The WARRIOR ( W omen’s Ischemi A T R eatment R educes Events I n Non- O bst R uctive CAD) trial (NCT03417388) enrolled symptomatic women with nonobstructive CAD—defined as <50% maximal stenosis—as determined by CCTA or invasive coronary angiography in ∼50 U.S. sites. After randomization, medications (atorvastatin or rosuvastatin; lisinopril or losartan; and aspirin) are provided to the IMT group. Follow-up of participants will occur at 3-, 6-, and 12-months following randomization during the 1st year and every 6 months thereafter for an average follow-up period of 2.5 years.

The NHL bisponsored WAT will recruit 200 participants from the WARRIOR trial including 100 IMT and UC each with nonobstructive plaque by entry CCTA. ( Figure 2 ) We will enroll women with coronary CCTA performed from 5 sites (University of Florida [Orlando, Jacksonville, Gainesville], the Lundquist Institute [formerly LA Biomed], and Cedars-Sinai Medical Center-Heart Institute) with strong experience in CCTA. After participant consent, anonymized DICOM image files of CCTA at trial entry will be assessed for image quality and plaque quantification at the Cedars–Sinai core laboratory for eligibility. Eligible women will undergo a second CCTA after 2 years in accordance with the Society of Cardiovascular CT Baseline and follow-up scans will be analyzed side by side through AI-enabled quantitative assessment of serial CCTA images; evaluating plaque burden, plaque composition, contrast density drop (CDD) and PCAT attenuation in patients treated with IMT and UC. Clinical data including fasting lipids, high-sensitivity C-reactive protein (hs-CRP) and the SAQ responses obtained as part of the WARRIOR trial visits will be available at baseline and follow-up for analysis of the ancillary study.

Ancillary study design.

The entire coronary artery tree including all segments with plaque will be analyzed. The number and distribution of plaques will be recorded. An experienced reader blinded to the treatment group will perform the analyses of vessels using multiplanar CCTA images. The proximal and distal limits of plaque will be identified on the baseline and follow-up CCTA viewed alongside each other after which an automated plaque quantification will be performed as previously described Automated vessel wall correction will be applied for fine adjustment of vessel wall outline. For each artery, maximum diameter stenosis (DS), remodeling index, and CDD values will be reported Scan-specific thresholds in HU for noncalcified plaque (NCP) and calcified plaque (CP) will be automatically generated and plaque components will be quantified using adaptive algorithms Low-density noncalcified plaque (LDNCP) will be defined as plaque with attenuation <30 HU. Quantitative DS (%) will be calculated by the software as the ratio between the narrowest lumen diameter and the mean of two healthy, nondiseased reference points. Positive remodeling quantified as remodeling index will be determined as the ratio of maximum vessel area to that at the proximal normal reference point. CDD will be calculated as the maximum percent difference in luminal contrast densities with respect to the proximal reference cross section without disease Maximum and mean CDD per patient will be quantified and all measurements will be automatically exported on a per-patient basis.

PCAT measurement for each CCTA will be performed with the use of an AI-enabled research plaque quantification software APQ. ^, PCAT density will be quantified in the proximal portion of the right coronary artery (RCA). For each measurement, PCAT will be sampled in three-dimensional layers, moving radially outwards from the outer vessel wall of the RCA in 1 mm increments. Adipose tissue is defined as all voxels with attenuation between −190 and −30 HU, and the mean PCAT density will be defined as the average CT attenuation in HU of the adipose tissue within the defined volume of interest. Mean PCAT density is assessed within an outer radial distance from the vessel wall equal to the average diameter of the vessel (typically 3 mm).

SAQ, an instrument for measuring health status in patients with CAD, is frequently used as an outcome in clinical trials and is endorsed as a performance measure for assessing the quality of CAD care A shorter 7-item version (SAQ7) quantifies symptoms, functional status, and quality of life, and it excludes the treatment satisfaction and angina stability domains An overall summary score will be derived as the mean of the 3 domain scores.

The MLRS will be computed from quantitatively measured CCTA plaque and vessel characteristics along with age, gender, medications, and risk factors Feature selection will be performed on all patient and quantitative CCTA measures, including PCAT CT attenuation, with information gain attribute ranking using stratified tenfold cross-validation. Prediction of recurrent angina will be performed using an ensemble classification approach. We will utilize state-of-the-art LogitBoost and extreme gradient boosting (XGBoost). Recurrent angina will be defined as a SAQ7 score <60 which represents weekly angina Both feature selection and LogitBoost will be conducted with stratified tenfold cross-validation. The data set will be randomly divided into 10 equally sized subsamples, each with the same number of events (recurrent angina). Training will be done in 9/10th of the data and tested on the remaining unseen 1/10th of the data. The MLRS− with continuous values from 0 to 1 will be computed directly by concatenating from all testing data folds and will correspond to the whole cohort. Optimized cutoff for MLRS will be calculated on a per-patient basis using the two-graph receiver operating characteristic ROC analysis as the intersection of the sensitivity and specificity graphs Multivariable models will be created to assess the ability of the MLRS to improve prediction of anginal symptoms compared with other methods. Model 1 will comprise clinical risk assessment including age, symptoms, baseline SAQ7 score, cardiac risk factors, medications, and maximal coronary artery stenosis. Model 2 will add MLRS to the covariates in Model 1 to assess continuous net reclassification improvement We have previously shown the feasibility of this approach to predict all-cause mortality and revascularization from quantitative assessment of CCTA plaque features. ^, Of note, the MLRS is mainly exploratory and intended to derive a prespecified model for external validation.

Bivariate associations between measured variables will be described using correlations or contingency tables. Total plaque volume (TPV), NCPV, LDNCPV, CPV and maximal CDD will be quantified and compared between the two groups at 2 years, adjusting for baseline differences. The association of NCPV, TPV, LDNCPV and CDD with study group at 2-years will be assessed using a linear regression model after adjusting for baseline values, as in ANCOVA. If imbalance between the randomized groups exists, the models will be adjusted for biological covariates such as age or body mass index (BMI) that might confound the results. The assumptions for the regression models will be examined through analysis of residuals. Standard variable selection methods such as stepwise and backward procedures will determine the best subset of covariates to include in the model. Collinearity will be explored using the condition index and after careful assessment of the correlation matrix. Mean PCAT density in HU will be compared between groups. Changes in the volume of PCAT will be examined by ANCOVA, along with linear regression modeling to adjust for potential imbalances in the randomization groups. Change in LDL, blood pressure and hs-CRP from baseline measurements will be compared between groups, the former as a measure of adherence to group assignment.

Change in CCTA variables include ∆TPV, high-risk plaque components (∆NCPV, ∆LDNCPV), ∆CDD and ∆PCAT density. Clinical improvement will be quantified by an increase in SAQ7 score and an increase in SAQ angina frequency domain score. Correlations among change in SAQ scores(∆SAQ), ∆TPV, ∆NCPV, ∆LDNCPV, ∆CDD and ∆PCAT density will be examined. As an exploratory analysis, multiple linear regression will be used to adjust the outcomes of change in SAQ for their baseline values and additional explanatory biological covariates including study group, age, BMI, risk factors and use of other medications and antianginals. The goal of the analysis is to associate changes in plaque with changes in SAQ, rather than detect differences in SAQ between treatment groups.

A mediation analysis will explore if changes in plaque act as a mediator between treatment and SAQ outcomes. Average direct effects and average causal mediation effects will be estimated as described by Imai et al. Pearson correlations will be examined for continuous variables, or Spearman correlations will be used if the distributions are non-normal. A binary indicator of SAQ improvement will be created for each woman based on ∆SAQ > 10 vs ∆SAQ ≤ 10. ∆SAQ > 10 represents a clinically significant improvement in angina. A logistic model will be created with SAQ improvement as the outcome and the explanatory factors already listed.

Association between MLRS and symptoms quantifiable by SAQ, will be evaluated in the entire WARRIOR cohort. Multivariable logistic models with follow-up angina status as the response variable (defining the event of angina as SAQ7 < 60) will be fit with the following set of covariates–(1) baseline SAQ7, (2) MLRS using baseline CCTA plaque and PCAT characterization, (3) baseline SAQ7 + MLRS. Use of antianginal medications will be used as a covariate in these models. Predictive performance will be evaluated with discrimination measures including AUC, Net Reclassification Index at the event rate, calibration measures such calibration-in-large, calibration slope and median absolute deviation as recommended by Harrell. Optimism bias will be corrected using bootstrap sampling with 1,000 replicates. AUC between models will be tested using a two-sided test. We will compare AUC between models with 95% bootstrap confidence intervals for ∆AUC.

Sensitivity analyses will be conducted in the substudy cohort compared to the full trial cohort including analysis with and without adjustment for baseline characteristics, and analysis with different methods of adjusting for baseline imbalance. eg, multivariable regression vs. propensity score method to address baseline group differences.

Based on our power calculations, with 100 patients in each group, we can detect a difference of 75mm3 in NCPV between the IMT and UC groups. A sample size of 200 participants (100 per group) would have 85% power to detect a Pearson correlation of 0.210 or greater between ∆SAQ and change in plaque measures, assuming a type I error rate of 0.05. Pilot data-based detailed power tables are show in the Supplement text and Supplementary Tables 1-3.