Development and validation of a clinical factor-based nomogram for predicting imaging-defined cardiopulmonary abnormality risk in an asymptomatic screening population

Abstract

BACKGROUND

Chronic cardiopulmonary diseases, including chronic obstructive pulmonary disease, interstitial lung disease, and coronary artery disease, represent a major global health burden. Low-dose computed tomography (LDCT) combined with artificial intelligence (AI) quantitative imaging enables the identification of cardiopulmonary imaging abnormalities in asymptomatic individuals.

AIM

To evaluate early risk factors for cardiopulmonary imaging abnormalities in asymptomatic middle-aged and elderly population using single-inspiratory phase LDCT combined with AI-based whole-lung quantitative analysis.

METHODS

A retrospective collection was conducted on 1035 asymptomatic individuals aged ≥ 40 years who underwent routine single-inspiratory-phase LDCT screening at Zhuzhou 331 Hospital in 2025. An AI platform was utilized to automatically extract airway wall area percentage, low-attenuation area percentage, interstitial lung abnormality, and coronary artery calcification score. Based on risk stratification criteria, the population was divided into a high-risk group (n = 689) and a low-risk group (n = 346) and randomly stratified into a training set (n = 724) and a validation set (n = 311) at a 7:3 ratio. Independent sample t-test or χ² test was applied to analyze between-group differences. Binary logistic regression was used to screen independently associated factors, and a multivariate logistic regression model was constructed after excluding collinear variables using variance inflation factor (VIF < 5), followed by the establishment of a nomogram prediction model. The receiver operating characteristic curve and DeLong test were employed to evaluate model discrimination. The Hosmer-Lemeshow test and calibration curves were used to assess goodness of fit. Decision curve analysis (DCA) was applied to evaluate clinical net benefit, and internal validation was performed using the bootstrap method (1000 resamplings).

RESULTS

Univariate analysis showed that smoking history, abnormal metabolic status, body mass index (BMI), age, and gender were significantly associated with cardiopulmonary imaging-defined high-risk status (P < 0.05), while work style showed a marginal association in univariate analysis (P = 0.043) but did not retain statistical significance in the multivariate model (P = 0.851). After VIF collinearity screening (all VIF < 5) and multivariate logistic regression analysis with forced entry of six candidate variables, smoking history [odds ratio (OR) = 1.968 per level, 95% confidence interval (CI): 1.665-2.325, P < 0.001], abnormal metabolic status (OR = 3.266, 95%CI: 2.259-4.723, P < 0.001), BMI (OR = 1.150 per unit, 95%CI: 1.079-1.226, P < 0.001), and age (OR = 1.028 per year, 95%CI: 1.005-1.052, P = 0.018) were identified as independently associated factors for cardiopulmonary imaging-defined high-risk status, while male gender demonstrated an (inverse) association (OR = 0.427, 95%CI: 0.276-0.660, P < 0.001) after adjustment for smoking and other covariates. The nomogram-based risk stratification model integrating the above five independently associated factors achieved an area under the curve (AUC) of 0.833 (95%CI: 0.787-0.879) in the validation set, significantly outperforming the baseline model containing only age and gender (AUC = 0.647, 95%CI: 0.582-0.712; DeLong test: Delta AUC = 0.186, z = 5.256, P < 0.001). The Hosmer-Lemeshow goodness-of-fit test confirmed satisfactory calibration (training set: χ² = 10.40, P = 0.238; validation set: χ² = 6.50, P = 0.591). Calibration curves and DCA demonstrated good agreement and positive clinical net benefit. Bootstrap internal validation (1000 resamples) confirmed model robustness (mean AUC = 0.792, 95%CI: 0.760-0.824).

CONCLUSION

The LDCT-based nomogram model combined with AI quantitative imaging analysis may assist in identifying individuals with cardiopulmonary imaging abnormalities in asymptomatic screening populations, potentially guiding further diagnostic evaluation. Prospective studies are warranted to establish its role in improving clinical outcomes.

Key Words: Computed tomography; Artificial intelligence; Cardiopulmonary imaging abnormalities; Middle-aged and elderly population; Opportunistic screening

Core Tip: This study developed and validated a clinical factor-based nomogram for predicting imaging-defined cardiopulmonary imaging abnormalities in 1035 asymptomatic screening individuals using artificial intelligence-assisted low-dose computed tomography quantitative analysis. Five independently associated factors were identified: Smoking history, abnormal metabolic status, body mass index, age, and gender. The nomogram achieved an area under the curve of 0.833 in the validation set. This model may assist early, non-invasive risk stratification for cardiopulmonary imaging abnormalities during routine lung cancer screening, potentially guiding timely preventive interventions in high-risk populations.

INTRODUCTION

Chronic obstructive pulmonary disease (COPD), interstitial lung disease (ILD), and coronary artery disease (CAD) collectively constitute a major chronic disease burden among the middle-aged and elderly population worldwide[1]. The Global Initiative for Chronic Obstructive Lung Disease (GOLD) 2025 report explicitly states that chest computed tomography (CT) has evolved from a mere diagnostic imaging tool into a key technique for identifying distinct clinical phenotypes of COPD, such as the airway-predominant and emphysema-predominant subtypes[2]. The cardiopulmonary system is physiologically interdependent, with shared risk factors and overlapping pathophysiological pathways.

Currently, low-dose CT (LDCT) of the chest is a routine imaging modality for lung cancer screening. However, conventional interpretation paradigms focus predominantly on pulmonary nodule detection, overlooking subclinical markers embedded within the vast imaging data, such as airway remodeling, parenchymal destruction, and coronary artery calcification. Adopting the concept of “opportunistic screening”[3], the present study retrospectively analyzes multimodal quantitative imaging data from 1035 patients to elucidate the mapping patterns between clinical factors and imaging phenotypes, and to develop a risk prediction model leveraging artificial intelligence (AI)-assisted quantitative analysis, thereby providing quantitative decision support for precision prevention in asymptomatic populations. However, systematic evaluation of cardiopulmonary imaging abnormalities using AI-assisted quantitative CT in asymptomatic screening populations remains limited.

MATERIALS AND METHODS

Study population

This study retrospectively enrolled individuals who underwent routine single-inspiratory-phase LDCT health examination screening at Zhuzhou 331 Hospital between March 1, 2025 and October 30, 2025. The study was approved by the hospital ethics committee (Approval No. 202503-K1-J1). Inclusion criteria were: (1) Age ≥ 40 years; (2) Underwent single-inspiratory-phase LDCT scanning with adequate image quality meeting AI segmentation requirements; and (3) Complete clinical and epidemiological survey data. Exclusion criteria were: (1) Previously diagnosed with COPD, ILD, or CAD; (2) History of pulmonary surgery or radiation therapy; (3) Concurrent severe heart failure, acute or chronic renal failure, or malignancy; and (4) Pregnant or lactating women. A total of 1035 subjects were ultimately enrolled, including 680 males (65.70%) and 355 females (34.30%), with a mean age of (52.6 ± 8.0) years.

Examination methods and AI quantitative workflow

LDCT examination: Scanning was performed using a GE Optima CT660 CT scanner. Subjects were placed in the supine position with both arms elevated, and inspiratory-phase scanning was conducted during breath-holding after deep inspiration. Scanning parameters: Tube voltage 120 kV, automatic tube current modulation (50 mA), FOV 350 mm × 350 mm, interslice gap 5 mm, slice thickness 5 mm, reconstruction slice thickness 0.625 mm, reconstruction interval 0.625 mm.

AI quantitative analysis: Data were processed using the SIMBIO-T 1.0.0 (Hangzhou Shimai Intelligent Technology Co., Ltd.) deep learning-based whole-lung quantitative analysis platform. The SIMBIO-T chest CT quantitative analysis system (software copyright registration No. 2024SR1939590) was used for automated analysis. All quantitative AI-generated results were reviewed by a radiologist; cases with suboptimal segmentation were reprocessed until meeting review standards. The quantitative analysis functions of this platform have been validated in multicenter settings.

(1) Airway remodeling assessment: Automatic segmentation of the bronchial tree was performed to precisely calculate the wall area percentage (WA%)[4], quantifying the degree of airway wall thickening; (2) Lung parenchymal destruction assessment: Based on density mask technique with a threshold of -950 HU, the whole-lung low-attenuation area percentage (LAA%)[5] was calculated as a quantitative indicator of emphysematous changes; (3) Interstitial lung abnormality (ILA) assessment: A deep learning texture analysis model was used to automatically detect and identify ILA patterns[6], generating a semi-quantitative ILA score to assess the severity of interstitial changes; and (4) Subclinical atherosclerosis assessment: Based on the same non-contrast CT data, a deep learning model was used to automatically localize the coronary arteries and calculate the Agatston coronary artery calcification score (CACS)[7] for cardiovascular disease risk stratification (Figure 1).

Figure 1 Representative outputs from the SIMBIO-T 1.0.0 artificial intelligence-based whole-lung quantitative analysis platform. (Hangzhou Shimai Intelligent Technology Co., Ltd.): Low-attenuation area percentage, interstitial lung abnormality, airway wall area percentage, and coronary artery calcification score.

Observation indicators and risk stratification definition

Clinical and epidemiological survey data were systematically collected, including gender, age, long-term work style, body mass index (BMI), smoking history, alcohol consumption history, sleep duration, New York Heart Association (NYHA) functional class, abnormal metabolic status, occupational exposure history, family history of chronic cardiopulmonary disease, family history of malignancy, and high-fat high-salt dietary habits.

Smoking index (pack-years) was calculated as (daily cigarettes smoked ÷ 20) × years of smoking, then categorized into four levels: Non-smoker (0 pack-years), mild (≤ 200), moderate (200 < smoking index < 400), and heavy (≥ 400). Alcohol consumption was self-reported and categorized by frequency: Non-drinker, rarely (< 1 time/month), light (2-4 times/month), moderate (2-3 times/week), and heavy (≥ 4 times/week). NYHA functional class was assessed by the examining physician according to standard criteria. High-fat high-salt dietary habits were defined by self-report of habitual intake of fried foods, pickled foods, or added salt at meals on most days of the week. Abnormal metabolic status was defined as any one of fasting blood glucose, blood lipids (total cholesterol or triglycerides), or blood pressure (systolic or diastolic) exceeding the respective normal reference range per institutional laboratory standards. Based on the composite risk definition established by CT quantitative imaging research consensus and prior studies, the population was stratified according to the composite risk of chronic cardiopulmonary comorbidity: Individuals meeting any one of the criteria: LAA% ≥ 6%[8], WA% > 60%[9], ILA+[6], or CACS > 100[10], were defined as the high-risk group for cardiopulmonary imaging abnormality phenotype (n = 689); the remaining individuals comprised the low-risk group (n = 346).

Statistical analysis

Statistical analysis was implemented using Python code. The statistical workflow was as follows: (1) Continuous data were expressed as mean ± SD, and between-group differences were compared using the independent sample t-test; categorical data were expressed as n (%), and between-group differences were compared using the χ² test; (2) Univariate logistic regression was used to screen risk factors (P < 0.05); (3) Variance inflation factor (VIF) was introduced for collinearity diagnosis, with VIF < 5 as the threshold to exclude collinear variables; (4) Multivariate logistic regression was applied to determine independently associated factors and calculate odds ratio (OR) and 95% confidence interval (CI); and (5) A nomogram prediction model was constructed, with its discrimination evaluated using the receiver operating characteristic curve and DeLong test, and compared with a baseline model containing only age and gender; the Hosmer-Lemeshow test and calibration curves were used to assess goodness of fit and calibration, decision curve analysis (DCA) was applied to evaluate clinical net benefit, and internal validation was performed using the bootstrap method (1000 resamplings). P < 0.05 was considered statistically significant.

The study followed the TRIPOD reporting guideline. No missing data were present in the final analytical dataset. With 482 events and six candidate variables (entered simultaneously), the events-per-variable ratio was 80.3, exceeding the recommended minimum of 10. The full model equation was as follows: Logit (P) = -4.846 + (-0.844 × gender) + (0.027 × age) + (0.140 × BMI) + (1.186 × metabolic) + (0.678 × smoking). Statistical analyses were performed using Python with the stats models and scikit-learn packages. The validation set was used only once for final model evaluation.

RESULTS

Differences in clinical characteristics between the high-risk and low-risk groups

Clinical baseline analysis of the 1035 enrolled subjects (Table 1) revealed: (1) The proportion of males was significantly higher in the high-risk group (69.376% vs 58.382%, χ² = 11.872, P < 0.001); (2) The mean age of the high-risk group was significantly greater than that of the low-risk group (53.546 ± 8.136 years vs 50.812 ± 7.361 years, t = 5.261, P < 0.001); (3) The mean BMI of the high-risk group was significantly higher (24.802 ± 3.070 kg/m² vs 23.135 ± 2.653 kg/m², t = 8.612, P < 0.001); (4) The proportion of subjects with abnormal metabolic status in the high-risk group (62.845%) was substantially greater than that of the low-risk group (26.301%, χ² = 121.603, P < 0.001); (5) The distribution of smoking history differed significantly between groups across all smoking index categories (χ² = 134.002, P < 0.001); and (6) Significant differences were observed between the two groups in the distribution of work style categories (χ² = 14.008, P = 0.001).

Table 1 Clinical baseline characteristics of 1035 asymptomatic health examination individuals aged 40 years and above, n (%).

Index	Total (n = 1035)	Low-risk group (n = 346)	High-risk group (n = 689)	t/χ2/z	P value
Age	52.632 ± 7.987	50.812 ± 7.361	53.546 ± 8.136	5.261	< 0.001
BMI	24.244 ± 3.040	23.135 ± 2.653	24.802 ± 3.070	8.612	< 0.001
Gender	-	-	-	11.872	< 0.001
Male	680 (65.700)	202 (58.382)	478 (69.376)	-	-
Female	355 (34.300)	144 (41.618)	211 (30.624)	-	-
Metabolic status	-	-	-	121.603	< 0.001
Abnormal	524 (50.628)	91 (26.301)	433 (62.845)	-	-
Normal	511 (49.372)	255 (73.699)	256 (37.155)	-	-
Smoking history	-	-	-	134.002	< 0.001
Non-smoker	465 (44.928)	242 (69.942)	223 (32.366)	-	-
Mild (smoking index ≤ 200)	134 (12.947)	32 (9.249)	102 (14.804)	-	-
Moderate (200 < smoking index < 400)	91 (8.792)	16 (4.624)	75 (10.885)	-	-
Heavy (smoking index ≥ 400)	345 (33.333)	56 (16.185)	289 (41.945)	-	-
Alcohol consumption history	-	-	-	3.191	0.526
Non-drinker	557 (53.816)	197 (56.936)	360 (52.250)	-	-
Rarely (< 1/month)	211 (20.386)	64 (18.497)	147 (21.335)	-	-
Light (2-4/month)	145 (14.010)	50 (14.451)	95 (13.788)	-	-
Moderate (2-3/week)	54 (5.217)	15 (4.335)	39 (5.660)	-	-
Heavy (≥ 4/week)	68 (6.570)	20 (5.780)	48 (6.967)	-	-
Work style	-	-	-	14.008	0.001
Sedentary behavior	368 (35.556)	148 (42.775)	220 (31.930)	-	-
Light physical activity	231 (22.319)	60 (17.341)	171 (24.819)	-	-
Moderate-to-high physical activity	436 (42.126)	138 (39.884)	298 (43.251)	-	-
High-oil high-salt diet	-	-	-	0.756	0.385
No	885 (85.507)	301 (86.994)	584 (84.761)	-	-
Yes	150 (14.493)	45 (13.006)	105 (15.239)	-	-
Sleep duration	-	-	-	1.043	0.594
< 6 hours	57 (5.507)	16 (4.624)	41 (5.951)	-	-
6-8 hours	546 (52.754)	188 (54.335)	358 (51.959)	-	-
> 8 hours	432 (41.739)	142 (41.040)	290 (42.090)	-	-
NYHA class	-	-	-	1.867	0.600
NYHA Class I	969 (93.623)	327 (94.509)	642 (93.179)	-	-
NYHA Class II	60 (5.797)	18 (5.202)	42 (6.096)	-	-
NYHA Class III	3 (0.290)	1 (0.289)	2 (0.290)	-	-
NYHA Class IV	3 (0.290)	0 (0.000)	3 (0.435)	-	-
Family history of heart disease	-	-	-	1.437	0.231
No	650 (62.802)	208 (60.116)	442 (64.151)	-	-
Yes	385 (37.198)	138 (39.884)	247 (35.849)	-	-
Occupational exposure history	-	-	-	0.019	0.891
No	984 (95.072)	328 (94.798)	656 (95.210)	-	-
Yes	51 (4.928)	18 (5.202)	33 (4.790)	-	-
Family history of lung cancer	-	-	-	0.000	1.000
No	978 (94.493)	327 (94.509)	651 (94.485)	-	-
Yes	57 (5.507)	19 (5.491)	38 (5.515)	-	-

Work style was self-reported and originally classified into five categories. Due to the very small sample size in the high-intensity physical activity category (< 10 cases), the five categories were merged into three: Sedentary behavior (reference), light physical activity, and moderate-to-high physical activity. Abnormal metabolic status was defined as any one of fasting blood glucose, blood lipids (total cholesterol or triglycerides), or blood pressure exceeding the normal reference range. Smoking history was categorized by smoking index (pack-years) into four levels. NYHA: New York Heart Association; BMI: Body mass index.

To further characterize the composite outcome, the distribution of component imaging abnormalities was as follows: Low-risk group (no imaging abnormality), n = 346 (33.4%); pulmonary imaging high-risk only (LAA% ≥ 6% or WA% > 60% or ILA⁺), n = 288 (27.8%); cardiovascular imaging high-risk only (CACS > 100), n = 45 (4.3%); and both pulmonary and cardiovascular imaging high-risk, n = 356 (34.4%). The high overall prevalence (66.6%) is attributable to the nature of the study cohort, comprising self-referred health examination participants who systematically differ from the general population in terms of demographic characteristics and risk factor profiles.

Validation of clinical characteristics balance between the training and validation sets

The 1035 subjects were randomly stratified into a training set (n = 724, 482 high-risk and 242 low-risk) and a validation set (n = 311, 207 high-risk and 104 low-risk) at a 7:3 ratio. No statistically significant differences were found in any clinical characteristics between the two sets (P > 0.05), indicating good balance in the random allocation (Table 2).

Table 2 Allocation of training and validation sets among 1035 asymptomatic health examination individuals aged 40 years and above, n (%).

Index	Training set (n = 724)	Validation set (n = 311)	t/χ2/z	P value
Age	52.481 ± 7.874	52.984 ± 8.247	-0.929	0.353
BMI	24.255 ± 2.983	24.220 ± 3.173	0.168	0.866
Sex	-	-	1.716	0.190
Male	466 (64.365)	214 (68.810)	-	-
Female	258 (35.635)	97 (31.190)	-	-
Metabolic status	-	-	0.070	0.791
Abnormal	369 (50.967)	155 (49.839)	-	-
Normal	355 (49.033)	156 (50.161)	-	-
Smoking history	-	-	5.055	0.168
Non-smoker	335 (46.271)	130 (41.801)	-	-
Mild (smoking index ≤ 200)	85 (11.740)	49 (15.756)	-	-
Moderate (200 < smoking index < 400)	68 (9.392)	23 (7.395)	-	-
Heavy (smoking index ≥ 400)	236 (32.597)	109 (35.048)	-	-
Alcohol consumption history	-	-	3.908	0.419
Non-drinker	394 (54.420)	163 (52.412)	-	-
Rarely (< 1/month)	154 (21.271)	57 (18.328)	-	-
Light (2-4/month)	95 (13.122)	50 (16.077)	-	-
Moderate (2-3/week)	38 (5.249)	16 (5.145)	-	-
Heavy (≥ 4/week)	43 (5.939)	25 (8.039)	-	-
Work style	-	-	1.285	0.526
Sedentary behavior	254 (35.083)	114 (36.656)	-	-
Light physical activity	157 (21.685)	74 (23.794)	-	-
Moderate-to-high physical activity	313 (43.232)	123 (39.550)	-	-
High-oil high-salt diet	-	-	2.127	0.145
No	611 (84.392)	274 (88.103)	-	-
Yes	113 (15.608)	37 (11.897)	-	-
Sleep duration	-	-	1.514	0.469
< 6 hours	36 (4.972)	21 (6.752)	-	-
6-8 hours	381 (52.624)	165 (53.055)	-	-
> 8 hours	307 (42.403)	125 (40.193)	-	-
NYHA class	-	-	3.284	0.350
NYHA Class I	679 (93.785)	290 (93.248)	-	-
NYHA Class II	41 (5.663)	19 (6.109)	-	-
NYHA Class III	3 (0.414)	0 (0.000)	-	-
NYHA Class IV	1 (0.138)	2 (0.643)	-	-
Family history of heart disease	-	-	0.200	0.655
No	451 (62.293)	199 (63.987)	-	-
Yes	273 (37.707)	112 (36.013)	-	-
Occupational exposure history	-	-	0.136	0.713
No	690 (95.304)	294 (94.534)	-	-
Yes	34 (4.696)	17 (5.466)	-	-
Family history of lung cancer	-	-	0.035	0.852
No	683 (94.337)	295 (94.855)	-	-
Yes	41 (5.663)	16 (5.145)	-	-

No statistically significant differences were observed between the training and validation sets for any variable (all P > 0.05), indicating adequate balance of random allocation. NYHA: New York Heart Association; BMI: Body mass index.

Screening for independent risk factors of chronic cardiopulmonary disease

Univariate logistic regression analysis of 13 clinical characteristics in the training set showed that gender (OR = 1.483, 95%CI: 1.078-2.039, P = 0.015), age (OR = 1.041 per year, 95%CI: 1.020-1.063, P < 0.001), BMI (OR = 1.187 per unit, 95%CI: 1.121-1.256, P < 0.001), abnormal metabolic status (OR = 4.015, 95%CI: 2.877-5.602, P < 0.001), smoking history (OR = 1.819 per level, 95%CI: 1.587-2.085, P < 0.001), and work style (OR = 1.199, 95%CI: 1.006-1.429, P = 0.043) were significantly associated with cardiopulmonary imaging-defined high-risk status. VIF collinearity diagnosis of the above six significant variables showed that all VIF values ranged from 1.032 to 1.397, confirming the absence of significant collinearity (Tables 3 and 4).

Table 3 Variance inflation factor analysis of clinical characteristics with significant differences in the study population.

Variable	VIF	Assessment (< 5)
Sex	1.397	Pass
Age	1.032	Pass
BMI	1.098	Pass
Metabolic status	1.111	Pass
Smoking history	1.334	Pass
Work style	1.101	Pass

All variance inflation factor values were below 5, indicating no significant multicollinearity among the clinical characteristics, confirming their suitability for inclusion in the multivariate regression model. BMI: Body mass index; VIF: Variance inflation factor.

Table 4 Univariate and multivariate binary logistic regression analysis of clinical characteristics.

Variable	Univariate analysis			Multivariate (enter) regression
	OR	95%CI	P value	OR	95%CI	P value
Sex	1.483	1.078-2.039	0.015	0.427	0.276-0.660	< 0.001
Age	1.041	1.020-1.063	< 0.001	1.028	1.005-1.052	0.018
BMI	1.187	1.121-1.256	< 0.001	1.150	1.079-1.226	< 0.001
Metabolic status	4.015	2.877-5.602	< 0.001	3.266	2.259-4.723	< 0.001
Smoking history	1.819	1.587-2.085	< 0.001	1.968	1.665-2.325	< 0.001
Work style	1.199	1.006-1.429	0.043	1.020	0.829-1.255	0.851
Alcohol consumption history	1.058	0.927-1.208	0.403
High-oil high-salt diet	1.326	0.852-2.062	0.211
Sleep duration	0.938	0.718-1.227	0.641
NYHA class	1.363	0.752-2.470	0.307
Family history of heart disease	0.775	0.565-1.063	0.113
Occupational exposure history	0.802	0.395-1.631	0.543
Family history of lung cancer	1.227	0.615-2.449	0.562

The multivariate analysis employed the forced entry (enter) method, incorporating all six variables identified as significant in univariate screening. Work style did not achieve statistical significance in the multivariate model (P = 0.851). The final nomogram incorporated five independently associated factors: Gender, age, body mass index, metabolic status, and smoking history. NYHA: New York Heart Association; BMI: Body mass index; OR: Odds ratio; CI: Confidence interval.

Multivariate logistic regression analysis with forced entry of all six candidate variables (Table 4) revealed the following: Smoking history demonstrated prominent association strength (OR = 1.968 per level, 95%CI: 1.665-2.325, P < 0.001), showing a monotonic dose-response gradient across smoking index categories; abnormal metabolic status (OR = 3.266, 95%CI: 2.259-4.723, P < 0.001), elevated BMI (OR = 1.150 per unit, 95%CI: 1.079-1.226, P < 0.001), and increasing age (OR = 1.028 per year, 95%CI: 1.005-1.052, P = 0.018) also significantly increased the probability of cardiopulmonary imaging-defined high-risk status; male sex showed an inverse association (OR = 0.427, 95%CI: 0.276-0.660, P < 0.001) after adjustment for smoking and other covariates, with female sex as the reference category; work style did not attain statistical significance in the multivariate model (OR = 1.020, 95%CI: 0.829-1.255, P = 0.851).

Performance evaluation and internal validation of the prediction model

A nomogram-based risk stratification model was constructed using the above five independently associated factors (Figure 2). The area under the curve (AUC) of this model was 0.788 (95%CI: 0.754-0.822) in the training set and 0.833 (95%CI: 0.787-0.879) in the validation set (Figure 3A). The DeLong test demonstrated that the full model significantly outperformed the baseline model containing only age and gender (AUC = 0.647, 95%CI: 0.582-0.712; Delta AUC = 0.186, z = 5.256, P < 0.001) (Figure 3B). The Hosmer-Lemeshow goodness-of-fit test confirmed satisfactory calibration in both the training set (χ² = 10.40, P = 0.238) and the validation set (χ² = 6.50, P = 0.591) (Figure 3C). Calibration curves demonstrated good agreement between estimated probabilities and observed frequencies. DCA showed that the model yielded a positive net benefit across a threshold probability range of 0.10 to 0.90 (Figure 3D). Bootstrap internal validation (1000 resamples) on the training set confirmed model robustness (mean AUC = 0.792, 95%CI: 0.760-0.824) (Figure 4).

Figure 2 Nomogram of the comprehensive clinical prediction model, including independent risk factors: Sex, age, Body mass index, metabolic status, and smoking history. Sex (1 = male, 0 = female); metabolic status (1 = abnormal metabolic status, 0 = normal metabolic status); smoking history [0 = non-smoker, 1 = mild (smoking index ≤ 200), 2 = moderate (200 < smoking index < 400), 3 = heavy (smoking index ≥ 400)]. BMI: Body mass index.

Figure 3 Performance evaluation of the comprehensive clinical prediction model. A: Receiver operating characteristic curve; B: DeLong test; C: Calibration curves of the training and validation sets; D: Decision curve analysis. AUC: Area under the curve; CI: Confidence interval.

Figure 4 Bootstrap internal resampling validation of the training set model (number of resamplings = 1000). AUC: Area under the curve; ROC: Receiver operating characteristic.

Based on the Youden index calculated in the training set, the optimal probability threshold for the nomogram was 0.724. At this cut point, the model achieved a sensitivity of 57.5%, specificity of 83.7%, positive predictive value (PPV) of 87.5%, and negative predictive value (NPV) of 49.7% in the validation set. The corresponding training set achieved 58.3% sensitivity, 81.0% specificity, 85.9% PPV, and 49.4% NPV (Table 5).

Table 5 Model performance metrics at optimal cut point (Youden Index).

Metric	Training set (n = 724) (%)	Validation set (n = 311) (%)
Optimal threshold (probability)	-	0.724
Sensitivity	58.3	57.5
Specificity	81	83.7
Positive predictive value	85.9	87.5
Negative predictive value	49.4	49.7

DISCUSSION

This study utilized a deep learning-based AI quantitative imaging analysis platform to define cardiopulmonary imaging-defined high-risk status in 1035 asymptomatic health examination individuals and constructed a visual risk stratification model based on independently associated clinical factors.

Upgrading the LDCT screening paradigm: AI-driven multi-dimensional imaging phenotypic analysis

According to the latest strategy of the GOLD 2026, COPD has been defined as a heterogeneous disease driven by genetic, developmental, and environmental factors, with emphasis on incorporating imaging biomarkers to define “low disease activity state”[11]. Conventional LDCT screening has largely been confined to pulmonary nodule detection, often overlooking the rich biological information embedded in airway, lung parenchymal, and cardiovascular structural changes[12]. This study employed a deep learning-based AI platform to automatically extract pulmonary and cardiovascular quantitative imaging indicators-WA%, LAA%, ILA, and CACS-extending the assessment dimensions of LDCT from solitary nodule detection to multi-dimensional cardiopulmonary quantitative analysis. This paradigm provides simultaneous early warning signals for cardiovascular risk, lung parenchymal damage, and airway remodeling without increasing radiation dose to the subjects, thereby maximizing the information yield of a single screening examination.

Clinical factors independently associated with cardiopulmonary imaging-defined high-risk status

Multivariate logistic regression analysis identified five factors independently associated with cardiopulmonary imaging-defined high-risk status: Smoking history (OR = 1.968 per level, 95%CI: 1.665-2.325, P < 0.001), and demonstrated a stepwise increasing trend across smoking index categories[13]; abnormal metabolic status (OR = 3.266, 95%CI: 2.259-4.723, P < 0.001)[14,15]; elevated BMI (OR = 1.150 per unit increase, 95%CI: 1.079-1.226, P < 0.001); older age (OR = 1.028 per year, 95%CI: 1.005-1.052, P = 0.018)[16]; and male sex (OR = 0.427, 95%CI: 0.276-0.660, P < 0.001, with female as reference), which showed an inverse association after adjustment for smoking and other covariates.

A noteworthy finding of this study is the reversal of the sex effect direction between univariate and multivariate analyses. In univariate analysis, male sex was positively associated with high-risk status (OR approximately 1.48), whereas in the multivariate model, after adjustment for smoking and other covariates, male sex emerged with an inverse association (OR = 0.427, 95%CI: 0.276-0.660, P < 0.001)[17]. This reversal is most plausibly explained by the strong confounding effect of smoking, although the possibility of over-adjustment cannot be entirely excluded. In this cohort, 74.4% (506/680) of males were smokers, compared with only 18.0% (64/355) of females. The crude univariate association for male sex predominantly reflected the confounding effect of differential smoking prevalence between sexes, a manifestation of the suppression effect (Simpson’s paradox analogue) in observational research. After statistically adjusting for smoking, the direction of the sex association reversed, revealing a residual male-inverse association. However, the possibility of residual confounding or over-adjustment cannot be entirely excluded. While the precise biological basis of this residual association cannot be determined from cross-sectional data, this finding underscores the importance of multivariable adjustment in observational studies and cautions against overinterpretation of crude (unadjusted) associations. Future prospective studies with more detailed behavioral and biological phenotyping are warranted to further elucidate the interplay between gender, smoking, and cardiopulmonary imaging phenotypes.

Elevated BMI was identified as an independently associated factor for cardiopulmonary imaging-defined high-risk status (OR = 1.150 per unit, 95%CI: 1.079-1.226, P < 0.001). The positive association between BMI and cardiopulmonary imaging abnormalities observed in this study suggests that BMI should be considered alongside other clinical factors when evaluating cardiopulmonary risk profiles in asymptomatic screening populations. Given that this study employed single-inspiratory-phase CT scanning, the potential influence of obesity-related lung volume changes on density-based CT indicators (such as LAA%)[18,19] could not be fully assessed; future studies with dual-phase CT acquisition may further clarify this relationship.

Visualized decision translation and clinical net benefit of the risk stratification model

The nomogram-based risk stratification model using the above five independently associated factors achieved an AUC of 0.833 (95%CI: 0.787-0.879) in the validation set and 0.788 (95%CI: 0.754-0.822) in the training set. The validation set AUC (0.833) was modestly higher than the training set AUC (0.788). As the two sets are derived from a single random split and are not independent, no formal statistical comparison is appropriate. The bootstrap-derived mean AUC of 0.792 (95%CI: 0.760-0.824) falls between the two points estimates, consistent with expected sampling variability. The Hosmer-Lemeshow test confirmed satisfactory calibration in both the training set (χ² = 10.40, P = 0.238) and the validation set (χ² = 6.50, P = 0.591). The DeLong test demonstrated that the full model significantly outperformed the baseline model containing only age and gender (AUC = 0.647, 95%CI: 0.582-0.712; delta AUC = 0.186, z = 5.256, P < 0.001). In clinical application, the nomogram model requires only five easily obtainable clinical indicators to complete risk assessment, age, gender, BMI, smoking history, and metabolic status, making it suitable for rapid risk stratification of asymptomatic middle-aged and elderly individuals in health examination settings, and facilitating the identification of high-risk individuals who may benefit from further LDCT examination with AI-assisted multi-dimensional imaging evaluation.

Limitations and prospects

This study has the following limitations: (1) Although LAA% ≥ 6% is a Fleischner Society-endorsed threshold and all scans were performed under end-inspiratory breath-hold with technologist verification, residual inter-subject variability in inspiratory effort remains an inherent limitation of single-phase CT densitometry[20]; (2) Study design: This is a cross-sectional observational study that can only establish associations between clinical factors and cardiopulmonary imaging-defined high-risk status and cannot infer causal relationships; (3) Selection bias: The study population was derived from a single-center health examination cohort, comprising self-referred health examination participants who may differ systematically from the general population (healthy-volunteer bias); multicenter external validation involving diverse scanners, AI platforms, and populations is required before clinical deployment; (4) Pulmonary function tests were not available to corroborate the CT-defined abnormalities with spirometric diagnoses; (5) CACS was derived from non-electrocardiography-gated LDCT with a standard lung screening reconstruction kernel rather than a dedicated sharp cardiac kernel, which may underestimate calcification burden; independent validation data for the WA% and CACS modules of the AI platform have not been published; (6) The cross-sectional design precludes assessment of longitudinal outcomes, such as progression to clinical disease; and (7) Metabolic status was captured as a binary composite variable without individual component values or waist circumference, precluding analysis using standard metabolic syndrome definitions.

CONCLUSION

In conclusion, this study integrated LDCT with AI quantitative imaging analysis technology, achieving early identification of independently associated clinical factors for cardiopulmonary imaging abnormalities in asymptomatic middle-aged and elderly health examination populations. A nomogram-based visual risk assessment model with good discrimination and clinical net benefit was constructed and validated. This model may assist clinicians in identifying high-risk individuals and guiding stratified management, thereby promoting optimal allocation of healthcare resources and improvement of patient outcomes.