U-shaped relationship between coronary artery calcification and underestimation of moderate coronary stenosis by coronary computed tomography angiography

Abstract

Background

Coronary computed tomography angiography (CCTA) is crucial for diagnosing coronary artery disease (CAD). However, its diagnostic accuracy is frequently influenced by factors such as coronary artery calcification (CAC), which can lead to both overestimation and critically, underestimation of stenosis severity. Underestimation of stenosis, particularly in the clinically ambiguous 50–70% range, may delay necessary invasive coronary angiography (ICA). Previous studies have lacked a systematic understanding of the entire spectrum of CAC score (CS) in relation to the risk of CCTA underestimation in this critical 50–70% stenosis range. Therefore, this study aimed to: (I) quantify the non-linear relationship between CS and CCTA underestimation of moderate coronary stenosis; (II) identify optimal CS thresholds for risk stratification of underestimation; and (III) determine clinical and imaging predictors to improve CCTA diagnostic accuracy.

Methods

This retrospective study included 216 patients who underwent both CCTA and ICA. Patients with CCTA-reported 50–70% stenosis were classified into an underestimation group (n=63; ICA stenosis >70%) and a non-underestimation group (n=153; ICA stenosis ≤70%). Demographic, clinical, and laboratory data were systematically collected, and CS was quantified using the Agatston method. Logistic regression and restricted cubic spline (RCS) analyses evaluated the association between CS and CCTA underestimation, identified non-linear relationships, and determined optimal CS cutoffs. Diagnostic performance and subgroup analyses were also conducted.

Results

The underestimation group exhibited significantly higher median CS (P=0.08), higher diabetes prevalence (P=0.007), elevated triglycerides (TG) (P=0.004) and fasting glucose (P=0.03), and lower high-density lipoprotein cholesterol (HDL-C) (P=0.004). RCS analysis revealed a significant U-shaped non-linear relationship between CS and underestimation risk (P-nonlinearity <0.05). Distinct cutoffs were identified: a lower threshold around 118 Agatston units (AU) and an upper threshold between 253 and 330 AU. Patients with CS <118 or >330 AU had the highest underestimation risk, whereas those with intermediate CS (118–330 AU) were prone to overestimation. In the fully adjusted model, higher CS remained an independent predictor of underestimation [adjusted odds ratio (OR) =1.002; 95% confidence interval (CI): 1.000–1.003, P=0.02]. Diagnostic models utilizing these thresholds demonstrated high sensitivity (82.5–87.3%) but limited specificity (24.8–30.7%). Subgroup analyses confirmed robustness, with pronounced effects in females, elderly patients, and those with calcified plaques (P=0.002).

Conclusions

Our study reveals a U-shaped non-linear relationship between CS and CCTA underestimation of moderate stenosis. Patients with CS <118 or >253–330 AU are at higher underestimation risks, while intermediate CS values are tilting towards overestimation. Despite a modest per-unit OR, the cumulative effect across clinically relevant CS ranges may hold clinical significance. Clinically, application of these CS thresholds offers a practical framework to identify patients requiring further ICA, minimizing missed diagnoses in high-risk groups and reducing unnecessary invasive procedures in intermediate-risk patients, thereby facilitating precise, individualized CAD management.

Highlight box.

Key findings

• For coronary computed tomography angiography (CCTA)-reported 50–70% stenosis, coronary artery calcification score (CS) is associated with CCTA underestimation in a non-linear dual-threshold pattern.

• The lower CS cutoff was consistently 118.07 Agatston units (AU) (paper thresholds), while the upper cutoff varied by model (270.87/253.50/329.90 AU).

What is known and what is new?

• Calcification can bias CCTA stenosis assessment, but a clinically actionable dual-threshold pattern across the CS spectrum has been unclear.

• We demonstrate a CS-underestimation dual-threshold relationship supported by restricted cubic spline modeling and diagnostic threshold analysis.

What is the implication, and what should change now?

• CS can be used as an adjunct interpretive reference when clinicians face borderline CCTA stenosis.

• Patients with very low CS (~<118 AU) or very high CS (above the model-dependent upper cutoff) may warrant additional evaluation beyond anatomical CCTA, while intermediate CS may represent a comparatively lower-risk range.

Introduction

Coronary artery disease (CAD) remains a leading cause of morbidity and mortality worldwide, necessitating accurate and timely diagnosis for optimal patient management. Coronary computed tomography angiography (CCTA) has emerged as a cornerstone non-invasive modality for CAD screening and diagnosis, offering high diagnostic accuracy for ruling out significant stenosis (1,2). However, its performance can be significantly impacted by several factors, notably coronary artery calcification (CAC), which is prevalent in patients with CAD and can compromise image quality and lumen assessment (3,4).

While CAC is known to affect CCTA accuracy, contributing to both overestimation (due to blooming artifacts or partial volume effects) and critically, underestimation of stenosis severity (5,6), underestimation poses a significant clinical challenge. Misinterpreting a hemodynamically significant stenosis as moderate or mild can lead to delayed invasive coronary angiography (ICA), suboptimal medical management, and increased risk of adverse cardiovascular events. This concern is particularly pronounced for patients with CCTA-reported stenosis in the 50–70% range, often termed the ‘borderline’ or ‘intermediate’ zone (7), where diagnostic discordance carries substantial implications for subsequent clinical decision-making. Accurate evaluation in this range is paramount; it can avert unnecessary ICA procedures and facilitate optimized non-invasive management (8,9), yet conversely, underestimation can significantly worsen patient outcomes.

Although the general influence of CAC-related artifacts on CCTA performance is acknowledged, a systematic understanding of the entire spectrum of CAC score (CS) in relation to the risk of CCTA underestimation in this critical 50–70% stenosis range remains elusive. Specifically, current literature lacks robust evidence demonstrating a reproducible non-linear relationship (e.g., a U-shaped pattern) between CS and underestimation risk, nor has it derived quantifiable, clinically actionable CS thresholds that could guide CCTA interpretation. Such thresholds are crucial for identifying patients at high risk for underestimation and warranting further investigation, thereby facilitating more precise and individualized patient management.

Therefore, this study aimed to (I) systematically characterize the relationship between CS and the underestimation of moderate coronary stenosis by CCTA using non-linear modeling; (II) derive clinically interpretable CS cutoff values that delineate distinct risk profiles for underestimation; and (III) explore associated clinical and imaging predictors to enhance the diagnostic precision of CCTA. We present this article in accordance with the STROBE reporting checklist (available at https://cdt.amegroups.com/article/view/10.21037/cdt-2026-0118/rc).

Methods

Study design and participants

This retrospective cohort study included 216 consecutive patients admitted to the Department of Cardiology at Beijing Friendship Hospital between January 2012 and December 2022. All patients presented with initial chest pain or discomfort and underwent both CCTA and ICA within 90 days. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of Beijing Friendship Hospital (approval No. 2018-P2-030-01). Written informed consent for the use of clinical data was formally waived for this retrospective analysis by the Ethics Committee (as approved).

CCTA and ICA procedures

CCTA was performed using a standardized protocol on various multi-detector CT scanners (e.g., 64-slice, 128-slice, or 256-slice from different manufacturers). CS was calculated using the Agatston method, which quantifies plaque burden in coronary arteries by area and density (10). ICA was performed via radial or femoral access. ICA angiograms were independently interpreted by two experienced interventional cardiologists, who were blinded to CCTA findings. Intracoronary nitroglycerin was administered prior to assessment to minimize vasospasm. In cases of disagreement, a consensus reading by a third senior cardiologist was sought.

Inclusion and exclusion criteria

Patients were eligible for inclusion if they: (I) were aged between 18 and 80 years; (II) presented with initial chest pain or discomfort as the primary reason for admission; (III) underwent both CCTA and ICA within 90 days; and (IV) had at least one coronary artery stenosis ranging from 50% to 70% in any major epicardial vessel [left main (LM), left anterior descending (LAD), left circumflex (LCX), or right coronary artery (RCA)] on CCTA.

Exclusion criteria included: (I) hemodynamic instability upon admission (e.g., cardiogenic shock or severe arrhythmias); (II) a diagnosis of classic myocardial infarction (ST-elevation or non-ST-elevation myocardial infarction) at the time of presentation; (III) CCTA-diagnosed coronary stenosis exceeding 70%; (IV) severe systemic diseases such as advanced liver failure, severe renal failure (estimated glomerular filtration rate <30 mL/min/1.73 m²), or active malignancies; and (V) patients with missing or incomplete CCTA or ICA data.

As illustrated in the patient inclusion flowchart (Figure 1), from an initial pool of 726 patients, 478 were excluded due to ineligible coronary stenosis (i.e., not within the 50–70% CCTA range). An additional 2 patients were excluded due to missing CCTA or ICA data, 13 patients due to the time interval between CCTA and ICA exceeding 90 days, and 17 patients due to missing biochemical data. Ultimately, 216 patients were included in the final analysis.

Figure 1.

Flowchart of patient inclusion and exclusion. Patients were screened from those who underwent both CCTA and ICA over a 10-year period (n=726). Among them, 493 patients were excluded based on predefined inclusion and exclusion criteria (stenosis not in 50–70% CCTA range: n=478; missing CCTA/ICA data: n=2; CCTA–ICA interval >90 days: n=13). After exclusion, 233 patients with CCTA-reported 50–70% stenosis and complete imaging data remained. Additional exclusions were applied for patients with missing biochemical values (n=17). Ultimately, 216 patients were included in the clinical analyses: 63 patients in the underestimated group and 153 patients in the non-underestimated group. Underestimation was defined as: ICA stenosis >70% with CCTA stenosis in the 50–70% range, whereas non-underestimation was defined as ICA stenosis ≤70%. ALT, alanine aminotransferase; APTT, activated partial thromboplastin time; AST, aspartate aminotransferase; BUN, blood urea nitrogen; CCTA, coronary computed tomography angiography; Cr, creatinine; HDL-C, high-density lipoprotein cholesterol; ICA, invasive coronary angiography; LAD, left anterior descending; LCX, left circumflex; LDL-C, low-density lipoprotein cholesterol; LM, left main; NSTEMI, non-ST-elevation myocardial infarction; PT, prothrombin time; RCA, right coronary artery; STEMI, ST-elevation myocardial infarction; TC, total cholesterol; TG, triglycerides; TNI, troponin I; TNT, troponin T.

Outcome definition

CCTA underestimation was stringently defined as: CCTA-reported stenosis in the 50–70% range, while subsequent ICA demonstrated a stenosis severity of >70% in the same coronary segment. Non-underestimation was defined as: CCTA-reported stenosis in the 50–70% range, with ICA demonstrating a stenosis severity of ≤70% in the same coronary segment. Figure 2 displays representative CCTA and ICA images demonstrating CCTA underestimation.

Figure 2.

Typical CCTA and angiography images demonstrating underestimation in CCTA patients. A representative case illustrates CCTA underestimation of coronary stenosis severity. In this patient, CCTA showed a CS of 604.74 AU and detected significant stenosis in the mid-segment of the LAD artery. The bright linear structures in the CCTA image represent calcified plaques. The red arrow indicates a calcified plaque in the mid-portion. The CCTA report estimated the stenosis severity to be within 50–70%; however, visual assessment suggests that the stenosis in the mid-LAD artery may exceed 70%. ICA confirmed that this patient represents a typical scenario of underestimation by CCTA. AU, Agatston units; CCTA, coronary computed tomography angiography; CS, coronary artery calcification score; ICA, invasive coronary angiography; LAD, left anterior descending.

Data collection and statistical analysis

Demographic and comprehensive clinical data were extracted from hospital medical records, including: sex, age, plaque type (categorized as calcified plaque, non-calcified plaque, and mixed plaque) (11), chest pain status, hypertension (HTN), diabetes mellitus (DM) (12), hyperlipidemia, smoking habits, alcohol consumption, body mass index (BMI), systolic blood pressure (SBP), diastolic blood pressure (DBP), liver and kidney function markers [alanine aminotransferase (ALT), aspartate aminotransferase (AST), creatinine (Cr), blood urea nitrogen (BUN)], glucose levels, lipid profiles [triglycerides (TG), total cholesterol (TC), low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C)], inflammatory markers [high-sensitivity C-reactive protein (hsCRP)], cardiac biomarkers [troponin T (TNT), troponin I (TNI)], and coagulation parameters [activated partial thromboplastin time (APTT), prothrombin time (PT), and D-dimer]. The primary outcome variable for this study was the presence of CCTA underestimation.

Baseline comparisons

Continuous variables were assessed for normality. Normally distributed data were presented as mean ± standard deviation (SD) and compared using independent sample t-tests. Non-normally distributed data were presented as median [interquartile range (IQR)] and compared using Mann-Whitney U tests. Categorical variables were presented as counts and percentages and compared using Chi-squared tests or Fisher’s exact tests as appropriate.

CS tertile stratification

Patients were categorized into three groups based on tertiles of their CS distribution. Tertile cutoffs were determined directly from the dataset as: T1 [0.00–7.64 Agatston units (AU)], T2 (69.14–144.07 AU), T3 (268.32–533.07 AU). One-way analysis of variance (ANOVA) or Kruskal-Wallis H tests were employed to compare clinical characteristics across these CS tertiles.

Logistic regression analysis

To systematically examine the association between CS and CCTA underestimation, three progressively adjusted logistic regression models were constructed:

• ❖ Model 1: unadjusted, including only CS as the independent variable.

• ❖ Model 2: adjusted for age and sex.

• ❖ Model 3: additionally adjusted for a comprehensive set of clinical confounders, including plaque type, HTN, DM, hyperlipidemia, smoking, alcohol consumption, BMI, ALT, AST, Cr, BUN, TG, TC, LDL-C, HDL-C, TNT, TNI, and D-dimer.

Odds ratios (ORs) and their corresponding 95% confidence intervals (CIs) were reported per 1 AU increase in CS. Collinearity among independent variables was assessed using the variance inflation factor (VIF). Variables with VIF >5 were considered for removal to ensure model robustness (Table S1,S2).

Restricted cubic spline (RCS) analysis

RCS models were applied to rigorously assess and visualize potential non-linear relationships between CS and the odds of underestimation. Three knots were placed at the 10^th, 50^th, and 90^th percentiles of the CS distribution. The fitted curve was expressed relative to a reference at OR =1. Model-based CS cutoff points were derived from the change points of the RCS curves, where the direction of the association shifted.

Diagnostic performance

The diagnostic performance of CS thresholds derived from RCS analysis was evaluated by calculating sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) for CCTA underestimation. The Youden index (sensitivity + specificity − 1) was calculated as an overall measure of diagnostic effectiveness.

Subgroup analyses

Subgroup analyses were conducted to evaluate whether the association between CS and CCTA underestimation differed across strata defined by key demographic and clinical variables. Effect modification was assessed via interaction tests.

All statistical analyses were performed using R statistical software (version 4.4.1). Two-sided tests were applied, with statistical significance defined as P<0.05. The R packages “table1”, “rms”, “ggplot2”, “visdat”, and “mice” were utilized for data analysis and visualization.

Table 1. Baseline clinical characteristics in patients with vs. without CCTA underestimation.

Variables	Non-underestimated (n=153)	Underestimation (n=63)	Statistic	P value	Test method
CS (AU)	96.67 (5.17–237.85)	108.88 (21.05–376.98)	4,078.000	0.08	Mann-Whitney U
Age (years)	63.00 (58.00–68.00)	63.00 (58.50–68.50)	4,685.500	0.75	Mann-Whitney U
Sex			1.86	0.02	Chi-squared test
Female	56 (36.6)	17 (27.0)
Male	97 (63.4)	46 (73.0)
Plaque type			2.12	0.27	Chi-squared test
Calcified plaque	40 (26.1)	11 (17.5)
Non-calcified plaque	31 (20.3)	11 (17.5)
Mixed plaque	82 (53.6)	41 (65.1)
HTN	105 (68.6)	50 (79.4)	2.53	0.15	Chi-squared test
DM	40 (26.1)	29 (46.0)	7.91	0.007	Chi-squared test
Hyperlipidemia	76 (49.7)	37 (58.7)	1.44	0.29	Chi-squared test
Smoking	84 (54.9)	36 (57.1)	0.09	0.88	Chi-squared test
Alcohol	78 (51.0)	28 (44.4)	0.76	0.47	Chi-squared test
BMI (kg/m2)	25.10 (23.12–27.49)	25.94 (23.50–28.06)	4,269.000	0.27	Mann-Whitney U
SBP (mmHg)	130.00 (119.00–138.00)	131.00 (117.50–139.00)	4,797.000	0.96	Mann-Whitney U
DBP (mmHg)	74.65±10.24	75.63±10.49	−0.633	0.53	T-test
HR (bpm)	70.00 (63.00–76.00)	72.00 (66.00–81.00)	4,073.000	0.07	Mann-Whitney U
ALT (U/L)	17.00 (13.00–24.00)	18.00 (14.50–24.50)	4,435.500	0.36	Mann-Whitney U
AST (U/L)	19.00 (16.30–23.80)	18.90 (17.20–23.35)	4,863.000	0.92	Mann-Whitney U
Cr (mmol/L)	72.10 (64.10–81.60)	71.40 (63.85–78.55)	5,162.000	0.41	Mann-Whitney U
BUN (mmol/L)	5.19 (4.56–6.03)	5.32 (4.53–6.20)	4,793.500	0.95	Mann-Whitney U
TG (mmol/L)	1.07 (0.89–1.48)	1.37 (0.98–2.04)	3,220.500	0.005	Mann-Whitney U
TC (mmol/L)	3.85 (3.35–4.73)	3.90 (3.27–4.65)	4,675.000	0.61	Mann-Whitney U
LDL-C (mmol/L)	2.28 (1.92–2.84)	2.23 (1.89–2.78)	4,569.000	0.52	Mann-Whitney U
HDL-C (mmol/L)	1.10 (0.95–1.32)	0.97 (0.85–1.17)	5,632.500	0.004	Mann-Whitney U
TNT (ng/mL)	0.01 (0.01–0.01)	0.01 (0.01–0.01)	4,329.500	0.70	Mann-Whitney U
TNI (ng/mL)	0.00 (0.00–0.00)	0.00 (0.00–0.01)	3518.500	0.002	Mann-Whitney U
D-dimer (mg/L)	0.20 (0.13–0.40)	0.19 (0.10–0.43)	4,754.500	0.60	Mann-Whitney U
hsCRP (mg/L)	0.95 (0.41–3.09)	0.80 (0.34–1.73)	4,857.500	0.17	Mann-Whitney U
Glucose (mmol/L)	5.25 (4.73–5.88)	5.49 (4.96–6.75)	3,810.500	0.03	Mann-Whitney U
APTT (s)	27.40 (25.60–30.60)	28.10 (25.90–30.20)	4,433.500	0.78	Mann-Whitney U
PT (s)	11.20 (10.80–11.70)	11.30 (10.90–11.60)	4,384.000	0.69	Mann-Whitney U

Values are presented as median (IQR) for non-normally distributed variables and mean ± SD for normally distributed variables, consistent with the normality flag in the analysis output. Statistical testing: Mann-Whitney U test for non-normally distributed variables and t-test for normally distributed variables. Categorical variables are summarized as n (%) and compared using the Chi-squared test (reported as the Chi-squared statistic and P value). Statistic is the test statistic for the selected test. P value is two-sided and obtained from the corresponding test. CCTA underestimation was defined as ICA stenosis >70% with CCTA stenosis in the 50–70% range. Non-underestimated was defined as ICA stenosis ≤70%. ALT, alanine aminotransferase; APTT, activated partial thromboplastin time; AST, aspartate aminotransferase; AU, Agatston units; BMI, body mass index; BUN, blood urea nitrogen; CCTA, coronary computed tomography angiography; Cr, creatinine; CS, coronary artery calcification score; DBP, diastolic blood pressure; DM, diabetes mellitus; HDL-C, high-density lipoprotein cholesterol; HR, heart rate; hsCRP, high-sensitivity C-reactive protein; HTN, hypertension; ICA, invasive coronary angiography; IQR, interquartile range; LDL-C, low-density lipoprotein cholesterol; PT, prothrombin time; SBP, systolic blood pressure; SD, standard deviation; TC, total cholesterol; TG, triglycerides; TNI, troponin I; TNT, troponin T.

Results

Baseline characteristics (Table 1)

Table 1 summarizes the baseline clinical and demographic characteristics of patients grouped by CCTA underestimation status. Significant differences were observed between the two groups in several key variables. Patients in the underestimation group had a significantly higher median CS [108.88 (21.05–376.98) vs. 96.67 (5.17–237.85) AU, P=0.08], a higher prevalence of DM (46.0% vs. 26.1%, P=0.007). Additionally, TG levels (P=0.004) and fasting glucose levels (P=0.03) were significantly higher in the underestimation group. Conversely, HDL-C levels were significantly lower (P=0.004), and TNI levels were notably higher (P=0.002) in this group. No significant differences were found in other variables, including age, sex, plaque type, HTN, smoking, alcohol consumption, BMI, SBP, DBP, and most other biochemical markers.

Characteristics across CS tertiles (Table 2)

Table 2. Clinical characteristics across CS tertiles.

Variables	T1 (n=72)	T2 (n=72)	T3 (n=72)	Statistic	P value
CS (AU)	0.00 (0.00–7.64)	98.90 (69.14–144.07)	369.19 (268.32–533.07)	192.630	<0.001
Age (years)	60.50 (53.00–66.25)	62.50 (58.00–66.00)	66.50 (61.00–70.25)	15.990	<0.001
Sex				1.56	0.56
Female	21 (29.2)	25 (34.7)	27 (37.5)
Male	51 (70.8)	47 (65.3)	45 (62.5)
Plaque type				85.59	0.001
Calcified plaque	8 (11.1)	21 (29.2)	22 (30.6)
Non-calcified plaque	39 (54.2)	3 (4.2)	0 (0.0)
Mixed plaque	25 (34.7)	48 (66.7)	50 (69.4)
HTN	49 (68.1)	49 (68.1)	57 (79.2)	2.78	0.23
DM	18 (25.0)	20 (27.8)	31 (43.1)	6.83	0.043
Hyperlipidemia	35 (48.6)	40 (55.6)	38 (52.8)	0.72	0.70
Smoking	40 (55.6)	41 (56.9)	39 (54.2)	0.11	0.95
Alcohol	38 (52.8)	34 (47.2)	34 (47.2)	0.55	0.74
BMI (kg/m2)	25.71 (24.27–27.99)	24.64 (22.85–26.99)	25.98 (22.65–27.98)	3.590	0.17
SBP (mmHg)	126.00 (115.00–138.00)	130.00 (119.75–138.75)	132.50 (121.75–139.25)	3.070	0.22
DBP (mmHg)	73.79±10.35	76.44±10.55	74.57±9.95	1.270	0.28
HR (bpm)	68.50 (63.00–77.00)	70.50 (63.00–75.50)	70.50 (64.75–81.00)	1.060	0.59
ALT (U/L)	17.00 (13.00–22.00)	18.00 (14.00–24.25)	17.00 (13.00–25.00)	0.830	0.66
AST (U/L)	18.25 (15.82–22.73)	19.45 (16.03–22.40)	19.35 (17.88–26.00)	6.140	0.047
Cr (mmol/L)	70.95 (62.00–80.35)	72.15 (64.65–81.65)	72.05 (65.20–79.50)	0.520	0.77
BUN (mmol/L)	5.41 (4.81–6.14)	5.20 (4.58–5.87)	5.03 (4.04–6.09)	1.890	0.39
TG (mmol/L)	1.14 (0.95–1.59)	1.11 (0.92–1.55)	1.16 (0.89–1.73)	0.060	0.97
TC (mmol/L)	3.78 (3.35–4.75)	3.92 (3.37–4.88)	4.06 (3.26–4.56)	0.640	0.73
LDL-C (mmol/L)	2.18 (1.89–2.77)	2.38 (1.91–2.92)	2.29 (1.93–2.78)	0.780	0.68
HDL-C (mmol/L)	1.06 (0.86–1.25)	1.08 (0.96–1.26)	1.07 (0.87–1.31)	0.820	0.67
TNT (ng/mL)	0.01 (0.01–0.01)	0.01 (0.01–0.01)	0.01 (0.01–0.01)	0.810	0.67
TNI (ng/mL)	0.00 (0.00–0.00)	0.00 (0.00–0.00)	0.00 (0.00–0.01)	11.200	0.004
D-dimer (mg/L)	0.19 (0.09–0.32)	0.21 (0.18–0.42)	0.20 (0.13–0.42)	3.420	0.18
hsCRP (mg/L)	0.93 (0.53–2.35)	0.93 (0.41–2.28)	0.79 (0.33–3.11)	0.670	0.71
glucose (mmol/L)	5.37 (4.85–5.98)	5.28 (4.66–6.02)	5.36 (4.83–6.75)	0.720	0.70
APTT (s)	29.10 (25.40–31.60)	27.50 (25.60–29.70)	27.40 (25.95–29.60)	1.790	0.41
PT (s)	11.50 (11.00–11.80)	11.30 (10.70–11.65)	11.10 (10.85–11.60)	4.820	0.09
eGFR (mL/min/1.73 m2)	84.46 (84.46–84.46)	76.91 (76.91–76.91)	N/A	1.000	0.32

Continuous variables are presented as mean ± SD if normally distributed, otherwise as median (IQR). Categorical variables are summarized as n (%) and compared using the Chi-squared test (reported as the Chi-squared statistic and P value). CS tertiles were defined as T1 (low), T2 (middle), and T3 (high), with T1, T2, and T3 each including n=72 patients (overall n=216). Statistic and P value are derived from one-way group comparisons across the three CS tertiles. For variables analyzed using ANOVA (approximately normal distribution), the table reports ANOVA F statistic and the corresponding ANOVA P value. For variables analyzed using Kruskal-Wallis (non-normal distribution), the table reports Kruskal-Wallis H statistic and the corresponding Kruskal-Wallis P value. For variables with N/A due to missing data, group comparison was performed using available observations only, and the statistic/P value corresponds to those available samples. ALT, alanine aminotransferase; ANOVA, analysis of variance; APTT, activated partial thromboplastin time; AST, aspartate aminotransferase; AU, Agatston units; BMI, body mass index; BUN, blood urea nitrogen; Cr, creatinine; CS, coronary artery calcification score; DBP, diastolic blood pressure; DM, diabetes mellitus; eGFR, estimated glomerular filtration rate; HDL-C, high-density lipoprotein cholesterol; HR, heart rate; hsCRP, high-sensitivity C-reactive protein; HTN, hypertension; IQR, interquartile range; LDL-C, low-density lipoprotein cholesterol; N/A, not available; PT, prothrombin time; SBP, systolic blood pressure; SD, standard deviation; TC, total cholesterol; TG, triglycerides; TNI, troponin I; TNT, troponin T.

Table 2 presents the clinical characteristics of patients stratified by CS tertiles. Significant differences were identified among the groups in several variables. Patients in the highest CS tertile (T3) were significantly older (P<0.001) and exhibited a higher prevalence of calcified plaques (30.6% in T3 vs. 11.1% in T1, P=0.001) and DM (43.1% in T3 vs. 25.0% in T1, P=0.043). Furthermore, the proportion of mixed plaques progressively increased across tertiles (69.4% in T3 vs. 34.7% in T1, P=0.001). No significant differences were observed in other variables, including sex, HTN, hyperlipidemia, smoking, alcohol consumption, BMI, blood pressure, and most biochemical markers, except for TNI, which was significantly elevated in the higher CS tertile (P=0.001). The distribution of CS by underestimation status is further illustrated in Figure S1, and the underestimation rate by CS tertile is shown in Figure S2.

Logistic regression analysis (Tables 3,4)

Table 3. Logistic regression between CS and CCTA underestimation.

Model	Variables	OR	95% CI lower	95% CI upper	P value	Number
Model 1	CS	1.002	1.001	1.003	0.004	216
Model 2	CS	1.002	1.001	1.004	0.002	216
Model 3	CS	1.002	1.000	1.003	0.053	194

Outcome: CCTA underestimation (coded as 1 for underestimation and 0 for non-underestimation). Model 1: unadjusted, including only CS as the independent variable. Model 2: adjusted for age and sex. Model 3: additionally adjusted for a comprehensive set of clinical confounders, including plaque type, HTN, DM, hyperlipidemia, smoking, alcohol consumption, BMI, ALT, AST, Cr, BUN, TG, TC, LDL-C, HDL-C, TNT, TNI, and D-dimer. “OR” is reported per 1 AU increase in CS. “Number” reflects the sample size used in each model (based on available data after covariate inclusion). ALT, alanine aminotransferase; AST, aspartate aminotransferase; AU, Agatston units; BMI, body mass index; BUN, blood urea nitrogen; CCTA, coronary computed tomography angiography; CI, confidence interval; Cr, creatinine; CS, coronary artery calcification score; DM, diabetes mellitus; HDL-C, high-density lipoprotein cholesterol; HTN, hypertension; LDL-C, low-density lipoprotein cholesterol; OR, odds ratio; TC, total cholesterol; TG, triglycerides; TNT, troponin T; TNI, troponin I.

Table 4. Collinearity (sensitivity) analysis for CS association with CCTA underestimation.

Model	Patients, n	CS OR (per 1 AU)	95% CI lower	95% CI upper	P value	VIF issue
Model 3 (original, TC + LDL-C)	194	1.0017	1	1.0033	0.053	TC =32.59, LDL-C =27.92
Model 3a (remove TC)	194	1.0016	1	1.0033	0.051	Checked, max VIF <5
Model 3b (remove LDL-C)	194	1.0016	1	1.0033	0.051	Checked, max VIF <5

This table summarizes a collinearity sensitivity analysis for the association between CS and CCTA underestimation in Model 3. Model 3: additionally adjusted for a comprehensive set of clinical confounders, including plaque type, HTN, DM, hyperlipidemia, smoking, alcohol consumption, BMI, ALT, AST, Cr, BUN, TG, TC, LDL-C, HDL-C, TNT, TNI, and D-dimer. The reference model included TC + LDL-C among lipid covariates; the sensitivity models (Model 3a and Model 3b) were constructed by removing TC or LDL-C, respectively, while keeping other covariates the same. “CS OR” denotes the OR for CCTA underestimation per 1 AU increase in CS (AU). “P value” refers to the two-sided P value for the CS term in each model. The “VIF issue” column reports the observed VIF related to collinearity (TC and LDL-C in the original model), and indicates that the maximum VIF was checked to be <5 after excluding TC or LDL-C in the sensitivity analyses. ALT, alanine aminotransferase; AST, aspartate aminotransferase; AU, Agatston units; BMI, body mass index; BUN, blood urea nitrogen; CCTA, coronary computed tomography angiography; CI, confidence interval; Cr, creatinine; CS, coronary artery calcification score; DM, diabetes mellitus; HDL-C, high-density lipoprotein cholesterol; HTN, hypertension; LDL-C, low-density lipoprotein cholesterol; OR, odds ratio; TC, total cholesterol; TG, triglycerides; TNI, troponin I; TNT, troponin T; VIF, variance inflation factor.

Table 3 summarizes the results of the logistic regression analysis evaluating the association between CS and the underestimation of CCTA findings. All three models consistently demonstrated a significant positive association between higher CS and the likelihood of CCTA underestimation. In Model 1 (unadjusted), the OR per 1 AU increase in CS was 1.002 (95% CI: 1.001–1.003; P=0.002). Model 2, adjusted for age and sex, yielded a slightly higher OR of 1.002 (95% CI: 1.001–1.004; P=0.001). Model 3, which incorporated comprehensive clinical confounders, also revealed a significant association with an OR of 1.002 (95% CI: 1.000–1.003; P=0.02). These findings collectively indicate that higher CS is independently associated with an increased likelihood of CCTA underestimation, even after accounting for multiple clinical factors. Table 4 details the collinearity sensitivity analysis for CS in Model 3, confirming that the removal of highly collinear lipid covariates (TC or LDL-C) did not materially alter the CS OR estimates and reduced the maximum VIF to <5, supporting the robustness of our findings.

RCS analysis and derived cutoffs (Figure 3, Table 5)

Figure 3.

RCS analysis of CS and the risk of CCTA underestimation. This figure presents the RCS dose-response relationship between CS (AU) and the risk of CCTA underestimation, with results shown for three logistic regression models: Model 1 (unadjusted), Model 2 (age + sex adjusted), and Model 3 (fully adjusted). The X-axis denotes CS (AU), while the Y-axis shows the OR for CCTA underestimation with a reference line at OR =1; the solid line represents the fitted RCS curve and the shaded region denotes the 95% CI. For each model, red dashed vertical lines indicate the lower and upper CS cutoffs (AU) derived from the RCS change points, and the reported P-nonlinearity quantifies evidence for non-linearity in the corresponding RCS model. Model 1: unadjusted, including only CS as the independent variable. Model 2: adjusted for age and sex. Model 3: additionally adjusted for a comprehensive set of clinical confounders, including plaque type, HTN, DM, hyperlipidemia, smoking, alcohol consumption, BMI, ALT, AST, Cr, BUN, TG, TC, LDL-C, HDL-C, TNT, TNI, and D-dimer. ALT, alanine aminotransferase; AST, aspartate aminotransferase; AU, Agatston units; BMI, body mass index; BUN, blood urea nitrogen; CCTA, coronary computed tomography angiography; CI, confidence interval; Cr, creatinine; CS, coronary artery calcification score; DM, diabetes mellitus; HDL-C, high-density lipoprotein cholesterol; HTN, hypertension; LDL-C, low-density lipoprotein cholesterol; OR, odds ratio; RCS, restricted cubic spline; TC, total cholesterol; TG, triglycerides; TNI, troponin I; TNT, troponin T.

Table 5. Diagnostic performance of CS for CCTA underestimation (different RCS thresholds).

Source	Model	Lower threshold (AU)	Upper threshold (AU)	TP, n	FP, n	FN, n	TN, n	Sensitivity (%)	Specificity (%)	PPV (%)	NPV (%)	Youden
Computed	Model 1	61.83	98.87	56	141	7	12	88.9	7.8	28.4	63.2	−0.033
	Model 2	89.42	414.89	41	86	22	67	65.1	43.8	32.3	75.3	0.089
	Model 3	75.53	99.73	57	144	6	9	90.5	5.9	28.4	60.0	−0.036
Paper	Model 1	118.07	270.87	54	114	9	39	85.7	25.5	32.1	81.2	0.112
	Model 2	118.07	253.50	55	117	8	36	87.3	23.5	32.0	81.8	0.108
	Model 3	118.07	329.90	51	105	12	48	81.0	31.4	32.7	80.0	0.123

Diagnostic thresholds are derived from RCS analyses and reported as lower and upper cutoffs. TP/FP/FN/TN correspond to classification outcomes for CCTA underestimation vs. ICA reference standard. Sensitivity = TP/(TP + FN); specificity = TN/(TN + FP); PPV = TP/(TP + FP); NPV = TN/(TN + FN). Youden index = sensitivity + specificity − 1. Both computed and paper threshold sets are presented in this reanalysis output; “source” indicates which threshold/operating point set was used to compute diagnostic performance (as provided in your reanalysis tables). “Computed” indicates thresholds and operating points derived from the current reanalysis dataset/model. “Paper” indicates thresholds adopted from the originally reported/paper-specified operating points, used here for comparison. Model 1: unadjusted, including only CS as the independent variable. Model 2: adjusted for age and sex. Model 3: additionally adjusted for a comprehensive set of clinical confounders, including plaque type, HTN, DM, hyperlipidemia, smoking, alcohol consumption, BMI, ALT, AST, Cr, BUN, TG, TC, LDL-C, HDL-C, TNT, TNI, and D-dimer. ALT, alanine aminotransferase; AST, aspartate aminotransferase; AU, Agatston units; BMI, body mass index; BUN, blood urea nitrogen; CCTA, coronary computed tomography angiography; Cr, creatinine; CS, coronary artery calcification score; DM, diabetes mellitus; FN, false negatives; FP, false positives; HDL-C, high-density lipoprotein cholesterol; HTN, hypertension; ICA, invasive coronary angiography; LDL-C, low-density lipoprotein cholesterol; NPV, negative predictive value; PPV, positive predictive value; RCS, restricted cubic spline ; TC, total cholesterol; TG, triglycerides; TN, true negatives; TNI, troponin I; TNT, troponin T; TP, true positives.

Figure 3 visually presents the RCS analysis depicting the dose-response relationship between CS and the risk of CCTA underestimation across the three logistic regression models. In all models, a significant U-shaped non-linear relationship was unequivocally observed (P-nonlinearity <0.05): both very low and very high CS values were associated with an increased risk of underestimation, while an intermediate CS range was associated with a comparatively lower risk of underestimation, often trending towards overestimation. Specifically, Model 1 (CS only) identified cutoff points at 118.07 AU (lower) and 270.87 AU (upper). Model 2, adjusted for age and sex, yielded cutoff points at 118.07 AU (lower) and 253.50 AU (upper). Model 3, fully adjusted for comprehensive clinical confounders, identified cutoff points at 118.07 AU (lower) and 329.90 AU (upper). These results underscore a robust and consistent non-linear association between CS and CCTA diagnostic performance, with a remarkably stable lower cutoff and a moderately model-dependent upper cutoff. A sensitivity analysis of the RCS model across different degrees of freedom (df =3, 4, and 5) is presented in Figure S3, confirming the consistency of the U-shaped pattern.

Table 5 details the diagnostic performance of these CS thresholds for identifying CCTA underestimation. In Model 1, sensitivity was 85.7%, specificity 26.8%, PPV 32.5%, and NPV 82.0%, with a Youden index of 0.125. Model 2 showed a sensitivity of 87.3%, specificity 24.8%, PPV 32.4%, and NPV 82.6%, with a Youden index of 0.121. Model 3 demonstrated a sensitivity of 82.5%, specificity 30.7%, PPV 32.9%, and NPV 81.0%, with a Youden index of 0.132. These findings indicate that while all three models exhibit moderate sensitivity, their specificity remains limited. Model 3 achieved the highest Youden index, suggesting a slightly better balance between sensitivity and specificity in the context of comprehensive adjustment.

Subgroup analyses (Table 6, Figure 4)

Table 6. Subgroup analyses (effect of CS on underestimation risk).

Subgroup variables	Number	Events	OR	95% CI lower	95% CI upper	P value	P-interaction
Sex							0.17
Female	73	19	1.003	1.001	1.006	0.007
Male	143	44	1.001	1.000	1.003	0.13
Elderly							0.67
Age ≤65 years	131	35	1.002	1.000	1.004	0.13
Age >65 years	85	28	1.002	1.000	1.004	0.02
DM							0.44
Non-DM	149	35	1.002	1.000	1.004	0.02
DM	67	28	1.001	0.999	1.003	0.23
HTN							0.78
Non-HTN	60	14	1.002	0.999	1.005	0.31
HTN	156	49	1.002	1.001	1.004	0.009
Smoking							0.19
Non-smoker	95	28	1.003	1.001	1.005	0.006
Smoker	121	35	1.001	0.999	1.003	0.21
Alcohol							0.60
Non-drinker	109	35	1.002	1.000	1.004	0.02
Drinker	107	28	1.002	1.000	1.004	0.10
Overweight							0.66
BMI <24 kg/m2	75	21	1.002	0.999	1.004	0.16
BMI ≥24 kg/m2	141	42	1.002	1.001	1.004	0.01
LDL							0.02
LDL ≥ median	114	34	1.004	1.002	1.006	0.001
LDL < median	102	29	1.000	0.998	1.002	0.89
HDL							0.22
HDL ≥ median	111	25	1.003	1.001	1.005	0.004
HDL < median	105	38	1.001	0.999	1.003	0.21
Plaque pattern							0.047
Non-calcified/mixed	165	53	1.002	1.000	1.003	0.03
Calcified	51	10	1.006	1.002	1.011	0.005
Plaque type							0.74
Calcified/non-calcified	94	22	1.002	1.000	1.005	0.07
Mixed	122	41	1.002	1.000	1.003	0.042

Subgroup analyses estimate the association between CS and CCTA underestimation within each stratum using logistic regression. For LDL-related subgroup analysis, the median cutoff was 3.4 mmol/L. For HDL-related subgroup analysis, the median cutoff was 1.16 mmol/L. “Events” indicates the number of underestimation cases within each subgroup level. “ORs” correspond to the effect of CS per 1 AU increase (as implemented in the subgroup model). “P value” is the two-sided P value for the subgroup-specific CS effect. “P-interaction” tests effect modification across subgroup levels. BMI, body mass index; CCTA, coronary computed tomography angiography; CI, confidence interval; CS, coronary artery calcification score; DM, diabetes mellitus; HDL, high-density lipoprotein; HTN, hypertension; LDL, low-density lipoprotein; OR, odds ratio.

Figure 4.

Subgroup analyses of the association between CS and CCTA underestimation risk. This forest plot summarizes subgroup-specific associations between CS (AU) and CCTA underestimation risk. The effect size is reported as the OR per 1 AU increase in CS, with corresponding 95% CIs for each subgroup level. A vertical dashed line at OR =1 indicates no association. The subgroup-level P value reflects the statistical significance of the CS effect within each stratum, while P interaction tests effect modification across subgroup categories. Subgroups include sex, age (elderly), DM, HTN, smoking, alcohol use, BMI (overweight), LDL-C and HDL-C (median-based cutoffs), and plaque type. AU, Agatston units; BMI, body mass index; CCTA, coronary computed tomography angiography; CI, confidence interval; CS, coronary artery calcification score; DM, diabetes mellitus; HDL, high-density lipoprotein; HDL-C, high-density lipoprotein cholesterol; HTN, hypertension; LDL, low-density lipoprotein; LDL-C, low-density lipoprotein cholesterol; OR, odds ratio.

As summarized in Table 6 and visually represented in Figure 4, subgroup analysis revealed a consistent and statistically significant correlation (P=0.002) between CS and CCTA underestimation across multiple subgroups. This correlation was particularly prominent in subgroups including females, elderly individuals (>65 years), non-diabetics, hypertensives, smokers, alcohol consumers, those who were overweight (BMI >24 kg/m²), individuals with low LDL-C (<3.4 mmol/L) and HDL-C (≤1.16 mmol/L) levels, and those with calcified plaques. Importantly, interaction effects were not statistically significant in most subgroups, indicating that these factors did not substantially modify the fundamental U-shaped relationship between CS and CCTA underestimation. These findings underscore the robustness of our primary observations across diverse patient populations.

Discussion

Our single-center retrospective study unveils a crucial non-linear dual-threshold (U-shaped) relationship between CS and the risk of CCTA underestimating moderate coronary stenosis (50–70% range). This novel finding provides quantifiable CS thresholds—a stable lower cutoff at 118.07 AU and model-dependent upper cutoffs (ranging from ~253 to 330 AU)—that delineate patient subgroups with elevated underestimation risk. This advanced understanding moves beyond qualitative observations of calcification artifacts towards a more precise, individualized interpretation of CCTA in a clinically challenging ‘gray zone’ (13).

The increased underestimation risk at low CS is mechanistically linked to the characteristics of early atherosclerotic plaques. These lesions are often non-calcified or minimally calcified and can be composed predominantly of lipid-rich components (14,15). CCTA may struggle to differentiate these ‘soft’ plaques from the vessel wall due to subtle attenuation differences. Furthermore, positive arterial remodeling, where the vessel expands to accommodate increasing plaque burden, can maintain a relatively normal luminal diameter despite significant plaque volume, thus masking the true severity of stenosis on visual CCTA assessment (16). As a result, CCTA can miss early or “soft” plaques, leading to an underrepresentation of the true extent of CAD in these patients (17).

Conversely, the elevated underestimation risk at high CS is primarily attributable to the well-documented limitations imposed by extensive calcification. Severe calcium deposits generate significant imaging artifacts, including blooming, beam-hardening, and partial-volume effects (18,19). These artifacts cause an artificial expansion of the high-density calcified lesions, obscuring adjacent lumen and impairing spatial and contrast resolution. Consequently, the true luminal narrowing is physically difficult to accurately assess, leading to a systematic underestimation of stenosis severity, even by experienced readers. This phenomenon is particularly pronounced when calcification occupies a substantial circumference of the vessel wall (20), and is increasingly relevant given the aging population and rising prevalence of CAD.

The intermediate CS range (between ~118 and ~253–330 AU), where the risk of underestimation is comparatively lower and there is a tendency towards overestimation, reflects a more nuanced interplay. In this range, moderate calcification may enhance plaque conspicuity, making lesions more readily visible than in low CS patients. However, the presence of visible but not overwhelming calcification can still introduce boundary uncertainty and partial-volume effects (21), leading to a degree of ambiguity in stenosis assessment. Radiologists, driven by the imperative to avoid missing significant disease, may adopt a more conservative interpretation, thereby tending to overestimate stenosis severity in this ‘alert zone’ (22).

Our findings significantly expand upon existing literature, which has predominantly focused on the independent impact of either low CS (e.g., non-calcified plaque burden) or high CS (due to blooming artifacts) on CCTA accuracy. By employing RCS analysis, we provide a continuous, non-linear characterization across the entire CS spectrum, revealing a distinct U-shaped pattern that was not previously well-quantified. The stability of the lower CS cutoff across different models is particularly noteworthy, suggesting a robust threshold for identifying patients with subtle plaque characteristics. While the upper cutoff showed some model dependency, this variability itself highlights the complex interplay of clinical confounders and imaging physics at very high calcium loads.

Clinical implications: these quantifiable CS thresholds offer a valuable tool for risk stratification and personalized management of patients undergoing CCTA. For patients with CCTA-reported 50–70% stenosis, if their CS falls below 118.07 AU or above the model-specific upper threshold, clinicians should have a heightened suspicion for potential underestimation. In these high-risk subgroups, further diagnostic evaluation, such as non-invasive functional assessment [e.g., fractional flow reserve derived from computed tomography (FFRct), stress myocardial perfusion imaging, or cardiac positron emission tomography] or early referral for ICA, should be strongly considered to prevent missed diagnoses and ensure appropriate revascularization decisions. Conversely, in patients with intermediate CS values, a CCTA finding of 50–70% stenosis may warrant a more cautious approach to ICA, potentially favoring further non-invasive functional assessment to rule out hemodynamically significant lesions and avoid unnecessary invasive procedures. This approach moves towards optimizing CCTA utility and improving patient outcomes by guiding appropriate resource allocation, thereby contributing to precision medicine in CAD management.

Limitations

This study has several strengths, including a well-defined cohort, comprehensive data collection, and robust statistical analyses. However, certain limitations should be acknowledged. First, its retrospective, single-center design may introduce selection bias and limit generalizability to other populations with different clinical characteristics or CCTA practices. Second, CCTA acquisition and reconstruction protocols may have evolved between 2012 and 2022, which could influence the extent of calcification artifacts and potentially affect CS threshold performance over time. Third, although ICA interpretation was performed by experienced clinicians with blinding to CCTA findings, and attempts were made to minimize inter-observer variability, visual stenosis grading by ICA itself remains susceptible to some degree of subjectivity, which was not fully quantified (e.g., via kappa statistics). Fourth, our study focused on a specific cohort of patients with CCTA-reported moderate stenosis (50–70%), and the findings may not be directly applicable to patients with lower or higher degrees of stenosis on CCTA. Finally, prospective, multi-center validation of the derived CS thresholds is imperative to confirm their external generalizability and clinical utility (23).

Conclusions

In conclusion, our study demonstrates a non-linear, dual-threshold (U-shaped) relationship between CS and CCTA underestimation of moderate coronary stenosis. We identified a consistently stable lower CS cutoff of 118.07 AU and model-specific upper cutoffs (ranging approximately from 253 to 330 AU), which stratify patients into higher or lower risk categories for CCTA underestimation. Patients with CS below 118.07 AU or above the model-specific upper cutoff are at higher risk for CCTA underestimation, while intermediate CS values are associated with comparatively lower risk, often tilting towards overestimation. While the per-unit OR is modest (OR ≈1.002 per AU), the cumulative effect across the clinically relevant CS range may be meaningful, particularly at the extremes of the CS distribution. These quantifiable CS thresholds provide a novel and practical framework for improving the diagnostic precision of CCTA, enabling more informed clinical decision-making, and facilitating personalized diagnostic and management strategies for patients with intermediate coronary lesions.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of Beijing Friendship Hospital (approval No. 2018-P2-030-01). Written informed consent for the use of clinical data was formally waived for this retrospective analysis by the Ethics Committee (as approved).

Data Sharing Statement

Available at https://cdt.amegroups.com/article/view/10.21037/cdt-2026-0118/dss

cdt-16-04-70-dss.pdf ^{(70.3KB, pdf)}

DOI: 10.21037/cdt-2026-0118