Prognostic Utility of Coronary CT Angiography-derived Plaque Information on Long-Term Outcome in Patients With and Without Diabetes mellitus

Purpose To investigate the long-term prognostic value of coronary CT angiography (cCTA)-derived plaque information on major adverse cardiac events (MACE) in patients with and without diabetes mellitus. Methods 64 patients with diabetes (63.3±10.1 years, 66% male) and suspected coronary artery disease (CAD) who underwent cCTA were matched with 297 patients without diabetes according to age, sex, cardiovascular risk factors, statin and antithrombotic therapy. Major adverse cardiac events (MACE) were recorded. cCTA-derived risk scores and plaque measures were assessed. The discriminatory power to identify MACE was evaluated using multivariable regression analysis and concordance indices (CIs). Results After a median follow-up of 5.4 years, MACE occurred in 31 patients (8.6%). In patients with diabetes, cCTA risk scores and plaque measures were signicantly higher compared to non-diabetic patients (all p<0.05). The following plaque measures were predictors of MACE using multivariable Cox regression analysis (hazard ratio [HR]) in patients with diabetes: segment stenosis score (HR 1.20, p<0.001), low-attenuation plaque (HR 3.47, p=0.05), and in non-diabetic patients: segment stenosis score (HR 1.92, p<0.001), Agatston score (HR 1.0009, p=0.04), and low-attenuation plaque (HR 4.15, p=0.04). A multivariable model showed signicantly improved C-index of 0.96 (95% CI 0.94-0.0.97) for MACE prediction, when compared to single measures alone. cCTA-derived stratication Data presented as medians with 25th and 75th percentile, mean ± standard deviation or percentages in parentheses (%). CAD = coronary artery disease. *Dened as blood pressure > 140 mmHg systolic, > 90 mmHg diastolic, or use of antihypertensive medication; † Dened as a total cholesterol of > 200mg/dl or use of anti-lipidemic medication.


Introduction
Coronary CT angiography (cCTA) is a well-established modality for the assessment of coronary artery disease (CAD) and non-invasive plaque quanti cation [1][2][3]. Recent studies have demonstrated the predictive value of coronary plaque measures (i.e. extent, composition, location) for improved risk strati cation [4,5].
Diabetes mellitus (DM) is a well-known cardiovascular risk factor for CAD with increased rates in morbidity and mortality [6]. Patients with diabetes present overall higher coronary plaque burden and are at higher risk for adverse cardiovascular events when compared to non-diabetic patients [7]. However, conventional risk scores recommended by societal guidelines are often challenging to apply effectively in patients with diabetes and do not necessarily meet the required prevention care of cardiovascular disease in this speci c patient population [8]. Recent investigations have demonstrated improved risk strati cation in diabetic patients provided by cCTA-derived risk scores and plaque quanti cation [9,10].
However, the impact of CT scores, especially high-risk plaque features, on major adverse cardiovascular events (MACE) in patients with diabetes has yet to be investigated.
Thus, we sought to evaluate the long-term prognostic value of cCTA-derived coronary plaque information on MACE in patients with and without diabetes mellitus.

Study population
This retrospective single-center study was approved by the institutional review board, and the need for written informed consent was waived due to the retrospective nature of this investigation. The study was performed in compliance with HIPAA. All consecutive patients with suspected or known CAD who underwent 64-slice cCTA as part of their clinical workup between April 2009 and October 2013 were included. 64 patients with diabetes mellitus were matched with 297 patients without diabetes according to the following parameters: age, sex, cardiovascular risk factors, statin and antithrombotic therapy. Portions of this patient population have been reported in a prior study [11]. However, the prior study focused on outcome prediction in a general patient population using machine learning principles, whereas the present study aims to investigate cCTA-derived plaque measures for risk strati cation in patients with and without diabetes.
MACE were recorded on follow-up. MACE were de ned as cardiac death (fatal myocardial infarction [MI]), non-fatal MI (ST-segment elevation [STEMI] and non-ST-segment elevation MI [NSTEMI]), and unstable angina leading to coronary revascularization (percutaneous coronary intervention [PCI] or coronary artery bypass grafting [CABG]), with more than 6 weeks between cCTA and invasive coronary angiography (ICA) with the revascularization procedure [12]. The patients' Framingham risk score was calculated to re ect clinical risk for cardiovascular events [13]. Diabetes was de ned as fasting glucose ≥ 126 mg/dl and/or treatment with insulin or oral hypoglycemic medication [8]. cCTA data with non-diagnostic image quality was excluded from further analysis. Likewise, patients were excluded if they underwent coronary revascularization within six weeks of the CT scan, or had a history of previous MI, PCI or CABG. Demographic and clinical data were collected from medical records.

Coronary CTA acquisition
A sixty-four-slice CT (Philips Brilliance 64, Philips Medical, Eindhoven, Netherlands) was used for image acquisition. Patients initially underwent a non-contrast enhanced calcium scoring scan (collimation, 32 x 1.2 mm; 120 kV tube voltage; tube current, 75 mA; 3 mm slice thickness with 1.5 mm increment). A retrospectively ECG-gated protocol in spiral technique was used for the subsequent contrast-enhanced cCTA with the following scan parameters: 120 kV, 600 mAs, temporal resolution of 165 ms, collimation 64 x 0.6 mm. Contrast enhancement was achieved by injecting 50-80 ml contrast agent at 4-6 ml/sec followed by a 30 ml saline bolus chaser. Weighted ltered back projection image reconstruction was performed in the cardiac phase with the least motion: section thickness of 0.75 mm, reconstruction increment of 0.5 mm, and a smooth convolution kernel.
2.3 Analysis of cCTA data, cCTA scores and plaque measures cCTA datasets were analyzed on a commercially available post-processing software (Philips Medical, Eindhoven, The Netherlands). Two observers, who were blinded to the patients' history, independently analyzed the lesion characteristics. All discordant cases were resolved by consensus. Transverse sections and automatically generated curved multiplanar reformations were used as the reference for diameter and area stenosis quanti cation. Average dimensions of non-affected vessel segments immediately proximal and distal to the lesion of interest were measured at points free of atherosclerotic plaque. The CAD-RADS™ (coronary artery disease reporting and data system) was used to determine the degree of stenosis: 1. none (0%) or minimal (1-24%), 2. mild (25-49% stenosis), 3. moderate (50-69% stenosis), 4. severe (70-99% stenosis), 5. total occlusion (100%). Obstructive CAD was de ned as ≥ 50% luminal stenosis [14]. A coronary plaque was de ned as a structure > 1 mm 2 located within or adjacent to the coronary artery lumen. Plaques with a CT attenuation of ≤30 Houns eld units (HU) were de ned as lowattenuation plaques [15]. On vessel cross-sections, the presence of positive vessel remodeling was measured as the ratio of the vessel area of the lesion over the proximal luminal reference area. A remodeling index ≥1.1 was de ned as positive vessel remodeling [15,16]. The presence of a positive napkin-ring sign, described as a low attenuating plaque core circumscribed by an area of higher attenuation, was evaluated [17]. Spotty calci cations were visually assessed as calci cations covering < 90° of the vessel circumference while being > 3 mm of length [18]. Segment involvement score and segment stenosis score were calculated as previously reported [19]. Presence of low-attenuation plaque, napkin-ring sign, positive remodeling, and spotty calci cations deemed "high-risk" plaque features were evaluated [20,21].

Statistical Analysis
MedCalc (MedCalc Software, version 15, Ostend, Belgium) and the python module scikit-survival [22] were used for statistical analysis. Continuous variables are displayed as mean ± standard deviation or median with interquartile range when not normally distributed. Normal distribution was assessed using Kolmogorov-Smirnov testing. Student t-test and Mann-Whitney U-test were used for parametric and nonparametric data, respectively. First, predictors of MACE were assessed for patients with and without diabetes using univariable and multivariable Cox proportional hazards analysis with backward elimination based on p-values as selection criterion. The resulting hazard ratios (HR) and 95% con dence intervals were reported. Second, the prognostic discriminatory capacity for predicting events in the overall dataset was evaluated by Cox proportional hazards analysis corrected for Diabetes mellitus, and concordance (C)-indices were determined as proposed by Harrell et al. [23]. To avoid over tting, a multivariable model was built utilizing recursive feature elimination with 5-fold cross-validation and Cindices as performance criterion. Improvement in the prediction performance of MACE was calculated using continuous net reclassi cation improvement (NRI) according to Pencina et al. [24]. Event rates were estimated by Kaplan-Meier curves and compared by log-rank test in patients with and without diabetes according to segment stenosis score (≥10 vs. <10), presence of low-attenuation plaque (yes/no), and high-risk plaque features (≥2 vs. <2). Statistical signi cance was assumed with a p-value ≤0.05.

Patient characteristics
A total of 361 patients were included: 64 diabetic patients (63.3 ± 10.1 years, 66% male) and 297 nondiabetics (61.6 ± 10.4 years, 64% male). MACE occurred after a median follow-up of 5.4 years (IQR 4.9-5.7 years) in 31 patients (8.6%), 21 with diabetes and 10 without diabetes. Baseline characteristics were well balanced in both groups (Table 1). However, Framingham risk score, which includes diabetes as a determining factor, was signi cantly higher in diabetic patients (  Data presented as medians with 25th and 75th percentile, mean ± standard deviation or percentages in parentheses (%). CAD = coronary artery disease. *De ned as blood pressure >140 mmHg systolic, >90 mmHg diastolic, or use of antihypertensive medication; † De ned as a total cholesterol of >200mg/dl or use of anti-lipidemic medication.
Diabetes was treated with oral antidiabetic medication in 47 patients (73%) and with insulin in 17 patients (27%  Quantitative analysis of coronary CT-derived markers in patients with diabetes and without diabetes.

Association of cCTA scores and plaque measures with MACE
Patients with diabetes who suffered MACE yielded higher cCTA scores and prevalence of plaque measures (Table 3). Similar results were demonstrated for patients without diabetes. Results of the univariable and multivariable Cox regression analysis are displayed in Table 4 and    Data presented as medians with 25th and 75th percentile or numbers with percentages (%). CAD = coronary artery disease.
Quantitative analysis of coronary CT-derived markers in patients with diabetes and without diabetes according to MACE.   Data presented as medians with 25th and 75th percentile or numbers with percentages (%); TPV = total plaque volume, CPV = calci ed plaque volume, NCPV = non-calci ed plaque volume, SSS = segment stenosis score, SIS = segment involvement score.
Univariate Cox proportional hazards regression analysis of coronary CT-derived markers in patients with diabetes and without diabetes for the prediction of MACE. ROC analysis demonstrated that the multivariable model including diabetes mellitus, segment stenosis score, segment involvement score, low-attenuation plaque, and positive remodeling resulted in a signi cantly improved C-index of 0.96 (95% CI 0.94-0.97) for MACE prediction, when compared to these parameters alone: low-attenuation plaque: C-index 0.82 (95% CI 0.75-0.90, p < 0.001), and ≥2 high-risk features: C-index 0.86 (95% CI 0.80-0.93, p = 0.003). Segment stenosis score yielded a C-index of 0.94 (95% CI 0.92-0.96, p = 0.049), comparable to that of the multivariable model (Fig. 1).
To assess the ability of appropriate reclassi cation of patient risk for MACE, the NRI was calculated. The NRI of the multivariable model compared to segment stenosis score was 0.45 (95% CI 0.08-0.81), 0.28 (95% CI 0.004-0.51) when compared to low-attenuation plaque, and 0.29 (95% CI 0.004-0.54) when compared to ≥2 high-risk features.
The Kaplan-Meier survival curves showed that patients with diabetes and addition of one of the following characteristics (≥ 2 high-risk plaque features, presence of low-attenuation plaque, or segment stenosis score > 10) had substantially higher event rates than patients without these ndings (p < 0.001) (Fig. 2). A case example of coronary stenosis on cCTA with corresponding high-risk plaque feature is shown in Fig. 3.

Discussion
The present study assessed the long-term prognostic value of cCTA-derived plaque information on MACE in patients with and without diabetes mellitus. Our results demonstrate the predictive value of cCTA measures for MACE in patients with diabetes, with segment stenosis score (HR 1.20, p < 0.001) and lowattenuation plaque (HR 3.47, p = 0.05), showing predictive power beyond cCTA stenosis grading and the Framingham risk score. These markers portend improved risk strati cation in both patients with and without diabetes. Additionally, a multivariable model showed a signi cantly improved C-index of 0.96 (95% CI 0.94-0.97) for MACE prediction, when compared to single measures alone (all p < 0.05).
Several prior studies have evaluated the prognostic value of cCTA-derived plaque information using semiautomatic plaque quanti cations and CT scores in patients with diabetes compared to patients without diabetes [7,9,10]. A recent study by van Hoogen et al. [9] and Hadamitzky et al. [25] showed that plaque related scores (i.e. segment stenosis score/segment involvement score) were signi cantly different in patients with and without diabetes and demonstrated predictive value. We also demonstrated a signi cant difference in CT scores between diabetic and non-diabetic patients (median segment  (Fig. 2), which is in line with ndings by Blanke et al. [27].
A major drawback is the necessity for manually performed time-consuming plaque analysis that hampers its applicability in a real-world clinical setting. Although semiautomatic plaque software has been used in prior investigations [1,26,28], manual adjustment is still required. Technical advances such as machinelearning applications may reduce this limitation and allow for improved risk strati cation in a timely and cost-effective manner [29,30].
This study has several limitations that deserve mentioning.
A relatively small number of patients with different types of diabetes as well as various medical therapies were included, which may incur selection bias.
Furthermore, the number of MACE in non-diabetic patients is very small compared to diabetics, which may be explained be the retrospective study design resulting in selection bias. Therefore, prospective studies on larger study cohorts are necessary to validate our ndings. Our results on multivariable analysis may be underpowered by the limited number of observations per variable included [31]. Therefore, the data generated in this study should only be considered hypothesis generating. Patient follow-up was performed using electronic medical records of the hospitals; potentially resulting in missed events that may have occurred outside the hospital system.
In conclusion, this study demonstrates that the presence of diabetes is associated with a signi cantly higher extent of CAD and plaque features, which have independent predictive values for MACE. cCTAderived plaque information portends improved risk strati cation of patients with diabetes beyond assessment of obstructive stenosis on cCTA alone.