events. Hence, we derived and independently validated an RS in FHS—an asymptomatic community-based cohort. The main findings of this study were twofold: first, we demonstrated that RS is an independent predictor (adjusted for AS, FRS) for future composite endpoint of clinical events. Notably, this significant association with composite endpoint of clinical events remained for RS when treating it both as a continuous and categorical (keeping the first tertile as a reference group) variable (Table 2). Second, the RS increases the discriminatory capacity of AS, significantly.
The strength of this study is twofold: first, we used radiomics to extract CAC features that are not quantifiable by conventional measurements, adding substantially greater data dimensionality. These features described coronary calcifications based on their intensity, shape, and textural differences. In particular, the radiomics analysis identified plaque textural differences that contribute most significantly in characterizing CAC. Second, we applied state-of-the-art machine learning methods to develop radiomic models and to derive an RS to predict clinical outcomes. The validation dataset indicates a strong predictive and discriminatory ability of our radiomic model for clinical outcomes. We showed that with every unit increase in RS, the participants have a significantly higher risk of having an event adjusted for FRS and FRS plus AS (HR = 1.7 and 1.8, respectively). The discriminatory analysis provided evidence for modest, incremental value of adding RS to AS. In participants with AS of less than 300, the incremental value of the RS was substantial for predicting clinical outcomes over and above AS (AUC = 0.68 vs 0.80, respectively; P = .03). This finding suggests that a more comprehensive analysis of CAC features may be especially useful in individuals with AS of less than 300 (6,18,27).
One of the most notable findings is that volume of CAC, the most important component of AS, is not part of the RS. The combination of 20 CAC features which RS is composed of does not include CAC volume. Instead, features describing textural differences were driving the RS. Thus, future studies may assess AS and RS in a complementary fashion. In our study, other calcium measures in addition to AS, such as CAC volume, mean attenuation, and the number of CAC segments, all previously (6-9,21) shown to predict cardiovascular disease events, had a significantly increasing trend with higher RS scores. However, in a univariate analysis, mean attenuation (AUC = 0.52) or number of segments with CAC (AUC = 0.67) both had a lower performance in comparison with RS.
Limitations
Our study had several limitations based on study design. The major limitation was the modest number of cardiovascular disease outcome events. Despite this limitation, the HR for RS in association with events and AUC of RS in predicting events supported our hypothesis and reached statistical significance. Nevertheless, external validation is still warranted to test the robustness of machine learning–based radiomics risk prediction models across different populations with a more diverse ethnicity and background. Another limitation of our study was the nature of the composite primary endpoint. Although this approach serves the purpose to explore general hypotheses about cardiovascular risk in general populations, further studies with statistical power to detect differences in more specific endpoints are still necessary to support the clinical utility of such a tool.
One of the contributing factors to the low number of events was exclusion of some participants with events because of suboptimal image quality. This procedural step was included according to previous studies that showed the influence of suboptimal image quality and motion artifacts on the overcalculation of AS (28,29). Therefore, in this proof-of-concept study of the prognostic value of radiomics of CAC, we only included images with good or excellent image quality as the information used for radiomics is much more likely to be susceptible to major motion artifacts. It is important to note that images included in this study were acquired approximately 15 years ago and from an image quality point of view, the selected images are much more generalizable to today’s quality of image acquisition using advanced scanner technology. Moreover, it is reassuring that there were no clinically meaningful differences in participants’ demographic and clinical risk factors between the included and excluded participants (Table E1 [supplement]).
We used a tertile-based RS categorical classification as an exploratory analysis of our results based on the clinical categorization of AS into low, medium, and high values. However, this was an arbitrary step taken without a priori knowledge of optimal thresholds. Future studies including an external validation cohort would be needed to establish prespecified thresholds. Finally, the FHS cohort is of white European ancestry, limiting the generalizability of the results to other races or ethnicities.
Conclusion
In this proof-of-concept study, we derived and independently validated a radiomic-based CAC score to predict clinical outcomes independent of the FRS and the AS. Further studies are needed to confirm these findings and determine a potential clinical impact on preventive patient management.
APPENDIX
SUPPLEMENTAL FIGURES
P.E. supported by the National Heart, Lung, and Blood Institute T32 fellowship (T32HL076136) and U.H. by the Research Training and Career Development K24 grant (K24HL13128).
Disclosures of Conflicts of Interest: P.E. Activities related to the present article: received grant funding from the Ruth L. Kirschstein Institutional National Research Award. Activities not related to the present article: disclosed no relevant relationships. Other relationships: disclosed no relevant relationships. C.P. disclosed no relevant relationships. B.F. disclosed no relevant relationships. J.E.S. disclosed no relevant relationships. A.I. disclosed no relevant relationships. R.Z. disclosed no relevant relationships. M.T.L. Activities related to the present article: received grant funding as a co-investigator on a clinical trial sponsored by Kowa and MedImmune/Astrazeneca. Activities not related to the present article: board member of Kowa and MedImmune/Astrazeneca as a coinvestigator on clinical trials; received grant funding from Nvidia Academic Program through a GPU grant. Other relationships: disclosed no relevant relationships. M.F. Activities related to the present article: received funding through the National Institutes of Health and The American Heart Association. Activities not related to the present article: disclosed no relevant relationships. Other relationships: disclosed no relevant relationships. R.S.V. Activities related to the present article: received grant funding through the National Institutes of Health. Activities not related to the present article: disclosed no relevant relationships. Other relationships: disclosed no relevant relationships. K.B. disclosed no relevant relationships. J.M.M. Activities related to the present article: received funding through the National Institutes of Health and the National Heart, Lung, and Blood Institute for role as statistician. Activities not related to the present article: disclosed no relevant relationships. Other relationships: disclosed no relevant relationships. R.B.D. Activities related to the present article: previously was a co-PI of the Framingham Heart Study; has not received funding since 2013 for this project. Activities not related to the present article: disclosed no relevant relationships. Other relationships: disclosed no relevant relationships. T.M. disclosed no relevant relationships. C.J.O. disclosed no relevant relationships. H.J.W.L.A. Activities related to the present article: received grant funding through the National Institutes of Health. Activities not related to the present article: received grant funding through the National Institutes of Health and Varian; owns stock in Sphera and Genospace. Other relationships: disclosed no relevant relationships. U.H. Activities related to the present article: received funding through the National Heart, Lung, and Blood Institute and Massachusetts General Hospital. Activities not related to the present article: consultant for Abbot, Duke University, and Recor Medical; provides expert testimony to Kowa, Medimmune, HeartFlow, Duke University (Abbot), Oregon Health & Science University. Other relationships: receives grant funding through The America Heart Association (13FTF16450001), the National Institutes of Health (5R01HL109711), and the National Heart, Lung, and Blood Institute (5K24HL113128, 5T32HL076136, 5U01HL123339); received payment from the RSNA for an RSNA workshop.
Abbreviations:
- AS
- Agatston score
- AUC
- area under the receiver operating characteristic curve
- CAC
- coronary artery calcium
- FHS
- Framingham Heart Study
- FRS
- Framingham risk score
- HR
- hazard ratio
- IQR
- interquartile range
- RS
- radiomic-based score
References
- 1.Detrano R, Guerci AD, Carr JJ, et al. Coronary calcium as a predictor of coronary events in four racial or ethnic groups. N Engl J Med 2008;358(13):1336–1345. [DOI] [PubMed] [Google Scholar]
- 2.Hoffmann U, Massaro JM, Fox CS, Manders E, O’Donnell CJ. Defining normal distributions of coronary artery calcium in women and men (from the Framingham Heart Study). Am J Cardiol 2008;102(9):1136–1141, 1141.e1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Sarwar A, Shaw LJ, Shapiro MD, et al. Diagnostic and prognostic value of absence of coronary artery calcification. JACC Cardiovasc Imaging 2009;2(6):675–688. [DOI] [PubMed] [Google Scholar]
- 4.Agatston AS, Janowitz WR, Hildner FJ, Zusmer NR, Viamonte M Jr, Detrano R. Quantification of coronary artery calcium using ultrafast computed tomography. J Am Coll Cardiol 1990;15(4):827–832. [DOI] [PubMed] [Google Scholar]
- 5.Alluri K, Joshi PH, Henry TS, Blumenthal RS, Nasir K, Blaha MJ. Scoring of coronary artery calcium scans: history, assumptions, current limitations, and future directions. Atherosclerosis 2015;239(1):109–117. [DOI] [PubMed] [Google Scholar]
- 6.Ferencik M, Pencina KM, Liu T, et al. Coronary Artery Calcium Distribution Is an Independent Predictor of Incident Major Coronary Heart Disease Events: Results From the Framingham Heart Study. Circ Cardiovasc Imaging 2017;10(10):e006592. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Blaha MJ, Budoff MJ, Tota-Maharaj R, et al. Improving the CAC Score by Addition of Regional Measures of Calcium Distribution: Multi-Ethnic Study of Atherosclerosis. JACC Cardiovasc Imaging 2016;9(12):1407–1416. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Criqui MH, Denenberg JO, Ix JH, et al. Calcium density of coronary artery plaque and risk of incident cardiovascular events. JAMA 2014;311(3):271–278. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Criqui MH, Knox JB, Denenberg JO, et al. Coronary Artery Calcium Volume and Density: Potential Interactions and Overall Predictive Value: The Multi-Ethnic Study of Atherosclerosis. JACC Cardiovasc Imaging 2017;10(8):845–854. [DOI] [PubMed] [Google Scholar]
- 10.Aerts HJ, Velazquez ER, Leijenaar RT, et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat Commun 2014;5(1):4006. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Coroller TP, Grossmann P, Hou Y, et al. CT-based radiomic signature predicts distant metastasis in lung adenocarcinoma. Radiother Oncol 2015;114(3):345–350. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Rios Velazquez E, Parmar C, Liu Y, et al. Somatic mutations drive distinct imaging phenotypes in lung cancer. Cancer Res 2017;77(14):3922–3930. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Kolossváry M, Kellermayer M, Merkely B, Maurovich-Horvat P. Cardiac Computed Tomography Radiomics: A Comprehensive Review on Radiomic Techniques. J Thorac Imaging 2018;33(1):26–34. [DOI] [PubMed] [Google Scholar]
- 14.Zhou Y, He L, Huang Y, et al. CT-based radiomics signature: a potential biomarker for preoperative prediction of early recurrence in hepatocellular carcinoma. Abdom Radiol (NY) 2017;42(6):1695–1704. [DOI] [PubMed] [Google Scholar]
- 15.D’Agostino RB Sr, Vasan RS, Pencina MJ, et al. General cardiovascular risk profile for use in primary care: the Framingham Heart Study. Circulation 2008;117(6):743–753. [DOI] [PubMed] [Google Scholar]
- 16.Kannel WB, Feinleib M, McNamara PM, Garrison RJ, Castelli WP. An investigation of coronary heart disease in families. The Framingham offspring study. Am J Epidemiol 1979;110(3):281–290. [DOI] [PubMed] [Google Scholar]
- 17.Splansky GL, Corey D, Yang Q, et al. The Third Generation Cohort of the National Heart, Lung, and Blood Institute’s Framingham Heart Study: design, recruitment, and initial examination. Am J Epidemiol 2007;165(11):1328–1335. [DOI] [PubMed] [Google Scholar]
- 18.Hoffmann U, Massaro JM, D’Agostino RB Sr, Kathiresan S, Fox CS, O’Donnell CJ. Cardiovascular Event Prediction and Risk Reclassification by Coronary, Aortic, and Valvular Calcification in the Framingham Heart Study. J Am Heart Assoc 2016;5(2):e003144. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Fedorov A, Beichel R, Kalpathy-Cramer J, et al. 3D Slicer as an image computing platform for the Quantitative Imaging Network. Magn Reson Imaging 2012;30(9):1323–1341. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Wu FZ, Wu MT. 2014 SCCT guidelines for the interpretation and reporting of coronary CT angiography: a report of the Society of Cardiovascular Computed Tomography Guidelines Committee. J Cardiovasc Comput Tomogr 2015;9(2):e3. [DOI] [PubMed] [Google Scholar]
- 21.Moselewski F, O’Donnell CJ, Achenbach S, et al. Calcium concentration of individual coronary calcified plaques as measured by multidetector row computed tomography. Circulation 2005;111(24):3236–3241. [DOI] [PubMed] [Google Scholar]
- 22.Detrano, R. et al. Coronary Calcium as a Predictor of Coronary Events in Four Racial or Ethnic Groups. 1336–1345 (2016). [DOI] [PubMed] [Google Scholar]
- 23.van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res 2017;77(21):e104–e107. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Grønnesby JK, Borgan O. A method for checking regression models in survival analysis based on the risk score. Lifetime Data Anal 1996;2(4):315–328. [DOI] [PubMed] [Google Scholar]
- 25.Robin X, Turck N, Hainard A, et al. pROC: an open-source package for R and S+ to analyze and compare ROC curves. BMC Bioinformatics 2011;12(1):77. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Heagerty PJ, Zheng Y. Survival model predictive accuracy and ROC curves. Biometrics 2005;61(1):92–105. [DOI] [PubMed] [Google Scholar]
- 27.Budoff MJ, Shaw LJ, Liu ST, et al. Long-term prognosis associated with coronary calcification: observations from a registry of 25,253 patients. J Am Coll Cardiol 2007;49(18):1860–1870. [DOI] [PubMed] [Google Scholar]
- 28.Brown SJ, Hayball MP, Coulden RAR. Impact of motion artefact on the measurement of coronary calcium score. Br J Radiol 2000;73(873):956–962. [DOI] [PubMed] [Google Scholar]
- 29.Horiguchi J, Fukuda H, Yamamoto H, et al. The impact of motion artifacts on the reproducibility of repeated coronary artery calcium measurements. Eur Radiol 2007;17(1):81–86. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.