Vessel-specific perivascular fat attenuation index derived from AI-CCTA and its association with impaired coronary flow reserve in patients with INOCA

Abstract

Objectives

To investigate the association between artificial intelligence (AI)-derived coronary computed tomography angiography (CCTA) features and impaired coronary flow reserve (CFR) in patients with ischemia and non-obstructive coronary arteries (INOCA).

Methods

Retrospective analysis of 101 suspected coronary artery disease (CAD) patients with non-obstructive stenosis (<50%) on CCTA who underwent cadmium-zinc-telluride single photon emission computed tomography (CZT-SPECT). Stratified by coronary flow reserve (CFR) into CFR < 2.0 and CFR ≥ 2.0 groups at patient and vessel levels. Compared AI-CCTA parameters between groups; identified predictors via logistic regression; diagnostic performance was evaluated using receiver operating characteristic (ROC) analysis with Bootstrap internal validation.

Results

At the vessel-level, the CFR < 2.0 group had lower coronary artery calcium score (CACS) (62.12 vs. 142.40 AU, P = 0.021) and higher perivascular fat attenuation index (FAI) (−78.69 ± 8.34 vs. −82.03 ± 8.56 HU, P = 0.009). FAI was independently predictor of CFR < 2.0 (OR = 1.043, 95%CI: 1.005 ∼ 1.084, P = 0.028). A combined model integrating AI-CCTA and clinical features showed an apparent AUC of 0.807, but Bootstrap validation yielded a corrected AUC of 0.648. Inverse spatial distributions of CACS (RCA > LAD > LCX) and FAI (LCX > LAD > RCA). Vessels with CFR < 2.0 were characterized by lower calcification and higher FAI.

Conclusions

FAI is independently associated with vessel-level CFR impairment in INOCA. The combined model demonstrates potential but requires external validation. The observed inverse spatial and functional relationship between CACS and FAI may reflect different stages or patterns of coronary atherosclerosis in non-obstructive CAD, warranting further investigation.

1. Introduction

Angina pectoris is a common manifestation of ischemic heart disease, historically attributed to obstructive coronary artery disease (CAD). However, recent studies [1] indicated that up to 70% of patients undergoing coronary angiography (CAG) for angina and myocardial ischemia showed no significant obstructive CAD. This condition, termed ischemia with non-obstructive coronary artery disease (INOCA), is defined by objective evidence of myocardial ischemia (e.g., abnormal ECG/stress tests) despite non-obstructive coronary stenosis (<50%). The primary pathophysiological mechanisms involve coronary microvascular dysfunction (CMD) and/or coronary vasospasm. Failure to diagnose and manage INOCA promptly leads to recurrent angina, impaired quality of life, and increased cardiovascular risk [2], [3], [4].

Currently, due to the absence of direct coronary microcirculation visualization, functional assessment of the microvasculature relies on invasive or non-invasive measurement of parameters such as microcirculation resistance index (IMR) and coronary flow reserve (CFR) [5]. Previously, positron emission tomography (PET) assessed CFR non-invasively [6], but its adoption was limited by cost and accessibility. Advances in cardiac-dedicated imaging now integrate cadmium-zinc-telluride single photon emission computed tomography (CZT-SPECT), enabling quantitative CFR evaluation via dynamic myocardial perfusion imaging (MPI) with sensitivity and specificity comparable to invasive methods [7], [8]. Despite non-obstructive stenosis, INOCA patients often show epicardial atherosclerosis [5]. Coronary computed tomography angiography (CCTA) can visualize lumen and plaque for CAD screening [9], but its application is hindered by anatomical complexity, motion/calcification artifacts, and high computational demands [10]. Artificial intelligence (AI) and deep learning automate multidimensional parameter extraction from CCTA can overcome traditional model constraints and enhancing diagnostic accuracy for CAD [10], [11].

To our knowledge, it appears that the prognostic significance of AI-CCTA in diagnosing impaired CFR has not been thoroughly investigated. The pathophysiological relationship between AI-CCTA features and impaired CFR remains unclear. Therefore, this study aims to explore the impact of epicardial characteristics from AI-CCTA on impaired CFR and their intrinsic relationship, and to facilitate personalized therapeutic decision-making for INOCA patients.

2. Methods

2.1. Study design

Retrospectively analyzing clinical and imaging data from suspected CAD patients undergoing both CCTA and CZT-SPECT dynamic MPI at our hospital from January 2024 to April 2025. Inclusion criteria encompassed age ≥ 18 years, CCTA and CZT-SPECT performed <3 months apart without revascularization, CCTA-confirmed stenosis <50% in all vessels, and complete data; while exclusion criteria covered prior revascularization, heart failure, myocardial infarction, severe arrhythmia, cardiomyopathy, or poor image quality. The screening diagram is shown in Fig. 1. Finally, 101 suspected CAD patients (303 epicardial major vessels) were included, using CZT-SPECT-derived CFR as the diagnostic standard to stratify groups at patient/vessel levels (CFR < 2.0 group vs. CFR ≥ 2.0 group) [12]. All participants provided written informed consent before examination. The study complied with the Helsinki Declaration and was approved by the hospital ethics committee (Ethics Number: [2024]-1230-1).

Fig. 1.

The Screening Schematic Diagram of this Study CAD: coronary artery disease; CCTA: coronary computed tomography angiography; CZT-SPECT: cadmium-zinc-telluride single photon emission computed tomography; CFR: coronary flow reserve.

2.2. CCTA

CCTA was performed using a GE Revolution 256-slice CT scanner with prospective ECG gating in supine position, administering 60–70 mL Ioversol contrast (350 mgI/mL) intravenously at 4.5–5 mL/s. Scanning parameters: Tube voltage: 100/120 kV; Tube current: Smart mA (300–740 mA) [13]. Images were processed via uCT-FFR intelligent analysis system (v1.5; United-Imaging Healthcare) for AI-assisted plaque evaluation. Typical cases are demonstrated in Fig. 2. Record the corresponding parameters of the overall and three main coronary arteries, namely the left anterior descending artery (LAD), left circumflex artery (LCX), and right coronary artery (RCA). The statistical parameters include: (1) coronary artery calcium scores (CACS) [14]: Defined using the Agatston calcification score in Agatston units (AU), with severity levels 0 (normal), 1–99 (mild), 100–400 (moderate), and >400 (severe). (2) Perivascular fat attenuation index (FAI) [15], [16]: Specified as the average CT attenuation of adipose tissue surrounding the coronary arteries, ranging from −190 to −30 HU, featuring a critical threshold of −70.1 HU (analysis range: 40 mm proximal for LAD and LCX; for RCA, proximal 10 mm excluded, analyzed at 10–50 mm to avoid aortic wall interference). (3) Signs of high-risk plaques [17]: (I) Positive remodeling: Assessed via remodeling index ≥1.1. (II) Low Attenuation Plaque: Non-calcified plaque with mean CT attenuation <30 HU. (III) Napkin-ring sign: High-density halo (≤130HU) encircling a low-density core. (IV) Spotty calcification: Calcified plaque with maximal length <3 mm and peak density >130 HU. (4) High-risk plaques: ≥ 2 signs of high-risk plaques. (5) Burden and volume of plaque types [18]: Calcified plaques (>350 HU), lipid plaques (−30 ∼ 30 HU), fibrous fatty plaques (30–130 HU), and fibrotic plaques (130–350 HU). Non-calcified plaques include lipid, fibrous fatty, and fibrotic types. (6) Plaque length, minimum luminal area, maximum diameter stenosis, maximum area stenosis, remodeling index, eccentricity index.

Fig. 2.

Quantitative analysis of CCTA coronary characteristic parameters This analysis details the CCTA characteristics of the RCA in CFR < 2.0 patient, with CZT-SPECT revealing a notably reduced CFR of 1.73. (A) The degree of stenosis affecting the proximal lumen of the RCA is approximately 47%; (B) The orange color indicates the RCA calcification marker, utilizing the Agatston calcification score, revealing a CACS of 793.75 AU for the RCA; (C) Quantitative analysis of RCA proximal plaques utilizing plaque analysis software reveals a mixed plaque composition; (D) The semi-automatic quantification of FAI around RCA plaques, conducted via post-processing software, yielded a value of −82.72 HU. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

2.3. CZT-SPECT dynamic MPI

CZT-SPECT (NM530c, GE Healthcare, Haifa, Israel) imaging was performed using ^99mTc-MIBI or ^99mTc-TF. The patient should not drink coffee, tea, or consume any food containing caffeine or theophylline within 24 h before the examination, stop taking routine cardiovascular medications such as beta blockers, calcium blockers, vasodilators, etc. And refrain from smoking on the day of the examination [19]. Using a one-day imaging protocol, resting imaging was performed first followed by adenosine stress imaging [20]. Finally, CFR measurements for the LAD, LCX, and RCA perfusion territories and global left ventricular CFR were obtained.

2.4. Statistical analysis

Statistical analysis was performed using SPSS 27.0. Normality of continuous variables was first assessed using statistical tests. Normally distributed data were expressed as ($\overline{\text{x}}\pm \text{s}$) and analyzed with Student’s t-test or ANOVA, while non-normally distributed data were reported as median (Q1, Q3) and compared using the Mann-Whitney U test. Categorical variables were presented as n (%) with group differences evaluated by Chi-square test. To address the non-independence of multiple vessel-level measurements within the same patient, generalized estimating equations (GEE) were employed. For multiple group comparisons, Bonferroni correction was applied. Predictors of impaired CFR were identified through logistic regression analyses. A sensitivity analysis was also performed using a more stringent CFR threshold of 2.5. To control for potential clinical confounding, a multivariable logistic regression analysis was performed at the vessel level, which included the FAI and patient-level covariates. Model performance was evaluated using receiver operating characteristic (ROC) curves with area under the curve (AUC), internally validated via 1000 Bootstrap iterations, and compared with the DeLong test. All variables in model exhibited variance inflation factors (VIF) <10, confirming the absence of multicollinearity. A two-tailed P-value ≤ 0.05 defined statistical significance.

3. Results

3.1. Comparison of clinical data between two groups

Among 101 patients with non-obstructive coronary artery stenosis, 41 (40.6%) were in the CFR < 2.0 group and 60 (59.4%) were in the CFR ≥ 2.0 group. Age ranged from 27 to 78 (mean 56.28 ± 10.90) years old, and 41 males (40.6%) and 60 females (59.4%), with no statistically significant difference between the groups (P > 0.05) in Table 1.

Table 1.

Comparison of clinical data between two groups.

Clinical Data	CFR ≥ 2.0 group (n = 60)	CFR < 2.0 group (n = 41)	P value
Age, years	55.57 ± 10.74	57.32 ± 11.17	0.431
Man, n (%)	21(35.0%)	20(48.8%)	0.166
BMI, kg/m2	25.62 ± 3.45	26.03 ± 3.74	0.571
Hypertension, n (%)	32(53.3%)	15(36.6%)	0.098
Diabetes, n (%)	11(18.3%)	4(9.8%)	0.234
Hyperlipidemia, n (%)	17(28.3%)	6(14.6%)	0.107
Smoking history, n (%)	10(16.7%)	12(29.3%)	0.132
Family history, n (%)	9(15.0%)	6(14.6%)	0.960

BMI: body mass index.

3.2. Comparison of characteristic parameters between two groups of CCTA at the patient-level

The CACS, lipid plaque burden and volume, positive remodeling, and high-risk plaques in the CFR ≥ 2.0 were higher, while other parameters including FAI were lower than CFR < 2.0 group, but the differences were not statistically significant (P > 0.05) in Table 2.

Table 2.

Comparison of coronary characteristic parameters between two groups of CCTA at the patient-level.

Characteristic parameters	CFR ≥ 2.0 group (n = 60)	CFR < 2.0 group (n = 41)	P value
CACS (AU)	509.40(315.90, 731.78)	447.09(230.98, 716.08)	0.714
FAI (HU)	−80.97(−85.50, −77.43)	−80.24(−83.68, −75.05)	0.159
Total Plaque Burden (%)	6.29(1.78, 13.45)	7.89(2.10, 17.18)	0.366
Calcified Plaque Burden (%)	0.25(0.01, 2.45)	0.86(0.07, 1.89)	0.323
Non-Calcified Burden (%)	5.67(1.49, 11.35)	7.41(1.82, 15.89)	0.408
Lipid Plaque Burden (%)	0.23(0.00, 0.81)	0.14(0.02, 0.68)	0.978
Fibrous Fatty Burden (%)	1.81(0.34, 3.93)	1.87(0.38, 6.12)	0.467
Fibrotic Plaque Burden (%)	3.27(0.78, 6.19)	3.80(1.20, 8.70)	0.300
Total Plaque Volume (mm3)	66.60(24.21, 124.02)	94.22(23.68, 174.29)	0.440
Calcified Plaque Volume (mm3)	2.78(0.16, 21.09)	7.11(1.00, 17.63)	0.271
Non-Calcified Volume (mm3)	52.33(21.00, 111.55)	72.00(17.42, 140.83)	0.436
Lipid Plaque Volume (mm3)	2.05(0.07, 8.34)	1.85(0.19, 7.32)	0.936
Fibrous Fatty Volume (mm3)	17.18(3.82, 42.04)	22.79(3.91, 58.59)	0.488
Fibrotic Plaque Volume (mm3)	30.78(10.64, 63.01)	41.95(12.28, 96.35)	0.400
Positive Remodeling	44(73.3%)	28(68.3%)	0.582
Low Attenuation Plaque	44(73.3%)	31(75.6%)	0.797
Napkin-ring Sign	10(16.7%)	12(29.3%)	0.132
Spotty Calcification	22(36.7%)	22(53.7%)	0.091
High-risk plaques	44(73.3%)	30(73.2%)	0.986
Minimum Luminal Area (mm2)	3.20(0.20, 5.13)	2.97(1.67, 4.86)	0.950
Maximum Diameter Stenosis (%)	40.09(26.07, 53.30)	45.46(27.21, 57.02)	0.685
Maximum Area Stenosis (%)	58.01(42.23, 71.67)	60.00(42.21, 75.05)	0.711
Remodeling Index (%)	1.13(1.03, 1.21)	1.12(1.08, 1.19)	0.893
Eccentricity Index (%)	0.83(0.00, 0.98)	0.88(0.66, 1.00)	0.137

CACS: coronary artery calcium scores; FAI: Perivascular fat attenuation index.

3.3. Comparison of characteristic parameters between two groups of CCTA at the vessel-level

303 major epicardial vessels underwent AI-CCTA analysis, including 85 vessels in the CFR < 2.0 group and 218 vessels in the CFR ≥ 2.0 group. The GEE model results showed that the CACS [142.40 (2.63, 300.30) AU] in the CFR ≥ 2.0 group was higher than the CFR < 2.0 group [64.12 (0.45, 188.23) AU], P = 0.021. The FAI (−82.03 ± 8.56 HU) was lower than the CFR < 2.0 group (−78.69 ± 8.34 HU), P = 0.009. In addition, other plaque parameters did not show significant differences between the two groups (P > 0.05), as shown in Table 3.

Table 3.

Comparison of coronary characteristic parameters between two groups of CCTA at the vessel-level.

Characteristic parameters	CFR ≥ 2.0 group (n = 218)	CFR < 2.0 group (n = 85)	P value
CACS (AU)	142.40(2.63, 300.30)	64.12(0.45, 188.23)	0.021
FAI (HU)	−82.03 ± 8.56	−78.69 ± 8.34	0.009
Total Plaque Burden (%)	0.00(0.00, 5.65)	0.00(0.00, 5.25)	0.639
Calcified Plaque Burden (%)	0.00(0.00, 0.28)	0.00(0.00, 0.46)	0.598
Non-Calcified Burden (%)	0.00(0.00, 4.91)	0.00(0.00, 4.57)	0.680
Lipid Plaque Burden (%)	0.00(0.00, 0.13)	0.00(0.00, 0.09)	0.482
Fibrous Fatty Burden (%)	0.00(0.00, 1.60)	0.00(0.00, 1.10)	0.731
Fibrotic Plaque Burden (%)	0.00(0.00, 2.77)	0.00(0.00, 2.69)	0.728
Total Plaque Volume (mm3)	0.00(0.00, 52.87)	0.00(0.00, 37.97)	0.464
Calcified Plaque Volume (mm3)	0.00(0.00, 3.15)	0.00(0.00, 4.96)	0.371
Non-Calcified Volume (mm3)	0.00(0.00, 47.12)	0.00(0.00, 33.94)	0.533
Lipid Plaque Volume (mm3)	0.00(0.00, 1.52)	0.00(0.00, 0.93)	0.658
Fibrous Fatty Volume (mm3)	0.00(0.00, 15.74)	0.00(0.00, 10.25)	0.532
Fibrotic Plaque Volume (mm3)	0.00(0.00, 28.43)	0.00(0.00, 20.89)	0.568
Positive Remodeling	74(33.9%)	26(30.6%)	0.556
Low Attenuation Plaque	80(36.7%)	30(35.3%)	0.810
Napkin-ring Sign	19(8.7%)	5(5.9%)	0.386
Spotty Calcification	37(17.0%)	18(21.2%)	0.344
High-risk plaques	75(34.4%)	27(31.8%)	0.641
Length (mm)	0.00(0.00, 14.70)	0.00(0.00, 16.05)	0.520
Minimum Luminal Area (mm2)	0.00(0.00, 4.25)	0.00(0.00, 4.73)	0.827
Maximum Diameter Stenosis (%)	0.00(0.00, 39.80)	0.00(0.00, 37.13)	0.771
Maximum Area Stenosis (%)	0.00(0.00, 56.46)	0.00(0.00, 54.75)	0.824

CACS: coronary artery calcium scores; FAI: Perivascular fat attenuation index.

3.4. Comparison of CACS and FAI of three main coronary arteries

The GEE model showed significant whole differences in CACS among the three main coronary arteries, with RCA having the highest CACS [247.22 (78.85, 477.65) AU], followed by LAD at [216.92 (125.39, 280.71) AU], and LCX having the lowest CACS [0.08 (0.00, 3.45) AU], adjusted P < 0.05. Significant differences were observed between LAD and LCX, as well as between LCX and RCA, in both the CFR < 2.0 and CFR ≥ 2.0 subgroups (adjusted P < 0.05), but there were no significant difference between LAD and RCA (adjusted P > 0.05); See Fig. 3.

Fig. 3.

Comparison of CACS among the whole, CFR < 2.0 group, and CFR ≥ 2.0 group in the three main Coronary Artery CACS: coronary artery calcium scores; LAD: left anterior descending artery; LCX: left circumflex artery; RCA: right coronary artery.

In addition, there were significant differences in FAI among the three vessels. Post hoc comparisons showed that LCX had the highest FAI, at −77.70 ± 8.05 HU, higher than LAD (−81.38 ± 8.06 HU) and RCA (−84.20 ± 8.55 HU), adjusted P < 0.001 for both. Meantime, LAD had a significantly higher FAI than RCA (adjusted P = 0.001). This FAI distribution pattern (LCX > LAD > RCA) also remained consistent in the subgroup analysis of the CFR ≥ 2.0 (adjusted P < 0.05). However, in the CFR < 2.0 subgroup, pairwise comparisons after treatment only showed that the FAI of LCX was higher than that of RCA (adjusted P = 0.005), while the differences between LAD and LCX, and LAD and RCA were no longer statistically significant (adjusted P > 0.05); See Fig. 4.

Fig. 4.

Comparison of FAI among the whole, CFR < 2.0 group, and CFR ≥ 2.0 group in the three main coronary artery FAI: Perivascular fat attenuation index; LAD: left anterior descending artery; LCX: left circumflex artery; RCA: right coronary artery.

The comparison of CACS and FAI between the two groups was conducted in LAD, LCX and RCA. The results showed that only the CFR ≥ 2.0 group in LAD had a higher CACS [225.54 (155.27, 289.06) AU] than the CFR < 2.0 group [164.78 (94.58, 225.67) AU], P = 0.019; However, among other groups, there were results where the CFR ≥ 2.0 group had a higher CACS and a lower FAI than the CFR < 2.0 group, but these differences were not statistically significant (P > 0.05), as shown in Table 4.

Table 4.

Comparison of CACS and FAI between CFR < 2.0 group and CFR ≥ 2.0 group in three branches of coronary artery.

		CACS (AU)	FAI (HU)
LAD	CFR ≥ 2.0 group	225.54(155.27, 289.06)	−82.00 ± 8.24
	CFR < 2.0 group	164.78 (94.58, 225.67)	−79.59 ± 7.40
	P value	0.019	0.191
LCX	CFR ≥ 2.0 group	0.08(0.00, 4.68)	−77.77 (−84.05, −74.22)
	CFR < 2.0 group	0.06 (0.00, 2.21)	−75.90(−80.86, −72.57)
	P value	0.890	0.260
RCA	CFR ≥ 2.0 group	266.67(66.49, 474.32)	−83.38 (−90.69, −79.18)
	CFR < 2.0 group	184.11 (81.16, 484.33)	−82.72(−88.07, −75.27)
	P value	0.894	0.232

CACS: coronary artery calcium scores; FAI: Perivascular fat attenuation index; LAD: left anterior descending artery; LCX: left circumflex artery; RCA: right coronary artery.

3.5. Logistic regression analysis of CCTA coronary characteristics at the vessel-level

Using the GEE model, the results showed that FAI (OR: 1.043, 95%CI: 1.005–1.084, P = 0.028) can be used as an independent predictor of CFR < 2.0, as shown in Table 5.

Table 5.

Logistic regression analysis of CCTA plaque characteristics at the vessel-level.

Predictors	Univariate analysis		Multivariate analysis
	OR (95%CI)	P value	OR (95%CI)	P value
CACS (10AU)	0.999(0.997 ∼ 1.000)	0.046	0.999(0.997 ∼ 1.001)	0.346
FAI (HU)	1.049(1.013 ∼ 1.086)	0.008	1.043(1.005 ∼ 1.084)	0.028
Total Plaque Burden (%)	0.991(0.953 ∼ 1.030)	0.634
Calcified Plaque Burden (%)	0.958(0.823 ∼ 1.116)	0.581
Non-Calcified Burden (%)	0.990(0.944 ∼ 1.038)	0.680
Lipid Plaque Burden (%)	0.845(0.503 ∼ 1.421)	0.526
Fibrous Fatty Burden (%)	0.979(0.866 ∼ 1.107)	0.734
Fibrotic Plaque Burden (%)	0.986(0.910 ∼ 1.067)	0.725
Total Plaque Volume (mm3)	0.999(0.996 ∼ 1.002)	0.872
Calcified Plaque Volume (mm3)	0.994(0.983 ∼ 1.005)	0.306
Non-Calcified Volume (mm3)	0.999(0.995 ∼ 1.003)	0.999
Lipid Plaque Volume (mm3)	0.990(0.943 ∼ 1.039)	0.678
Fibrous Fatty Volume (mm3)	0.997(0.986 ∼ 1.007)	0.545
Fibrotic Plaque Volume (100 mm3)	0.998(0.992 ∼ 1.004)	0.560
Positive Remodeling	0.858(0.510 ∼ 1.442)	0.562
Low Attenuation Plaque	0.941(0.571 ∼ 1.551)	0.811
Napkin-ring Sign	0.655(0.235 ∼ 1.823)	0.417
Spotty Calcification	1.314(0.751 ∼ 2.299)	0.338
High-risk plaques	0.946(0.486 ∼ 1.840)	0.870
Length (mm)	0.996(0.981 ∼ 1.011)	0.609
Minimum Luminal Area (mm2)	1.009(0.932 ∼ 1.092)	0.826
Maximum Diameter Stenosis (%)	1.002(0.991 ∼ 1.012)	0.769
Maximum Area Stenosis (%)	1.001(0.993 ∼ 1.009)	0.823
Vascular type
LAD VS RCA	1.176(0.761 ∼ 1.817)	0.466
LCX VS RCA	1.878(1.287 ∼ 0.742)	0.001	1.106(0.577 ∼ 2.122)	0.761

CACS: coronary artery calcium scores; FAI: Perivascular fat attenuation index; LAD: left anterior descending artery; LCX: left circumflex artery; RCA: right coronary artery.

3.6. Sensitivity analysis using a more stringent CFR threshold of 2.5 at the vessel-level

Vessels with CFR < 2.5 exhibited significantly higher FAI compared to those with CFR ≥ 2.5 [-79.06(−84.26, −73.99) vs. −82.38(−87.30, −75.76) HU, P = 0.003], while the difference in CACS was no longer significant (P = 0.071) (Supplementary Table S1). In a multivariable logistic regression model, FAI remained independently associated with CFR < 2.5 (OR = 1.043, 95% CI: 1.007–1.080, P = 0.019) (Supplementary Table S2).

3.7. Multivariable regression analysis adjusted for clinical confounders at the vessel-level

The results showed that FAI remained a significant independent predictor of CFR < 2.0 (adjusted OR: 1.046, 95% CI: 1.011–1.082, P = 0.010). Concurrently, statin use was identified as a strong independent protective factor (adjusted OR: 0.306, 95% CI: 0.120 ∼ 0.781, P = 0.013) (Supplementary Table S3).

3.8. Analysis of the diagnostic efficacy of low CFR based on different models

The results showed that both clinical features and CCTA coronary parameters with a single parameter had unsatisfactory diagnostic efficacy for impaired CFR, with AUC < 0.65. In view of the above situation, we further constructed a joint diagnostic model with multidimensional parameters. The clinical features model (VIF < 5), showed a certain diagnostic ability for low CFR, with AUC value of 0.722. The CCTA joint feature model selected representative parameters, including CACS, FAI, high-risk plaque signs, calcified plaque volume, non-calcified plaque volume, minimum luminal area, maximum diameter stenosis, remodeling index, and eccentricity index (VIF < 10), with AUC value of 0.745. The combination of CCTA and clinical features model (VIF < 10) predicted the highest AUC for CFR < 2.0, with a value of 0.807, 95%CI of 0.720 ∼ 0.894, sensitivity of 75.6%, and specificity of 80.0%, as shown in Fig. 5. After 1000 Bootstrap internal validations, the adjusted AUC values for the three were 0.630, 0.631, and 0.648, respectively. Pairwise comparison of the original AUCs using DeLong test revealed no statistically significant differences between any two models (all P > 0.05) (Supplementary Table S4).

Fig. 5.

ROCs of low CFR with different predictive models CCTA: coronary computed tomography angiography.

4. Discussion

In the present study, we adopted a two-level (patient and vessel) analytical approach to investigate INOCA. Patients and individual coronary vessels were stratified based on a clinically relevant threshold of CFR < 2.0, operationally defining the CFR < 2.0 group. We propose that this group represents a distinct “low-CFR INOCA phenotype”, which focuses on a key functional derangement common across multiple potential microvascular etiologies [21], providing a pragmatic imaging-based framework for risk stratification prior to invasive assessment. Within this framework, our study demonstrates a vessel-level association between an elevated FAI and low-CFR INOCA phenotype. From a pathophysiological perspective, this association raises the hypothesis that epicardial perivascular inflammation may not remain confined to the large epicardial vessels, but could extend inward to affect the coronary microcirculation. Researches have shown [22], [23] that perivascular adipose tissue-derived adipokines are pro-inflammatory molecules that induce endothelial oxidative stress, could impair endothelial function and nitric oxide bioavailability, leading to coronary vasoconstriction and vascular smooth muscle cell proliferation. These mechanisms, initially characterized in epicardial arteries, are biologically plausible contributors to microvascular dysfunction and impaired myocardial perfusion. This proposed mechanistic link—from epicardial perivascular inflammation to downstream microvascular dysfunction—remains a conceptual model derived from our observational data. Thus, our findings should be interpreted as hypothesis-generating, providing a rationale for future translational and mechanistic investigations.

Notably, this association proved robust in a sensitivity analysis employing a stricter definition for preserved microvascular function (CFR ≥ 2.5), where FAI remained a significant independent predictor, reinforcing its role as a stable imaging biomarker for the low-CFR phenotype. Crucially, the association between elevated FAI and impaired CFR persisted after rigorous adjustment for clinical confounders, including age, gender, BMI, diabetes, statin use, and smoking, confirming its independence. Furthermore, the analysis revealed that statin therapy was independently associated with preserved microvascular function. This novel finding aligns with the anti-inflammatory and pleiotropic effects of statins [24] and underscores the importance of accounting for medication history when evaluating imaging biomarkers of vascular inflammation.

In addition, this study discovered a distinct “calcification-inflammation dissociation” pattern in non-obstructive CAD patients when comparing CACS and FAI across the three main coronary arteries. Specifically, CACS showed heterogeneity with RCA > LAD > LCX, whereas FAI exhibited the opposite pattern of LCX > LAD > RCA, reflecting differences in atherosclerotic plaque progression between vessels. This dissociation likely stems from anatomical factors like vessel bifurcation and tortuosity, as well as hemodynamic changes such as endothelial shear stress [25]. Although contrasting with previous reports indicating “LAD has the highest calcification burden” in obstructive CAD patients with stable angina or acute coronary syndrome, where LAD demonstrates prominent hemodynamic burden and plaque vulnerability [25], [26], [27], this highlights INOCA as an independent pathological entity, not merely a milder CAD form, characterized by diffuse coronary microvascular dysfunction and/or spasm [2]. Here, RCA's role as a high-risk site for microvascular spasm may promote calcification through chronic ischemia–reperfusion injury and shear stress changes, placing it in a late stable phase [28], [29], while LAD is more likely to be in a relatively more active pathological process. Conversely, for LCX with a CACS of 0 or lower, the measured FAI was higher. This indicates that LCX's tortuous anatomy predisposes it to endothelial dysfunction and sustained inflammation owing to unique mechanical forces [30]. Therefore, as revealed by Mátyás et al. [31], biologically active high risk atherosclerotic plaque that may not be detected only by CACS, we need to integrate anatomy and inflammatory imaging to obtain comprehensive evaluation value.

At the same time, subgroup analysis of the three coronary arteries in CFR < 2.0 and CFR ≥ 2.0 groups reveals key insights into vascular patterns. No difference in CACS between LAD and RCA existed in either group, indicating high calcification represents an inherent anatomical property of these vessels rather than a direct microcirculation determinant [32]. For FAI, it showed a clear gradient across vessels in the CFR ≥ 2.0 subgroup, confirming inherent inflammatory differences; in CFR < 2.0 subgroup, this difference was largely weakened (with only statistical difference between LCX and RCA), indicating that low-CFR INOCA phenotype may associate with the entire coronary system entering a more homogeneous inflammatory state. Under ischemic stress, inflammation in LAD and RCA increases sharply, potentially reaching the elevated baseline level of LCX and masking inherent inter-vessel differences. Thus, plaque stabilization/calcification drivers for vessels of non-obstructive CAD patients may differ from or oppose perivascular inflammation drivers. This suggests that the LCX region may show active lesions characterized mainly by inflammation earlier, while the RCA region may more likely exhibit late, stable lesions dominated by calcification. But this hypothesis requires direct evidence from longitudinal or histopathologic studies for verification. In addition, our speculation is not entirely consistent with the results of prognostic studies such as the CRISP-CT study [33], which found that FAI around the LCX does not have predictive value for mortality. This suggests that the clinical significance of local FAI may vary depending on the study population, endpoint events, and disease stage. Our findings focus on the functional abnormality phenotype, and its relationship with long-term prognosis remains to be further explored.

It is worth noting that our study identified a consistent “calcification-inflammation dissociation” phenomenon in low-CFR INOCA phenotype across patient, vessel, and coronary branch levels, characterized by lower CACS but higher FAI compared. This cross-level consistency reveals that inflammatory activity surrounding microvasculature—not stable calcified plaques—directly drives microvascular dysfunction. The spatial dissociation is more like under a “static” anatomical distribution. In different vascular environments, the development stage and dominant characteristics of atherosclerosis are different. Some vessels tend to calcification (such as RCA), while others are more inclined to maintain an inflammatory state (such as LCX); Functional dissociation, on the other hand, is similar to dissociation under a “dynamic” pathophysiological influence. The root of myocardial ischemia is more related to the abnormal function of microvasculature [3] and active inflammation of surrounding tissues [23], independent of macroscopic calcified plaque burden. Although statistical significance varied locally, the phenomenon’s reproducibility across hierarchical levels strongly indicates it is non-random and mechanistically relevant to this phenotype pathogenesis.

Finally, this study evaluated the diagnostic efficacy of clinical features and CCTA coronary parameters for low-CFR INOCA phenotype using the ROC analysis. We found that the diagnostic value of a single clinical factor or coronary parameter was limited (AUC < 0.65), while after strict collinearity control, integrating multidimensional and multiparameter joint models can significantly improve the diagnostic value of this phenotype. However, after accounting for overfitting via Bootstrap internal validation, the corrected discriminative ability of all models was more modest, and the combined model did not demonstrate a statistically significant superiority over the simpler clinical or CCTA models alone in our cohort. This indicates that the initially high AUC of the combined model was likely inflated and should be interpreted as preliminary. Therefore, the primary value of the combined model lies not in its current validated performance, but in its proof-of-concept that integrating clinical and advanced CCTA features holds potential for identifying the low-CFR phenotype. This multi-parametric approach warrants further optimization and external validation in larger, prospective cohorts to determine if a statistically and clinically significant advantage can be achieved.

This study still has certain limitations: (1) It is a single center retrospective study with a relatively small number of included cases, and there may be some bias in the enrolled patients. The extrapolation of the conclusions needs further verification through large scale and multi center prospective studies; (2) CAG was not used to evaluate the degree of coronary artery stenosis. However, CCTA assisted by AI not only has high negative predictive value, but also can obtain various coronary characteristic parameters [34], supporting its use as the reference standard in this study; (3) Due to limitations in clinical application, invasive measurement of IMR [35] or specific tests such as acetylcholine stimulation test were not performed on the patients to exclude epicardial or microvascular spasm [36]. In this study, the CFR ≥ 2.0 group may have mixed patients with completely normal microcirculation function and vascular spasm, which may dilute the differences that should exist between the two groups to some extent, leading to underestimation of the correlation between some parameters; (4) This study did not involve follow-up data, and survival data will be introduced in the future to verify the potential association between the above risk factors and cardiovascular adverse events in INOCA patients.

In summary, this study demonstrates that AI-derived CCTA features—FAI—are independently associated with impaired CFR at the vessel level in INOCA patients. The observed inverse spatial distribution of CACS and FAI suggests heterogeneous disease phenotypes across different coronary territories. These findings support the hypothesis that epicardial perivascular inflammation may extend to affect the coronary microcirculation, though this mechanistic link remains conceptual and requires direct investigation. A multi-parameter model integrating AI-CCTA and clinical features showed diagnostic potential for the low-CFR phenotype, but its performance was modest after internal validation and was not statistically superior to simpler models; external validation is therefore essential. Collectively, this study provides a hypothesis-generating foundation for the non-invasive phenotyping of INOCA patients using AI-CCTA, which may inform future strategies for personalized risk stratification and therapeutic targeting.

CRediT authorship contribution statement

Jing Ni: Writing – original draft, Software, Resources, Project administration, Methodology, Investigation. Ting Wang: Writing – review & editing, Validation, Resources, Methodology, Investigation, Formal analysis, Data curation. Haoran Guo: Writing – review & editing, Validation, Resources, Investigation. Ajay Kumar Chaudhary: Writing – review & editing, Validation, Resources, Investigation, Data curation. Zekun Pang: Writing – review & editing, Validation, Resources, Project administration, Investigation. Fukai Zhao: Writing – review & editing, Validation, Resources, Investigation, Formal analysis. Yue Chen: Writing – review & editing, Validation, Supervision, Resources, Investigation. Jiao Wang: Writing – review & editing, Validation, Resources, Investigation, Funding acquisition. Jianming Li: Writing – review & editing, Supervision, Resources, Project administration, Methodology, Funding acquisition, Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following are the Supplementary data to this article: