Artificial intelligence-based chest computed tomography assessment of bone, vascular, and epicardial adipose changes in type 2 diabetes

Abstract

BACKGROUND

Type 2 diabetes mellitus (T2DM) is a complex multisystem disorder in which chronic inflammation often drives both bone loss and accelerated atherosclerosis. Although coronary artery calcification (CAC) and epicardial adipose tissue (EAT) are established cardiovascular risk markers, their integrated relationship with bone mineral density (BMD) in T2DM remains underexplored.

AIM

To develop an artificial intelligence (AI)-based framework to simultaneously quantify BMD, CAC, and EAT from routine chest computed tomography and investigate their associations in T2DM.

METHODS

This retrospective study included 4589 participants, including 1369 patients with T2DM and 3220 controls, from five tertiary hospitals. An AI framework was used to quantify thoracic vertebral BMD, the Agatston coronary artery calcium score, EAT volume, and EAT density. Propensity score matching was performed to control for confounding bias, and multivariable regression models were used to assess independent associations. Restricted cubic splines were used to evaluate nonlinear relationships between imaging biomarkers. Subgroup analyses were conducted to examine the consistency of associations across disease duration and glycemic control groups.

RESULTS

After adjustment for age, sex, and metabolic risk factors, T2DM remained independently associated with higher EAT density (β = 4.131, P < 0.001) and severe CAC (odds ratio = 3.656, P < 0.001). Restricted cubic splines analysis revealed a significant negative linear association between BMD and EAT volume (P < 0.001) and a positive linear association between EAT density and severe CAC (P = 0.007). Notably, a J-shaped nonlinear relationship was observed between BMD and EAT density (P for nonlinearity < 0.001). Subgroup analysis showed that the negative correlation between BMD and EAT volume was significantly stronger in patients with T2DM than in controls, particularly among those with moderate glycemic control and shorter disease duration.

CONCLUSION

T2DM was associated with adverse cardiovascular-bone profiles. Automated AI-based opportunistic screening using chest computed tomography provides a “one-stop” tool for early identification of high-risk multisystem phenotypes in T2DM, thereby facilitating more precise individualized interventions.

Key Words: Type 2 diabetes mellitus; Chest computed tomography; Artificial intelligence; Bone mineral density; Coronary artery calcification; Epicardial adipose tissue

Core Tip: We developed an artificial intelligence-based opportunistic chest computed tomography framework that simultaneously quantifies bone mineral density, coronary artery calcification, and epicardial adipose tissue in type 2 diabetes mellitus. The study reveals a linear bone mineral density-epicardial adipose tissue (EAT) volume association, a linear EAT density-severe coronary artery calcification association, and a J-shaped bone mineral density-EAT density relationship, supporting routine chest computed tomography as a one-stop tool for multisystem risk stratification in diabetes.

INTRODUCTION

Type 2 diabetes mellitus (T2DM) affects more than 500 million adults worldwide, representing a substantial global health burden[1]. Beyond glucose dysregulation, T2DM is recognized as a systemic inflammatory disorder that contributes to both bone fragility and accelerated atherosclerosis[2,3]. Hyperglycemia and oxidative stress impair bone quality while promoting endothelial dysfunction and vascular calcification[4,5]. This “bone-vascular axis” suggests that osteoporosis and atherosclerosis may share common pathogenic mechanisms and pathophysiological processes[6]. Integrated assessment of these multiorgan systems is therefore essential for optimizing management in this high-risk population.

Bone mineral density (BMD) is commonly assessed using dual-energy X-ray absorptiometry (DXA) to predict fracture risk[7]. Cardiovascular risk is frequently quantified using coronary artery calcium score (CACS) on dedicated cardiac computed tomography (CT)[8]. However, DXA screening remains underutilized in patients with T2DM, and many high-risk individuals do not receive timely evaluation[9]. Furthermore, separate bone and cardiac imaging increases healthcare costs and clinical fragmentation, and integration into a unified automated diagnostic framework remains limited in daily practice.

Noncontrast chest CT is frequently performed in T2DM patients for various indications, providing an underutilized source of health data. Although deep learning enables automated measurement of thoracic BMD, CACS, and epicardial adipose tissue (EAT) from these scans, the interactions among these biomarkers in T2DM remain incompletely understood[10,11]. Understanding the relationships among bone loss, ectopic fat accumulation, and arterial calcification is essential for risk stratification beyond traditional clinical factors. This study aimed to develop an artificial intelligence (AI)-based framework for the simultaneous quantification of thoracic BMD, CACS, and EAT from chest CT. By applying this “one-stop” tool to a large multicenter cohort, we sought to investigate the associations among these biomarkers in T2DM patients compared with nondiabetic controls.

MATERIALS AND METHODS

Study population

This multicenter retrospective study included patients who underwent non-contrast chest CT between January 2019 and January 2024 at five tertiary hospitals. The institutional review boards approved the study protocol of all participating centers, and the requirement for informed consent was waived because of the retrospective nature of the study (ethics approval No: TJ-IRB202412092).

The T2DM cohort comprised patients with diagnosed T2DM who were hospitalized for endocrine-related conditions. The nondiabetic (non-T2DM) control group comprised individuals who underwent chest CT for health screening or noncardiac indications during the same period. The absence of diabetes in controls was confirmed through electronic medical record review.

The exclusion criteria for both groups were as follows: (1) Established cardiovascular disease, including myocardial infarction or ischemia-related revascularization; (2) Severe motion or respiratory artifacts compromising quantitative analysis; (3) Incomplete clinical or imaging data; (4) Diseases affecting bone metabolism, including hyper/hypoparathyroidism and Paget disease; (5) A history of malignancy; and (6) A history of prediabetes or glucose-lowering medication use in controls. After screening, 1369 patients with T2DM and 3220 non-T2DM controls were included in the final analysis (Figure 1).

Figure 1 Flowchart showing the study design. T2DM: Type 2 diabetes mellitus; CT: Computed tomography.

Data collection and definitions

Baseline demographics, cardiovascular risk factors, and laboratory data were extracted from clinical records. Variables included age, sex, body mass index (BMI), smoking status, alcohol consumption, hypertension, and dyslipidemia.

T2DM was diagnosed according to the 2024 American Diabetes Association criteria[12]: (1) Documented diabetes history; (2) Fasting plasma glucose ≥ 7.0 mmol/L; (3) Glycated hemoglobin ≥ 6.5%; and (4) Current glucose-lowering medication use. T2DM patients were categorized by disease duration as short (< 5 years), medium (5-10 years), or long (> 10 years)[13,14]. Glycemic control (GC) was classified as good (glycated hemoglobin < 7%), moderate (7%-9%), or poor (> 9%)[15].

Coronary artery calcification (CAC) severity was graded as none (0), mild (1-100), moderate (101-300), or severe (> 300)[16]. Severe CAC (SCAC) was defined as CACS > 300[17].

Chest CT acquisition

Chest CT scans were performed using multidetector CT systems (Siemens Healthineers, GE Healthcare, and Philips Healthcare). Each examination was conducted during a single breath-hold at the end of quiet inspiration to minimize motion artifacts. The scanning range extended from the thoracic inlet to the level of the diaphragm (i.e., from the lung apices to the costophrenic angles), fully covering the lung fields and the cardiac region. Scan parameters were as follows: (1) 100-120 kVp with automatic tube current modulation; and (2) Images were reconstructed using a standard body algorithm with a slice thickness of 1.0-1.5 mm.

AI-driven image biomarker quantification

To assess skeletal status, the analysis relied strictly on the measurement of volumetric BMD (vBMD) using quantitative CT (QCT), rather than material decomposition techniques derived from dual-energy or spectral CT. To ensure quantitative consistency across different CT scanners, this QCT approach was based on prospective calibration using a solid-state calibration phantom containing a hydroxyapatite reference standard (0 mg/cm³, 50 mg/cm³, 100 mg/cm³, 150 mg/cm³, 200 mg/cm³, and 300 mg/cm³) (Supplementary Table 1). For segmentation, thoracic vertebrae T1-T12 were segmented using a three-dimensional (3D) U-Net model (TotalSegmentator) (Supplementary Figure 1). To isolate trabecular bone regions for accurate assessment, cortical bone was excluded by applying a threshold of > 300 Hounsfield unit (HU). Regions corresponding to the vertebral venous plexus and focal osteosclerosis were carefully excluded during segmentation. The final vBMD value was calculated as the mean vBMD across all segmented thoracic vertebrae.

A custom two-stage 3D CNN was employed. Stage 1 localized the cardiac region using a modified 3D U-Net with a “bottleneck” design; stage 2 segmented calcified plaques (> 130 HU) and classified them into coronary branches, including the left anterior descending artery, left circumflex artery, right coronary artery, and left main coronary artery. Total CACS was calculated using the Agatston method (Supplementary Figures 2 and 3). A dedicated 3D U-Net model segmented the pericardial sac. EAT was defined as adipose tissue (-190 to -30 HU) within the pericardial mask. Total EAT volume (cm³) and mean EAT density (HU) were calculated. Detailed model architectures are provided in Supplementary Table 2.

This multitask deep learning framework was developed and validated in a separate multicenter study of 11188 participants, demonstrating strong agreement with clinical reference standards. The framework consisted of three core components: (1) Skeletal assessment, including automatic segmentation of thoracic vertebrae and vBMD calculation, with validation showing a Dice coefficient of 0.975 and a correlation of r = 0.77 with DXA T-scores (Supplementary Figure 4); (2) Coronary artery calcium assessment, using a two-stage deep learning network for automatic cardiac localization, calcified plaque segmentation, and branch classification, with validation achieving weighted kappa coefficients of 0.75-0.88 (Supplementary Table 3, Supplementary Figure 5); and (3) EAT assessment (Supplementary Figure 6), including automatic pericardial segmentation and quantification of EAT volume and density, with validation yielding a Dice coefficient of 0.935. All imaging data were de-identified before processing.

Statistical analysis

Statistical analyses were performed using R (version 4.2.0) and SPSS (version 26.0). Continuous variables are presented as the median (interquartile range) or mean ± SD, and categorical variables are presented as n (%). Intergroup comparisons were performed using the Mann-Whitney U test or the Kruskal-Wallis test for non-normally distributed continuous variables and the χ² test for categorical variables. Post hoc pairwise comparisons were performed using Mann-Whitney U tests with Bonferroni correction when the Kruskal-Wallis test was significant. Propensity score matching (PSM) was performed to minimize selection bias and balance baseline characteristics. A 1:1 nearest-neighbor algorithm was used with a caliper width of 0.2 × SD of the logit propensity score, without replacement. Covariate balance was assessed using standardized mean differences; with standardized mean difference < 0.1 indicating good balance. Multivariable linear and logistic regression models were used to evaluate the independent associations of T2DM with BMD, EAT volume, EAT density, and SCAC risk. A stepwise adjustment strategy was used: (1) Model 1 was unadjusted; (2) Model 2 was adjusted for age and sex; and (3) Model 3 was additionally adjusted for smoking, alcohol consumption, hypertension, dyslipidemia, and BMI. Restricted cubic splines (RCS) were used to explore nonlinear relationships, with knots placed at the 5%, 35%, 65%, and 95%. Subgroup analyses were performed to evaluate the consistency of associations across the T2DM and non-T2DM groups, as well as across subgroups stratified by T2DM duration and GC. Results were visualized using forest plots. All tests were two-sided, and P < 0.05 was considered statistically significant.

RESULTS

Baseline characteristics of study groups

A total of 4589 subjects were included [median age, 55.0 years (interquartile range, 45.0, 63.0); 58.07% male (n = 2665)]. The cohort comprised 1369 T2DM patients and 3220 non-T2DM controls. Baseline characteristics are shown in Table 1. T2DM patients were older, had a higher BMI, and had higher prevalences of male sex, hypertension, and dyslipidemia than non-T2DM controls. T2DM patients also exhibited lower BMD, higher EAT volume and density, and greater CAC burden (all P < 0.001).

Table 1 Baseline characteristics of patients with non-type 2 diabetes mellitus and type 2 diabetes mellitus, n (%)/median (interquartile range).

Variables	Total	Non-T2DM patients	T2DM patients	P value
n	4589	3220	1369
Baseline characteristic
Age (years)	55.00 (45.00, 63.00)	53.00 (41.00, 60.00)	60.00 (53.00, 67.00)	< 0.001
BMI (kg/m2)	23.73 (21.48, 25.95)	23.62 (21.40, 25.84)	24.02 (21.88, 26.22)	< 0.001
Male sex	2665 (58.07)	1756 (54.53)	909 (66.40)	< 0.001
Smoke	946 (20.61)	650 (20.19)	296 (21.62)	0.271
Alcohol	709 (15.45)	543 (16.86)	166 (12.13)	< 0.001
Hypertension	2377 (51.80)	1248 (38.76)	1129 (82.47)	< 0.001
Dyslipidemia	2598 (56.61)	1646 (51.12)	952 (69.54)	< 0.001
Laboratory findings
HDL-cholesterol (mmol/L)	1.09 (0.90, 1.34)	1.14 (0.95, 1.39)	0.95 (0.78, 1.21)	< 0.001
LDL-cholesterol (mmol/L)	2.67 (2.04, 3.37)	2.78 (2.19, 3.49)	2.35 (1.74, 3.06)	< 0.001
Total cholesterol (mmol/L)	4.40 (3.68, 5.36)	4.53 (3.84, 5.45)	4.05 (3.31, 5.12)	< 0.001
Triglyceride (mmol/L)	1.52 (1.03, 2.35)	1.44 (1.00, 2.21)	1.70 (1.13, 2.65)	< 0.001
Fast glucose (mmol/L)	5.44 (4.89, 6.52)	5.18 (4.79, 5.69)	8.01 (6.10, 11.20)	< 0.001
CT findings
BMD (mg/cm3)	173.37 (149.18, 196.47)	176.72 (151.09, 200.71)	167.90 (144.78, 187.61)	< 0.001
EAT volume (cm3)	139.89 (95.68, 196.14)	132.59 (91.86, 184.31)	160.57 (111.34, 215.94)	< 0.001
EAT density (HU)	-77.06 (-82.28, -72.03)	-77.79 (-82.55, -72.99)	-75.07 (-81.41, -69.36)	< 0.001
CACS	0.00 (0.00, 60.53)	0.00 (0.00, 5.02)	61.20 (0.00, 330.42)	< 0.001
CAC category				< 0.001
0	2671 (58.20)	2286 (70.99)	385 (28.12)
1-100	948 (20.66)	566 (17.58)	382 (27.90)
101-300	448 (9.76)	203 (6.30)	245 (17.90)
> 300	522 (11.38)	165 (5.12)	357 (26.08)

T2DM: Type 2 diabetes mellitus; HDL: High-density lipoprotein; LDL: Low-density lipoprotein; BMD: Bone mineral density; EAT: Epicardial adipose tissue; HU: Hounsfield unit; CACS: Coronary artery calcium score; CT: Computed tomography; BMI: Body mass index; CAC: Coronary artery calcification.

Imaging biomarker characteristics within the T2DM cohort

T2DM patients were stratified by disease duration as short (n = 355), medium (n = 311), and long (n = 703). Baseline characteristics are shown in Table 2. Patients with long-duration T2DM were older and had a higher prevalence of hypertension. BMD decreased with diabetes duration (P = 0.013), whereas EAT density and CACS increased (both P < 0.001). EAT volume did not differ across duration groups.

Table 2 Baseline characteristics of patients with type 2 diabetes mellitus stratified by duration, n (%)/median (interquartile range).

Variables	Short-duration	Intermediate-duration	Long-duration	P value
n	355	311	703
Baseline characteristic
Age (years)	57.00 (49.00, 64.00)	59.00 (52.00, 67.00)	61.00 (55.00, 69.00)	< 0.001
BMI (kg/m2)	24.09 (21.60, 26.26)	24.37 (22.03, 26.66)	23.81 (21.88, 26.10)	0.190
Male sex	223 (62.82)	214 (68.81)	472 (67.14)	0.220
Smoke	84 (23.66)	72 (23.15)	140 (19.91)	0.285
Alcohol	38 (10.70)	41 (13.18)	87 (12.38)	0.594
Hypertension	273 (76.90)	242 (77.81)	614 (87.34)	< 0.001
Dyslipidemia	256 (72.11)	206 (66.24)	490 (69.70)	0.257
Laboratory findings
HDL-cholesterol (mmol/L)	0.98 (0.79, 1.28)	0.96 (0.78, 1.21)	0.94 (0.78, 1.16)	0.084
LDL-cholesterol (mmol/L)	2.53 (1.86, 3.43)	2.47 (1.79, 3.00)	2.23 (1.68, 2.97)	< 0.001
Total cholesterol (mmol/L)	4.40 (3.45, 5.70)	4.11 (3.43, 5.10)	3.91 (3.20, 4.86)	< 0.001
Triglyceride (mmol/L)	1.89 (1.21, 3.21)	1.66 (1.14, 2.40)	1.63 (1.08, 2.46)	0.001
HbA1c (%)	7.00 (6.30, 7.10)	7.00 (6.30, 7.00)	6.90 (6.35, 7.10)	0.008
Fast glucose (mmol/L)	7.49 (5.97, 10.52)	7.65 (6.04, 10.56)	8.51 (6.25, 11.72)	0.002
CT findings
BMD (mg/cm3)	171.70 (148.95, 191.93)	168.83 (145.90, 185.79)	166.44 (141.91, 186.12)	0.013
EAT volume (cm3)	157.95 (107.31, 214.17)	167.29 (114.66, 220.84)	160.16 (110.63, 214.58)	0.491
EAT density (HU)	-76.84 (-83.17, -70.48)	-74.91 (-81.98, -69.52)	-74.36 (-80.03, -69.10)	< 0.001
CACS	12.87 (0.00, 145.79)	34.09 (0.00, 260.27)	116.96 (7.63, 496.13)	< 0.001
CAC category				< 0.001
0	139 (39.15)	99 (31.83)	147 (20.91)
1-100	111 (31.27)	89 (28.62)	182 (25.89)
101-300	49 (13.80)	52 (16.72)	144 (20.48)
> 300	56 (15.77)	71 (22.83)	230 (32.72)

HDL: High-density lipoprotein; LDL: Low-density lipoprotein; BMD: Bone mineral density; EAT: Epicardial adipose tissue; HU: Hounsfield unit; CACS: Coronary artery calcium score; CT: Computed tomography; BMI: Body mass index; CAC: Coronary artery calcification; HbA1c: Glycated hemoglobin.

Figure 2 illustrates the distribution of CAC severity across disease-duration groups. CAC severity increased with disease duration. Moderate-to-severe calcification (CACS > 100) was most prevalent in the long-duration group, whereas patients with short-duration T2DM predominantly had absent or mild calcification. Figure 3 shows the distributions of CT-derived parameters across disease-duration groups. CACS increased with disease duration (all P < 0.05). BMD and EAT density differed significantly only between the short-duration and long-duration groups (P ≤ 0.01).

Figure 2 Distribution of coronary artery calcification severity across type 2 diabetes mellitus duration groups. CAC: Coronary artery calcification; T2DM: Type 2 diabetes mellitus.

Figure 3 Comparison of bone mineral density, severe coronary artery calcification, and epicardial adipose tissue density among type 2 diabetes mellitus patients with different durations. A: Severe coronary artery calcification; B: Bone mineral density; C: Epicardial adipose tissue. Data are presented as the median (interquartile range). Comparisons among the three groups were performed using one-way analysis of variance, with P values indicated at the top of each panel. SCAC: Severe coronary artery calcification; BMD: Bone mineral density; EAT: Epicardial adipose tissue; HU: Hounsfield unit.

Comparisons after PSM

Two rounds of PSM were conducted to minimize confounding. First, non-T2DM controls were matched 1:1 with each T2DM duration subgroup (Table 3). After matching, CACS and EAT density remained higher in all T2DM subgroups (all P ≤ 0.006), whereas differences in BMD and EAT volume were no longer significant.

Table 3 Comparison between non-type 2 diabetes mellitus patients and type 2 diabetes mellitus patients with different durations after matching, n (%)/median (interquartile range).

Variables	Non-T2DM vs short-duration T2DM		P value	Non-T2DM vs intermediate-duration T2DM		P value	Non-T2DM vs long-duration T2DM		P value
	Non-T2DM	Short-duration T2DM		Non-T2DM	Intermediate-duration T2DM		Non-T2DM	Long-duration T2DM
n	355	355		308	308		664	664
Age (years)	57.00 (50.00, 64.00)	57.00 (49.00, 64.00)	0.937	59.00 (51.75, 66.00)	59.00 (52.00, 67.00)	0.746	59.00 (54.00, 67.00)	60.00 (54.00, 68.00)	0.133
BMI (kg/m2)	24.02 (21.64, 26.13)	24.09 (21.60, 26.26)	0.874	24.22 (22.35, 26.45)	24.32 (22.03, 26.68)	0.878	24.12 (22.20, 26.23)	23.88 (21.88, 26.18)	0.218
Male sex	203 (57.18)	223 (62.82)	0.125	221 (71.75)	211 (68.51)	0.379	457 (68.83)	447 (67.32)	0.556
Smoke	75 (21.13)	84 (23.66)	0.418	84 (27.27)	71 (23.05)	0.227	137 (20.63)	133 (20.03)	0.785
Alcohol	37 (10.42)	38 (10.70)	0.903	52 (16.88)	41 (13.31)	0.216	82 (12.35)	86 (12.95)	0.741
Hypertension	274 (77.18)	273 (76.90)	0.929	239 (77.60)	239 (77.60)	1.000	574 (86.45)	575 (86.60)	0.936
Dyslipidemia	263 (74.08)	256 (72.11)	0.554	213 (69.16)	203 (65.91)	0.390	458 (68.98)	452 (68.07)	0.723
Fast glucose (mmol/L)	5.25 (4.85, 5.78)	7.49 (5.97, 10.52)	< 0.001	5.24 (4.86, 5.79)	7.66 (6.02, 10.56)	< 0.001	5.24 (4.84, 5.79)	8.57 (6.26, 11.71)	< 0.001
CACS	0.00 (0.00, 27.55)	12.87 (0.00, 145.79)	< 0.001	0.00 (0.00, 34.71)	34.00 (0.00, 257.59)	< 0.001	0.00 (0.00, 66.40)	113.28 (7.31, 470.36)	< 0.001
CAC category			< 0.001			< 0.001			< 0.001
0	212 (59.72)	139 (39.15)		180 (58.44)	98 (31.82)		357 (53.77)	143 (21.54)
1-100	87 (24.51)	111 (31.27)		72 (23.38)	89 (28.90)		164 (24.70)	174 (26.20)
101-300	32 (9.01)	49 (13.80)		33 (10.71)	51 (16.56)		67 (10.09)	139 (20.93)
> 300	24 (6.76)	56 (15.77)		23 (7.47)	70 (22.73)		76 (11.45)	208 (31.33)
BMD (mg/cm3)	167.53 (145.37, 193.04)	171.70 (148.95, 191.93)	0.430	169.01 (145.01, 189.85)	169.32 (146.29, 185.90)	0.874	163.40 (141.82, 185.09)	166.87 (142.84, 186.62)	0.722
EAT volume (cm3)	149.17 (106.21, 202.46)	157.95 (107.31, 214.17)	0.338	166.60 (126.41, 218.08)	166.26 (114.52, 219.02)	0.495	162.74 (120.21, 213.48)	158.94 (109.21, 213.74)	0.144
EAT density (HU)	-78.71 (-83.49, -74.02)	-76.84 (-83.17, -70.48)	0.006	-80.56 (-85.52, -75.47)	-74.88 (-81.88, -69.48)	< 0.001	-79.95 (-84.58, -75.21)	-74.17 (-80.02, -69.11)	< 0.001

T2DM: Type 2 diabetes mellitus; BMI: Body mass index; CACS: Coronary artery calcium score; CAC: Coronary artery calcification; BMD: Bone mineral density; EAT: Epicardial adipose tissue; HU: Hounsfield unit.

Second, PSM was performed across T2DM duration groups for within-group comparisons (Table 4). Patients with long-duration T2DM had higher CACS than those with short-duration T2DM (median, 79.28 vs 18.96, P < 0.001) and medium-duration T2DM (median, 121.49 vs 35.05; P < 0.001). CACS did not differ between the short-duration and medium-duration groups. EAT density was higher in patients with long-duration T2DM than in those with short-duration T2DM (-74.38 HU vs -76.30 HU; P = 0.004); whereas BMD and EAT volume did not differ across duration groups.

Table 4 Comparison between type 2 diabetes mellitus patients with different durations after matching, n (%)/median (interquartile range).

Variables	Short-duration vs intermediate-duration		P value	Short-duration vs long-duration		P value	Intermediate-duration vs long-duration		P value
	Short-duration	Intermediate-duration		Short-duration	Long-duration		Intermediate-duration	Long-duration
n	267	267		310	310		305	305
Age (years)	58.00 (51.00, 66.00)	58.00 (52.00, 65.00)	0.878	57.00 (50.00, 65.00)	59.00 (51.25, 66.00)	0.114	59.00 (52.00, 67.00)	60.00 (54.00, 68.00)	0.082
BMI (kg/m2)	24.04 (21.51, 26.09)	24.37 (22.01, 26.69)	0.148	24.14 (21.54, 25.96)	23.88 (21.89, 26.13)	0.785	24.38 (22.03, 26.67)	23.74 (21.89, 26.12)	0.147
Male sex	181 (67.79)	178 (66.67)	0.782	196 (63.23)	202 (65.16)	0.615	209 (68.52)	201 (65.90)	0.490
Smoke	61 (22.85)	63 (23.60)	0.838	67 (21.61)	74 (23.87)	0.502	69 (22.62)	66 (21.64)	0.770
Alcohol	27 (10.11)	32 (11.99)	0.490	35 (11.29)	42 (13.55)	0.394	39 (12.79)	37 (12.13)	0.806
Hypertension	208 (77.90)	206 (77.15)	0.836	246 (79.35)	251 (80.97)	0.615	241 (79.02)	250 (81.97)	0.358
Dyslipidemia	188 (70.41)	173 (64.79)	0.165	220 (70.97)	222 (71.61)	0.859	200 (65.57)	209 (68.52)	0.438
Fast glucose (mmol/L)	7.49 (6.04, 10.59)	7.66 (6.04, 10.61)	0.621	7.43 (5.97, 10.43)	8.61 (6.52, 11.93)	< 0.001	7.65 (6.04, 10.65)	8.72 (6.52, 11.22)	0.015
HbA1c (%)	7.00 (6.30, 7.10)	7.00 (6.35, 7.00)	0.289	7.00 (6.30, 7.10)	7.00 (6.40, 7.27)	0.775	7.00 (6.30, 7.00)	6.90 (6.40, 7.10)	0.574
CACS	21.14 (0.00, 161.95)	32.04 (0.00, 213.55)	0.426	18.96 (0.00, 164.64)	79.28 (0.00, 326.17)	< 0.001	35.05 (0.00, 265.64)	121.49 (7.85, 492.69)	< 0.001
CAC category			0.697			0.001			0.003
0	96 (35.96)	88 (32.96)		112 (36.13)	84 (27.10)		94 (30.82)	61 (20.00)
1-100	85 (31.84)	80 (29.96)		99 (31.94)	79 (25.48)		89 (29.18)	79 (25.90)
101-300	39 (14.61)	46 (17.23)		46 (14.84)	62 (20.00)		52 (17.05)	67 (21.97)
> 300	47 (17.60)	53 (19.85)		53 (17.10)	85 (27.42)		70 (22.95)	98 (32.13)
BMD (mg/cm3)	166.10 (147.47, 190.49)	170.11 (147.51, 187.26)	0.900	171.64 (149.81, 191.01)	170.02 (144.57, 188.12)	0.222	168.83 (145.61, 185.70)	167.88 (141.91, 186.09)	0.445
EAT volume (cm3)	155.47 (105.92, 220.46)	159.70 (113.23, 210.34)	0.969	156.71 (107.58, 215.17)	151.50 (104.02, 210.74)	0.508	168.09 (115.08, 221.18)	157.13 (110.40, 212.37)	0.276
EAT density (HU)	-75.93 (-82.31, -69.26)	-75.20 (-82.29, -69.59)	0.833	-76.30 (-83.13, -69.80)	-74.38 (-79.66, -69.31)	0.004	-74.95 (-81.93, -69.51)	-74.64 (-79.70, -69.28)	0.204

BMI: Body mass index; HbA1c: Glycated hemoglobin; CACS: Coronary artery calcium score; CAC: Coronary artery calcification; BMD: Bone mineral density; EAT: Epicardial adipose tissue; HU: Hounsfield unit.

Multivariable regression analysis

In unadjusted linear regression (Table 5), T2DM was associated with lower BMD and higher EAT volume. However, these associations lost significance after adjustment for age, sex, and cardiovascular risk factors (models 2 and 3). In contrast, the association between T2DM and EAT density remained robust across all models (model 3: β = 4.131, 95%CI: 3.528-4.733, P < 0.001).

Table 5 Multivariate linear regression analysis of type 2 diabetes mellitus with bone mineral density, epicardial adipose tissue volume, and epicardial adipose tissue density.

Model	BMD		EAT volume		EAT density
	β (95%CI)	P value	β (95%CI)	P value	β (95%CI)	P value
Model 1	-9.579 (-11.617 to -7.542)	< 0.001	24.122 (19.546-28.697)	< 0.001	2.255 (1.711-2.799)	< 0.001
Model 2	1.411 (-0.411 to 3.233)	0.129	4.304 (-0.070 to 8.679)	0.054	3.611 (3.055-4.167)	< 0.001
Model 3	0.506 (-1.477 to 2.489)	0.617	-3.363 (-8.056 to 1.331)	0.16	4.131 (3.528-4.733)	< 0.001

Model 1: Unadjusted. Model 2: Adjusted for age and sex. Model 3: Further adjusted for smoking status, alcohol consumption, hypertension, dyslipidemia, and body mass index. BMD: Bone mineral density; EAT: Epicardial adipose tissue.

Multivariable logistic regression (Table 6) showed that T2DM was a strong independent predictor of SCAC (CACS > 300). In the fully adjusted model, T2DM patients had a 3.656-fold higher risk of SCAC than non-T2DM controls (95%CI: 2.903-4.604, P < 0.001).

Table 6 Multivariate logistic regression analysis of type 2 diabetes mellitus with severe coronary artery calcification.

Variables	SCAC
	OR (95%CI)	P value
Model 1	6.532 (5.360-7.960)	< 0.001
Model 2	4.294 (3.487-5.288)	< 0.001
Model 3	3.656 (2.903-4.604)	< 0.001

Model 1: Unadjusted. Model 2: Adjusted for age and sex. Model 3: Further adjusted for smoking status, alcohol consumption, hypertension, dyslipidemia, and body mass index. SCAC: Severe coronary artery calcification; OR: Odds ratio.

Restricted cubic spline analysis

RCS analysis (Figure 4) revealed complex nonlinear relationships among the biomarkers. BMD and EAT volume showed a linear negative association (P < 0.001; Figure 4D). EAT density and SCAC risk showed a linear positive association (P = 0.007; Figure 4C). The relationship between BMD and EAT density was particularly distinctive, showing a significant J-shaped nonlinear association (P for nonlinearity < 0.001; Figure 4E). No significant associations were observed between BMD or EAT volume and SCAC risk.

Figure 4 Restricted cubic spline analyses of the associations between computed tomography-derived imaging biomarkers. The β coefficients and odds ratios are represented by solid lines, and the 95%CIs are indicated by shaded areas. Knots were positioned at the 5%, 35%, 65%, and 95% of each distribution. A: Bone mineral density (BMD) and severe coronary artery calcification (SCAC); B: Epicardial adipose tissue (EAT) volume and SCAC; C: EAT density and SCAC; D: BMD and EAT volume; E: BMD and EAT density. SCAC: Severe coronary artery calcification; BMD: Bone mineral density; EAT: Epicardial adipose tissue; OR: Odds ratio.

Subsequently, a piecewise linear regression model was used to explore the dose–response relationship between BMD and EAT attenuation (Table 7). Standard linear regression showed a negative correlation. However, the likelihood ratio test indicated that the two-piecewise model provided a significantly better fit (P < 0.001), suggesting a nonlinear relationship. Threshold effect analysis revealed an inflection point at 174.91 mg/cm³. Below this threshold, BMD was negatively associated with EAT attenuation (β = -0.12, 95%CI: -0.13 to -0.10, P < 0.001), whereas above this threshold, no significant association was observed (P = 0.060), corroborating a J-shaped pattern.

Table 7 Threshold effect analysis of the relationship between bone mineral density and epicardial adipose tissue density.

Model	Threshold	β (95%CI)	P value
Model 1: Standard linear regression	-	-0.06 (-0.07 to -0.06)	< 0.001
Model 2: Piecewise linear regression	-	-	-
Threshold	174.91	-	-
< 174.91	-	-0.12 (-0.13 to -0.10)	< 0.001
≥ 174.91	-	0.02 (-0.00 to 0.04)	0.060
Likelihood ratio test P	-	-	< 0.001

Model 1: Simple linear regression. Model 2: Piecewise linear regression fitted after threshold identification using the maximum likelihood method. The likelihood ratio test compares model 2 with model 1, P value < 0.05 indicates a significant threshold effect.

Subgroup analysis

Stratified analyses assessed the consistency of associations across GC and T2DM duration strata (Figure 5). The negative association between BMD and EAT volume was more pronounced in patients with T2DM than in those without T2DM (all P for interaction < 0.05). This negative association was prominent in patients with good and moderate GC, with the strongest effect observed in the moderate GC group. Furthermore, this association weakened with increasing diabetes duration.

Figure 5 Forest plot. A: Forest plot of subgroup analyses for the association of bone mineral density with epicardial adipose tissue density. Data are presented as β coefficients with 95%CI derived from linear regression models; B: Forest plot of subgroup analyses for the association of bone mineral density with severe coronary artery calcification. Data are presented as odds ratio with 95%CI derived from logistic regression models. P values indicate the statistical significance of the associations within each subgroup, while P for interaction assesses whether subgroup variables modify the association. GC: Glycemic control; OR: Odds ratio; T2DM: Type 2 diabetes mellitus.

In contrast, the association between EAT density and SCAC was consistent across all subgroups. No significant effect modification was observed (all P for interaction > 0.05), indicating the robustness of EAT density as an SCAC risk marker.

DISCUSSION

This study leveraged a fully automated AI framework for opportunistic screening using chest QCT in a large multicenter cohort. We simultaneously quantified three cross-system imaging biomarkers (bone density, epicardial fat, and coronary calcification) and investigated their association patterns in T2DM. The main findings were as follows: (1) T2DM patients exhibited higher EAT density and CAC burden after adjustment for cardiovascular risk factors; (2) T2DM was an independent risk factor for high EAT density and SCAC, but not for BMD or EAT volume after multivariable adjustment; and (3) Complex interrelationships were observed, including a linear negative association between BMD and EAT volume, a linear positive association between EAT density and SCAC, and a J-shaped association between BMD and EAT density. These findings provide imaging-based evidence for the shared pathophysiology of skeletal and cardiovascular comorbidities in T2DM.

Although DXA remains the clinical gold standard for BMD assessment[18], we used QCT because it provides three-dimensional volumetric density measurements, enables selective trabecular bone assessment, and is less susceptible to degenerative artifacts and vascular calcification overlap, features particularly relevant to T2DM populations. The opportunistic use of routine chest CT eliminates additional radiation exposure and costs. The strong correlation between our QCT-derived BMD and DXA T-scores (r = 0.77) supports the clinical validity of QCT-derived BMD.

Although electrocardiographic-gated cardiac CT represents the gold standard for CACS[8], we used a non-gated routine chest CT. Our AI framework was validated against gated CT references (kappa = 0.75-0.88), and prior studies support the prognostic value of CAC derived from non-gated chest CT[19,20]. This opportunistic approach maximizes clinical applicability without imposing an additional imaging burden.

These results confirm and extend previous findings on cardiovascular risk in T2DM. Previous studies have reported higher CACS, EAT volume, and EAT density in T2DM patients than in non-T2DM individuals[21,22], as well as diabetes duration as an independent risk factor for SCAC[17]. PSM and multivariable regression demonstrated that the associations of T2DM with EAT density and SCAC persisted after controlling for confounders. This finding suggests that T2DM-specific pathophysiological mechanisms, including hyperglycemia, insulin resistance, and advanced glycation end products, may drive EAT dysfunction and vascular calcification[23,24]. RCS analysis confirmed that increased EAT density, rather than EAT volume, may drive the progression of coronary atherosclerosis. Increased EAT density is considered an imaging manifestation of adipose tissue inflammation and fibrosis, reflecting its pathological transformation from myocardial protective tissue into an inflammatory substrate[25,26]. Gao et al[27] and Liu et al[28] confirmed that higher EAT density independently predicts high-risk coronary plaques better than EAT volume.

A compelling finding is the complex relationship among BMD, EAT volume, and EAT density. Consistent with previous studies, the linear negative correlation between BMD and EAT volume suggests that bone loss and visceral fat accumulation share common upstream drivers, including chronic inflammation, insulin resistance, and related metabolic disorders[29-31]. However, the J-shaped relationship between BMD and EAT density reveals more complex pathophysiological crosstalk. We speculate that at normal or mildly decreased BMD levels, bone metabolism and EAT function remain in relative homeostasis. However, when BMD decreases to a low level (< 174.91 mg/cm³), pathological changes emerge[32,33] such as progressive bone loss accompanied by bone marrow fat expansion and exacerbated systemic inflammation, which may synergistically drive the transformation of EAT into pro-inflammatory and fibrotic tissue[34], ultimately increasing CT density. This finding provides imaging-based evidence that skeletal deterioration may be directly linked to epicardial fat dysfunction, offering new insights into the “bone-heart” dialogue.

Subgroup analysis showed that the negative association between BMD and EAT volume was more pronounced in patients with T2DM, particularly those with moderate GC and shorter disease duration. This finding may be attributed to the more rigorous monitoring and treatment these patients received. Multiple studies have demonstrated that glucagon-like peptide-1 receptor agonists (GLP-1 RAs), as first-line hypoglycemic agents for T2DM, can significantly and rapidly reduce EAT thickness[35], and sodium-glucose cotransporter 2 inhibitors can also markedly decrease EAT volume in patients with T2DM[36,37]. Regarding bone metabolism, evidence shows that patients treated with GLP-1 RAs have a lower risk of osteoporosis than those not treated[38]. In contrast, the positive association between EAT density and SCAC remained robust across all subgroups, suggesting that once EAT inflammation/fibrosis is initiated, its detrimental effect on the vasculature may be persistent, difficult to reverse, and not significantly modulated by GC or disease duration.

This study has several limitations that should be considered when interpreting the results. First, the cross-sectional design precludes causal inference. For example, it remains unclear whether low BMD leads to increased EAT attenuation or whether a common upstream driver simultaneously affects both, and this study cannot determine the temporal sequence or direction of these relationships. Second, selection bias should be considered. The non-T2DM control group was predominantly recruited from individuals undergoing routine health check-ups with chest CT, which may represent a healthier segment of the general population, potentially leading to overestimation of the differences between the diabetic and nondiabetic groups. Third, although we adjusted for various potential confounding factors, unmeasured confounders may still exist, such as detailed medication histories, particularly among patients with T2DM, including metformin, sodium-glucose cotransporter 2 inhibitors, and GLP-1 RAs, which may affect BMD, ectopic adipose tissue deposition, and CAC. Lifestyle factors and specific inflammatory markers may also influence the interpretation of the results. Fourth, the relatively small sample size in certain subgroup analyses, such as the poor GC group, may have reduced statistical power and impaired the assessment of association precision in these specific populations. Fifth, regarding imaging assessment, this study evaluated only CACS and did not include the degree of coronary stenosis, noncalcified plaque volume, or other functional indicators, thereby failing to fully reflect the structural and functional phenotypes of coronary atherosclerosis. Furthermore, although EAT density is considered an imaging biomarker of inflammation, this study was unable to correlate it with serum inflammatory markers, such as interleukin-6 and C-reactive protein, for validation. Sixth, menopausal status was not formally assessed because of limitations in the retrospective data. Given its impact on bone health and adipose tissue distribution, this unmeasured variable may modify associations in female patients and therefore warrants cautious interpretation. Finally, despite adjustments for multiple comparisons, the associations identified in this study lack external validation or mechanistic support, and the underlying biological mechanisms of these hierarchical relationships remain unclear, thereby somewhat weakening the conclusions. Future research should use prospective cohort designs to elucidate causal relationships and incorporate serological markers, more comprehensive imaging assessments, and menopausal status data to validate and extend the combined predictive value identified in this study. Until such data are available, we recommend cautious interpretation of our findings in female patients and encourage consideration of menopausal status in clinical risk assessment.

CONCLUSION

This study developed an AI-based model to assess BMD, CACS, and EAT from chest CT images. For the first time, it systematically characterized, from an imaging perspective, the linear negative association between BMD and EAT volume, as well as the J-shaped nonlinear association between BMD and EAT density, revealing complex nonlinear interrelationships among the skeletal, adipose, and vascular systems in T2DM. These findings highlight the need to move beyond single-organ perspectives in T2DM management and integrate multidimensional skeletal, adipose, and vascular information. Opportunistic multi-indicator assessment based on routine chest CT holds promise as a clinical screening tool for earlier and more precise cross-system risk stratification in T2DM, supporting the future integration of these routine chest CT metrics into T2DM management guidelines to promote personalized prevention strategies.

ACKNOWLEDGEMENTS

We sincerely thank all the research participants of the present study for their invaluable contributions.