Clustering characteristics of upper gastrointestinal cancer risk behaviours and their association with social determinants of health: a latent class analysis

Abstract

Research on risk behaviours (RBs) clustering linked to upper gastrointestinal cancer (UGC) is scarce. This study focused on rural residents aged 40–69 years in high-risk areas of UGC, and sought to identify clustering patterns among 11 RBs related to UGC and their associations with social determinants of health (SDOH). Using data from the rural UGC screening in Jiangsu Province, latent class analysis was used to identify distinct RB classes, and multinomial logistic regression was applied to assess their associations with SDOH. Among the 45,036 participants, 58.56% had two or more RBs, with overweight/obesity being the most prevalent RB. Four distinct RB latent classes were identified: overall healthy behaviour (55.39%), smoking-alcohol dominant (7.91%), unhealthy diet-dominant (9.37%), and nutrient insufficiency (27.33%) clusters. Men, those with advanced age, and those with lower educational attainment were more likely to belong to the smoking-alcohol dominant, unhealthy diet-dominant, and nutrient insufficiency clusters compared to the overall healthy behaviour cluster. Married individuals and those from households with 4–6 members were less likely to be classified into the smoking-alcohol dominant cluster, whereas individuals from such households were more likely to be in the nutrient insufficiency cluster. Individuals from households with an annual income of Chinese Yuan (CNY) ≥ 60,000 were more likely to be classified into the smoking-alcohol dominant and unhealthy diet-dominant clusters, whereas those with an income of CNY 30,000–59,999 had lower odds of being in the nutrient insufficiency cluster but higher odds of being in the unhealthy diet-dominant cluster. This study indicates that RBs for UGC are highly co-occurring and exhibit distinct clustering within the target screening population. Tailored health intervention strategies based on RB clustering patterns during UGC screening may effectively reduce the UGC burden.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-53727-6.

Introduction

Upper gastrointestinal cancers (UGCs), mainly represented by gastric cancer (GC) and oesophageal cancer (EC), are among the leading malignant tumours worldwide¹. According to the latest Global Cancer Statistics (2022) report, GC and EC account for 4.9 and 2.6% of global cancer cases, and 6.8% and 4.6% of cancer-related deaths, respectively, with East Asia bearing the highest disease burden^1,2. Data from the China National Cancer Centre indicate that in 2022, China had 358,700 new GC cases (5th) and 224,000 new EC cases (7th), with 260,400 GC-related deaths (3rd) and 187,500 EC-related deaths (5th) across all cancer types³. Despite recent declines in incidence and mortality, UGC remains a critical public health challenge^4–6.

Epidemiological evidence has identified multiple risk factors for UGC. For GC, established risk factors include advanced age, male sex, smoking, ethnicity, low socioeconomic status (SES), family history of cancer, inadequate physical activity levels, obesity, low dietary fibre intake, and other unhealthy dietary habits⁵. EC is closely associated with low SES, smoking, alcohol consumption, regular consumption of hot beverages, nitrite exposure, gastroesophageal reflux disease, and obesity⁶. However, most studies have evaluated these factors in isolation^7–14, thus failing to capture their clustering patterns and synergistic/compensatory effects. Risk behaviours (RBs) rarely occur independently. For example, co-occurring smoking and alcohol consumption may cause synergistic mucosal damage exceeding individual effects¹⁵, while protective behaviours (e.g., a high-fibre diet in smokers) may potentially mitigate harm. Both are obscured in single-factor analyses. Therefore, ignoring such clustering may bias UGC risk estimation and hinder targeted prevention.

UGC risk factor clustering refers to the co-occurrence of interrelated factors, which is critical for understanding aetiology. A Netherlands cohort-based study evaluated smoking, body mass index (BMI), physical activity, Mediterranean diet adherence, and alcohol consumption to assess GC/EC subtype risk¹⁶, but lacked coverage of UGC-related factors (e.g., tooth loss, fried/pickled foods intake) and a comprehensive analysis of RB combination patterns. Beyond individual RBs, social determinants of health (SDOH), non-medical factors that influence health outcomes, including income, education, employment, job insecurity, and social security, shape health behaviours^17,18. However, previous studies have focused on individual SDOH indicators (e.g., SES) and upper gastrointestinal disorders^19,20, rather than on how multi-dimensional SDOH interact to shape UGC RB clustering. In fact, SDOH may act as a “macro-level driver” of behavioural clustering: for instance, individuals with low education levels and low household income are more likely to exhibit clustered high-risk behaviours (e.g., smoking and low fruit and vegetable consumption) owing to limited health literacy and restricted access to healthy food options^21–23; conversely, high-income groups may present distinct behavioural clusters (e.g., lower smoking rates alongside high dietary fibre intake)^24,25. Clarifying the relationship between SDOH and RB clustering is crucial for bridging the gap between macro social factors and micro individual behaviours, and for formulating equitable UGC prevention strategies.

A comprehensive understanding of the clustering characteristics of modifiable RBs among individuals aged 40–69 years residing in UGC high-risk areas (recognised as the high-risk population for UGC in China) and their associations with SDOH can yield valuable insights for effective cancer prevention in high-prevalence regions. Latent class analysis (LCA), a person-centred analytical approach, is uniquely suited to this purpose, as it can distinguish unobserved (latent) subgroups of individuals with similar RB profiles within heterogeneous populations²⁶. Therefore, using cross-sectional data from rural UGC screening programmes in Jiangsu Province, this study aims to: (1) analyse the prevalence and latent classes of modifiable RBs for UGC in high-risk populations; and (2) investigate the associations between multi-dimensional SDOH and latent classes of RBs. This study will fill a research gap in UGC risk factor clustering analysis and provide empirical evidence to inform the formulation of stratified, equitable UGC prevention strategies.

Materials and methods

Study design and participants

This study was a secondary analysis based on data from the Rural Early Diagnosis and Early Treatment of UGC Project in Jiangsu Province from January 2017 to August 2021. Briefly, the project invited eligible residents aged 40–69 years from high-prevalence townships or villages (defined by local GC and EC incidence data) to participate voluntarily in UGC screening using a cluster sampling design. Detailed information regarding participant inclusion and exclusion criteria, screening procedures, and standardised data quality control protocols has been fully reported in a previous publication²⁷. The project was specifically designed to systematically collect data on key risk factors for UGC and SDOH in high-risk populations. Its core objectives included identifying RBs associated with UGC in rural areas of Jiangsu Province and facilitating early detection and treatment of UGC. This project was led by the Jiangsu Provincial Centre for Disease Control and Prevention. The inclusion and exclusion criteria for participants remained consistent across all years and regions. All investigators received uniform training to ensure consistent data collection procedures. The project team conducted annual quality control audits to minimise inter-site and inter-year variability in data collection. Whilst the dataset was previously used to explore obesity-related issues²⁷, its standardised data collection framework covered the full spectrum of all UGC-related RBs (e.g., smoking, alcohol consumption, dietary habits) and SDOH indicators (e.g., education, income) required for the present study. Systematic data cleaning and multi-step participant exclusion were conducted for this cross-sectional screening dataset. Individuals from non-high-prevalence regions, those with unsubmitted screening data, inconsistent basic information, ineligible age, and missing key indicators were excluded²⁷. The current study included 45,036 participants aged 40–69 years recruited from UGC high-prevalence rural areas in Jiangsu Province. The original study complied with the Declaration of Helsinki and was approved by the Ethics Committee of Yangzhong People’s Hospital (approval No. KY202402). All participants provided written informed consent prior to enrolment.

Data collection and definitions

This study included relevant variables from the original programme, including: screening year and region, gender (man/woman); age group (stratified into 40–49, 50–59, and 60–69 years); marital status (married and other statuses [unmarried/divorced/separated/widowed]); education level (no formal schooling, primary school, junior high school, and senior high school or above); household size (0–3, 4–6, and ≥ 7 members); annual household income (categorised into three groups via tertile division: ≤29,999, 30,000–59,999, and ≥ 60,000 Chinese Yuan [CNY]); smoking (yes/no) and alcohol consumption (yes/no); self-reported intake frequencies of fresh foods (vegetables, fruits, meat, eggs, milk, soya products) and unhealthy foods (pickled, fried, hot, and mouldy foods). Height and weight were collected via standardised physical examination protocols. All questionnaire items were adopted from the nationally unified screening tool for the national UGC Early Diagnosis and Early Treatment Programme, which has been routinely applied in large-scale population screening across multiple provinces in China. Standardised field survey protocols, unified investigator training, and rigorous quality control were implemented across the entire project to ensure consistent data collection and overall dataset validity. Detailed methods for the collection and measurement of these variables have been reported elsewhere²⁷.

LCA variable definitions

Smoking and alcohol consumption were categorised as “yes” or “no” based on self-reported daily sustained habits, irrespective of the type. Inadequate intake of fresh vegetables, fruits, meat, eggs, milk, and soya products was defined as “yes (intake frequency < 2 times/week)” versus “no (intake frequency ≥ 2 times/week)”. Intake of pickled, fried, hot, and mouldy foods was classified as “yes (intake frequency ≥ 2 times/week)” versus “no (intake frequency < 2 times/week)”. BMI was calculated as weight divided by height squared (kg/m²). According to the Chinese adult standard, BMI ≥ 24 kg/m² was defined as overweight or obese²⁸.

Statistical analysis

All analyses were performed using SPSS (version 27.0) and Mplus (version 8.3). Categorical variables were presented as frequencies and proportions, and group comparisons were assessed using the chi-square test. Continuous variables were reported as mean ± SD, and the two-sample t-test was used to compare groups. The SDOH indicators included gender, age, marital status, education, household size, and annual household income. The modifiable RBs included smoking and alcohol consumption, inadequate intake of fresh vegetables, fruits, meat, eggs, milk, and soya products, intake of pickled, fried, hot, and mouldy foods, and being overweight or obese. Mplus software was used to fit LCA models with 2 to 5 latent classes within a latent mixture model framework. All 11 modifiable RBs were coded as binary variables (1 = present, 0 = absent) for LCA model fitting, consistent with the variable definitions in the LCA Variable Definitions section. Latent class membership probabilities were estimated independently of age, gender, and other factors (i.e., no covariates were included in the class-formation stage) to ensure that latent classes reflected only RB clustering. Models were estimated using robust maximum likelihood (MLR) with numerical integration, the default estimation approach for categorical latent class models in Mplus. To avoid local optima, models were estimated with 20 initial random starts and 4 final optimisations, the default setting in Mplus 8.3. The best log-likelihood was successfully replicated, ensuring stable model convergence. Mplus TECH11 and TECH14 outputs were specified to yield the Lo-Mendell-Rubin adjusted likelihood ratio test (LMRT) and bootstrapped likelihood ratio test (BLRT) for model comparison. Each participant was assigned to a specific latent class according to the highest posterior class probability derived from the final optimal LCA model. Optimal model selection was based on the fitting indices [Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), adjusted BIC (aBIC), entropy, LMRT, BLRT] and clinical plausibility and interpretability²⁹. The lower the values of the AIC, BIC, and aBIC, the better the model fit. The closer the entropy is to 1, the better the classification performance; an entropy value ≥ 0.8 is generally considered to reflect satisfactory classification quality. A P < 0.05 for LMRT and BLRT indicates that the Kth model is better than the (K-1)th model^29,30. To mitigate potential confounding effects of temporal and geographic heterogeneity, screening year and screening region were included as covariates (alongside SDOH indicators) in the multinomial logistic regression model. This model was used to explore the independent associations between SDOH and latent class membership of modifiable RBs. A two-tailed P < 0.05 was considered statistically significant in all analyses.

Results

Basic information on the participants

Table 1 presents the basic characteristics of participants stratified by gender. The mean age for men and women was 57.47 ± 7.61 and 56.03 ± 7.45 years, respectively. The majority of participants were women (25,181, 55.91%), married (man: 18,793, 94.65%; woman: 23,790, 94.48%), and the majority had four to six household members (man: 10,582, 53.30%; woman: 14,053, 55.81%). More men had junior high school education (8,836, 44.50%), while more women had primary school education (8,931, 35.47%). More men reported an annual household income of at least CNY 60,000 (7,100, 35.76%), while more women had an annual household income between CNY 30,000 and 59,999 (8,851, 35.15%). Among participants, the prevalence rates of RBs were as follows: smoking (19.26%), alcohol consumption (17.25%), inadequate intake of fresh vegetables (3.37%), fruits (36.53%), meat, eggs, milk (27.48%), soya products (30.26%), and consumption of pickled (15.37%), fried (3.00%), hot (5.72%), and mouldy foods (0.28%), with 54.00% being overweight or obese. There were significant differences in the distribution of screening year, screening region, age, education, household size, annual household income, smoking, alcohol consumption, inadequate intake of fresh fruits, intake of fried foods, and overweight or obesity between men and women (all P < 0.05). Of the total population, more than half (58.56%) had ≥ 2 RBs. Among individual RBs, overweight or obesity was the most prevalent, while pickled food intake combined with overweight or obesity was the most common two-RB combination (Supplementary Table 1).

Table 1.

Basic information of the study participants.

Variable	Total (n = 45,036)	Man (n = 19,855)	Woman (n = 25,181)	P-value
Screening year				< 0.001
2017	4,242 (9.42)	1,944 (9.79)	2,298 (9.13)
2018	12,658 (28.11)	5,605 (28.23)	7,053 (28.01)
2019	9,965 (22.13)	4,311 (21.71)	5,654 (22.45)
2020	10,616 (23.57)	4,547 (22.90)	6,069 (24.10)
2021	7,555 (16.78)	3,448 (17.37)	4,107 (16.31)
Screening region (city)				< 0.001
Huai’an	17,869 (39.68)	7,813 (39.35)	10,056 (39.93)
Yancheng	12,687 (28.17)	5,696 (28.69)	6,991 (27.76)
Taizhou	8,392 (18.63)	3,776 (19.02)	4,616 (18.33)
Yangzhou	4,473 (9.93)	1,829 (9.21)	2,644 (10.50)
Zhenjiang	1,615 (3.59)	741 (3.73)	874 (3.47)
Age (years)	56.67 ± 7.56	57.47 ± 7.61	56.03 ± 7.45	< 0.001
Age group (years)				< 0.001
40–49	8,846 (19.64)	3,511 (17.68)	5,335 (21.19)
50–59	17,799 (39.52)	7,152 (36.02)	10,647 (42.28)
60–69	18,391 (40.84)	9,192 (46.30)	9,199 (36.53)
Marital status				0.416
Never married/divorced/separated/widowed	2,453 (5.45)	1,062 (5.35)	1,391 (5.52)
Married	42,583 (94.55)	18,793 (94.65)	23,790 (94.48)
Education				< 0.001
No formal schooling	8,287 (18.40)	1,396 (7.03)	6,891 (27.37)
Primary school	14,922 (33.13)	5,991 (30.17)	8,931 (35.47)
Junior high school	16,273 (36.13)	8,836 (44.50)	7,437 (29.53)
Senior high school or above	5,554 (12.33)	3,632 (18.29)	1,922 (7.63)
Household size				< 0.001
0–3	18,180 (40.37)	8,297 (41.79)	9,883 (39.25)
4–6	24,635 (54.70)	10,582 (53.30)	14,053 (55.81)
≥ 7	2,221 (4.93)	976 (4.92)	1,245 (4.94)
Annual household income (CNY)				< 0.001
≤ 29,999	13,953 (30.98)	5,918 (29.81)	8,035 (31.91)
30,000–59,999	15,688 (34.83)	6,837 (34.43)	8,851 (35.15)
≥ 60,000	15,395 (34.18)	7,100 (35.76)	8,295 (32.94)
Smoking				< 0.001
No	36,360 (80.74)	12,022 (60.55)	24,338 (96.65)
Yes	8,676 (19.26)	7,833 (39.45)	843 (3.35)
Alcohol consumption				< 0.001
No	37,268 (82.75)	12,726 (64.09)	24,542 (97.46)
Yes	7,768 (17.25)	7,129 (35.91)	639 (2.54)
Inadequate fresh vegetables				0.556
No	43,519 (96.63)	19,175 (96.58)	24,344 (96.68)
Yes	1,517 (3.37)	680 (3.42)	837 (3.32)
Inadequate fresh fruits				< 0.001
No	28,586 (63.47)	11,976 (60.32)	16,610 (65.96)
Yes	16,450 (36.53)	7,879 (39.68)	8,571 (34.04)
Inadequate meat, eggs, and milk				0.561
No	32,662 (72.52)	14,427 (72.66)	18,235 (72.42)
Yes	12,374 (27.48)	5,428 (27.34)	6,946 (27.58)
Inadequate soya products				0.195
No	31,410 (69.74)	13,785 (69.43)	17,625 (69.99)
Yes	13,626 (30.26)	6,070 (30.57)	7,556 (30.01)
Pickled foods				0.515
No	38,115 (84.63)	16,779 (84.51)	21,336 (84.73)
Yes	6,921 (15.37)	3,076 (15.49)	3,845 (15.27)
Fried foods				0.005
No	43,685 (97.00)	19,209 (96.75)	24,476 (97.20)
Yes	1,351 (3.00)	646 (3.25)	705 (2.80)
Hot foods				0.189
No	42,459 (94.28)	18,751 (94.44)	23,708 (94.15)
Yes	2,577 (5.72)	1,104 (5.56)	1,473 (5.85)
Mouldy foods				0.546
No	44,912 (99.72)	19,797 (99.71)	25,115 (99.74)
Yes	124 (0.28)	58 (0.29)	66 (0.26)
Overweight or obesity				< 0.001
No	20,718 (46.00)	8,829 (44.47)	11,889 (47.21)
Yes	24,318 (54.00)	11,026 (55.53)	13,292 (52.79)

Data are expressed as mean ± SD or n (%).

CNY Chinese Yuan.

An independent-samples t-test was used to compare continuous variables, and a chi-square test was used to compare categorical variables between groups.

Identification of latent classes

Table 2 demonstrates the latent class models for the 11 RBs based on LCA. Overall, AIC, BIC, and aBIC decreased progressively with increasing number of classes, and both LMRT and BLRT P-values were < 0.001. Entropy values were higher in the two- and four-class models. When a five-class model was fitted, the separation between classes one and two was unclear with respect to the probabilities of smoking and alcohol consumption. Similarly, no clear distinction was observed between class two and class five across several RBs, including inadequate intake of fresh fruits, meat, eggs, milk, and soya products (Supplementary Fig. 1). This is unfavourable for clinical interpretation, and entropy was lower than that in the two- and four-class models. Therefore, based on the combined judgement of the various indicators, the four latent classes were considered optimal.

Table 2.

Model fit statistics for LCA models with 2–5 classes.

Statistic	2 classes	3 classes	4 classes	5 classes
Number of free parameters	23	35	47	59
Log-likelihood	-184,680.343	-181,236.518	-179,027.788	-178,145.071
AIC	369,406.687	362,543.037	358,149.575	356,408.142
BIC	369,607.137	362,848.069	358,559.190	356,922.340
aBIC	369,534.043	362,736.839	358,409.824	356,734.838
Entropy	0.811	0.689	0.792	0.755
LMRT	< 0.001	< 0.001	< 0.001	< 0.001
BLRT	< 0.001	< 0.001	< 0.001	< 0.001
Class proportions	0.302	0.287	0.079	0.068
	0.698	0.208	0.094	0.086
		0.505	0.554	0.091
			0.273	0.549
				0.206

LCA latent class analysis. AIC Akaike Information Criterion. BIC Bayesian Information Criterion. aBIC Adjusted Bayesian Information Criterion. LMRT Lo-Mendell-Rubin adjusted likelihood ratio test. BLRT Bootstrapped likelihood ratio test.

Based on the characteristics of the RBs, Class one was named the smoking-alcohol dominant cluster (3,563 cases, 7.91%). Nearly all individuals in this class reported smoking (98.80%) and consuming alcohol (98.30%). Additionally, 1.60% reported inadequate fresh vegetables, 41.10% reported inadequate fresh fruits, 19.90% reported inadequate meat, eggs, and milk, 24.30% reported inadequate soya products, 17.60% reported consuming pickled foods, 2.10% reported consuming fried foods, 4.50% reported consuming hot foods, 0.10% reported consuming mouldy foods, and 50.90% reported being overweight or obese (Fig. 1 and Supplementary Table 2).

Fig. 1.

Conditional probabilities of 11 RBs across four latent classes. RBs risk behaviours. Conditional probability represents the prevalence of each risk behaviour within each corresponding latent class.

Class two, the unhealthy diet-dominant cluster (4,219 cases, 9.37%), was characterised by a relatively high intake of pickled (70.80%), fried (23.40%), and hot foods (33.30%). Besides, there were lower or similar prevalence rates of smoking (18.20%), alcohol consumption (22.30%), inadequate fresh vegetables (1.50%), inadequate fresh fruits (39.60%), inadequate meat, eggs, and milk (11.30%), inadequate soya products (16.10%), mouldy foods (2.10%), and overweight or obesity (64.40%) (Fig. 1 and Supplementary Table 2).

Class three (overall healthy behaviour cluster) constituted 55.39% (n = 24,945) of the sample, with a small proportion reporting smoking (9.80%), alcohol consumption (7.00%), inadequate fresh vegetables (0.00%), inadequate fresh fruits (10.70%), inadequate meat, eggs, and milk (2.40%), inadequate soya products (7.90%), intake of pickled (8.60%), fried (0.20%), hot (2.10%), mouldy foods (0.00%), and being overweight or obese (53.60%, similar to other types) (Fig. 1 and Supplementary Table 2).

Class four was named the nutrient insufficiency cluster (12,309 cases, 27.33%). This classification was characterised by the fact that the majority of the participants had inadequate fresh fruits (82.00%), meat, eggs, and milk (82.70%), soya products (79.20%), while smoking (15.60%), alcohol consumption (12.10%), inadequate fresh vegetables (10.80%), intake of pickled (5.00%), fried (0.20%), hot (1.60%), mouldy foods (0.10%), and overweight or obesity (51.40%) were less or similar to other types (Fig. 1 and Supplementary Table 2).

Notably, the prevalence of overweight/obesity was high across all four latent classes (50.90% to 64.40%). This aligns with the epidemiological characteristics of the study population²⁷. Therefore, latent class names focused on differences within clusters of RBs rather than on overweight/obesity, which was common. This approach ensured valid and specific class labels.

Association of latent classes of RBs with SDOH

Table 3 shows the association of the latent classes with SDOH. Multinomial logistic regression analysis, using the overall healthy behaviour cluster as the reference group, showed that women and participants with higher education displayed a lower likelihood of belonging to the smoking-alcohol dominant, unhealthy diet-dominant, and nutrient insufficiency clusters than their counterparts. By contrast, older participants generally exhibited a higher probability of falling into all three RB clusters. Participants with 4–6 household members were less likely to be in the smoking-alcohol dominant cluster than those from 0 to 3 member households (AOR = 0.877, 95%CI 0.808–0.952), but had a higher likelihood of being in the nutrient insufficiency cluster (AOR = 1.068, 95%CI 1.016–1.123). Participants with an annual household income of CNY 30,000–59,999 showed a lower risk of nutrient insufficiency cluster than those with lower household incomes (AOR = 0.823, 95%CI 0.777–0.872), yet a higher likelihood of belonging to the unhealthy diet-dominant cluster (AOR = 1.127, 95%CI 1.036–1.226). Participants with an annual household income of CNY ≥ 60,000 had an elevated likelihood of being in all RB clusters except the nutrient insufficiency cluster. Additionally, married participants demonstrated a lower likelihood of falling into the smoking-alcohol dominant cluster (AOR = 0.850, 95%CI 0.723-1.000).

Table 3.

Associations between latent classes of RBs and SDOH.

Variable	Smoking-alcohol dominant cluster (n = 3,563)			Unhealthy diet-dominant cluster (n = 4,219)			Nutrient insufficiency cluster (n = 12,309)
	AOR	95% CI	P-value	AOR	95% CI	P-value	AOR	95% CI	P-value
Gender
Man	1.000			1.000			1.000
Woman	0.020	(0.016, 0.024)	< 0.001	0.549	(0.509, 0.591)	< 0.001	0.641	(0.610, 0.673)	< 0.001
Age (years)
40–49	1.000			1.000			1.000
50–59	1.356	(1.211, 1.517)	< 0.001	1.259	(1.143, 1.388)	< 0.001	1.119	(1.050, 1.194)	0.001
60–69	1.461	(1.305, 1.637)	< 0.001	1.107	(0.998, 1.227)	0.055	1.113	(1.040, 1.191)	0.002
Marital status
Never married/divorced/separated/widowed	1.000			1.000			1.000
Married	0.850	(0.723, 1.000)	0.049	1.024	(0.885, 1.185)	0.746	0.957	(0.866, 1.057)	0.388
Education
No formal schooling	1.000			1.000			1.000
Primary school	0.717	(0.618, 0.831)	< 0.001	0.569	(0.518, 0.625)	< 0.001	0.833	(0.779, 0.890)	< 0.001
Junior high school	0.679	(0.586, 0.786)	< 0.001	0.410	(0.370, 0.454)	< 0.001	0.464	(0.431, 0.500)	< 0.001
Senior high school or above	0.609	(0.517, 0.717)	< 0.001	0.361	(0.316, 0.412)	< 0.001	0.321	(0.291, 0.355)	< 0.001
Household size
0–3	1.000			1.000			1.000
4–6	0.877	(0.808, 0.952)	0.002	0.936	(0.870, 1.007)	0.078	1.068	(1.016, 1.123)	0.010
≥ 7	0.996	(0.835, 1.187)	0.964	0.999	(0.865, 1.153)	0.984	1.005	(0.897, 1.125)	0.935
Annual household income (CNY)
≤ 29,999	1.000			1.000			1.000
30,000–59,999	1.063	(0.965, 1.171)	0.217	1.127	(1.036, 1.226)	0.006	0.823	(0.777, 0.872)	< 0.001
≥ 60,000	1.255	(1.126, 1.399)	< 0.001	1.713	(1.563, 1.877)	< 0.001	1.020	(0.955, 1.089)	0.560

Reference group: the overall healthy behaviour cluster.

Adjusted for screening year, screening region, gender, age, marital status, education, household size, and annual household income.

RBs risk behaviours. SDOH social determinants of health. AOR adjusted odds ratio. CI confidence interval. CNY Chinese Yuan.

Discussion

The study describes the prevalence and latent classes of 11 UGC-related RBs and explores associations between the identified latent classes and SDOH. The most prevalent RBs were overweight/obesity, inadequate fresh fruit intake, and inadequate soya products intake. LCA yielded four latent classes among all participants, with 58.56% of the sample presenting co-occurring RBs. Men, older individuals, and those with lower educational attainment were more likely to be in the smoking-alcohol dominant, unhealthy diet-dominant, or nutrient insufficiency clusters than the overall healthy behaviour cluster. Married participants were less likely to be in the smoking-alcohol dominant cluster, while household size and annual household income showed heterogeneous effects across latent classes.

Most existing studies on chronic disease-related RBs focus only on smoking, drinking, physical inactivity, obesity, and inadequate fruit/vegetable intake, or on cardiovascular disease risk factor clustering and their association with diseases^31–36. Research focusing on UGC-specific risk factors remains scarce. In this study, overweight/obesity was the most prevalent RB (54.00%), consistent with a regional study in China (52.6%), which, despite sharing a comparable target population with the present study, did not investigate the clustering of multiple RBs³⁷. An earlier study of the same target population in Yangzhong City found that 45.3% of participants were overweight or obese. However, inadequate intake of fresh fruits (94.5%) was the most prevalent RB in that investigation³⁸. These differences may reflect divergent indicator definitions, population characteristics, and exposure levels. By contrast, physical inactivity was the most prevalent RB in a Korean study³⁹. It is well known that there is a significant inverse association between physical activity and obesity, a finding that partially supports the observations of the present study⁴⁰. Notably, high prevalence rates of overweight or obesity were observed across all latent classes identified in the present study. This is a concerning finding, given that obesity is associated with at least 13 distinct cancer types^41,42. Additionally, more than half (58.56%) of participants had ≥ 2 RBs, consistent with previous literature^32,38. Pickled food intake and overweight/obesity were the most common two-RB combinations in this study, while other studies reported different leading pairs, reflecting study design, population characteristics as well as lifestyle variations^31,43.

The present study further categorises 11 RBs based on LCA and explores their association with SDOH, with some new findings. For example, participants with 4–6 household members displayed a lower likelihood of being in the smoking-alcohol dominant cluster than those with 0–3 household members. One study found that the more extended household members they felt close to, the greater the odds of not smoking and the more likely they were to report a healthy BMI⁴⁴. This is mainly related to positive social support among extended household members^38,44. For example, household members can communicate frequently to remind or advise others to improve bad habits or maintain a healthy lifestyle⁴⁵. A family-oriented intervention study has also reported that family meals facilitate family interaction and communication, particularly discussions among household members regarding alcohol consumption⁴⁶. Furthermore, family intactness and cohesion have been confirmed to significantly reduce the likelihood of smoking and alcohol consumption among household members^47,48. This protective effect against RBs is particularly pronounced in rural areas, where traditional family ties are stronger and daily interactions among household members are more frequent⁴⁹. In contrast, participants from 0 to 3 member households (mainly elderly living alone, empty-nesters, or childless couples) lack such household-level supervision. The present study also found that married individuals had a lower likelihood of being classified into the smoking-alcohol dominant cluster, a finding that further corroborates the aforementioned observations.

The present study further identified that a larger household size (4–6 members) constitutes an independent risk factor for the nutrient insufficiency cluster. This association may be attributed to the following factors: first, with household income fixed (a factor adjusted for in the present study), an increase in household size directly reduces the per capita disposable food budget. Constrained by this budget, household food purchasing strategies tend to prioritise staple foods with the lowest cost per calorie (e.g., cereals) over those with higher nutritional cost per unit (e.g., fruits, lean meat, dairy products), thereby being associated with inadequate nutritional intake^50,51. Second, larger households face greater meal preparation complexity, and, coupled with work and labour commitments, this leads to time poverty among household members, reducing the time allocated to meal preparation and compromising dietary quality^52,53. Finally, as the primary breadwinners and caregivers of the household, individuals in larger households may compromise their own dietary quality to ensure the nutritional needs of other household members (e.g., children), ultimately being associated with inadequate nutrition for themselves^54,55. In this study, participants with an annual household income of CNY 30,000–59,999 had a lower likelihood of being classified into the nutrient insufficiency cluster than those with an income of less than CNY 30,000, which may be associated with their greater purchasing power⁵⁶. However, the likelihood of belonging to the smoking-alcohol dominant and unhealthy diet-dominant clusters increased when annual household income exceeded CNY 60,000. This indicates that household economic status is a double-edged sword in contrast to educational attainment and does not necessarily foster the adoption of healthy behaviours, consistent with the findings from other studies^38,57,58. It also implies that the probability of engaging in unhealthy lifestyle and dietary behaviours rises if health literacy does not improve in tandem with increased purchasing power; thus, individuals with high household incomes are potential targets for interventions to address adverse RBs. Furthermore, the present study found that women and higher educational attainment were protective factors against RB clustering, whereas advanced age was a risk factor for this outcome, consistent with previous research findings^31,32.

Given that our study was conducted in rural populations, targeted UGC prevention strategies should be developed based on distinct RB clustering profiles and their relationships with SDOH. Specifically, these strategies should account for the population prevalence, modifiability of RBs, and local intervention feasibility. A stratified, rural‑adapted hierarchical intervention framework is therefore warranted. Notably, being male gender, advanced age, and low educational attainment represent consistent high-risk characteristics across all risk clusters, and these groups should be regarded as the core target population for universal rural prevention and control efforts. Beyond the aforementioned shared high-risk traits, the nutrient insufficiency cluster accounts for up to 27.33% of the study population and is independently associated with large household size and low annual household income. Owing to its strong reversibility and broad population reach, this cluster warrants designation as the highest-priority intervention group. Given the limited access to health information among rural residents, routine simple nutrition counselling and personalised dietary guidance should be integrated into UGC screening workflows. Intervention compliance and coverage may be enhanced through context-appropriate strategies, including village physician outreach, community engagement, and home-based counselling. This approach prioritises addressing gaps in the intake of fresh fruits and vegetables, as well as meat, eggs, milk, and soya products, with the aim of progressively advancing dietary health equity for rural vulnerable populations. The unhealthy diet-dominant cluster was more common among participants with higher annual household incomes, with excessive consumption of pickled, hot, and fried foods as its core risk factors. This pattern aligns with the traditional dietary norms of rural areas³⁸. Targeted health education can be delivered concurrently during screening, tailored to local dietary practices to disseminate evidence-based dietary guidance. In conjunction, grassroots regulatory authorities should be engaged to strengthen food safety oversight and promote healthy cooking practices, thereby enabling seamless integration of screening and targeted intervention and reducing exposure to dietary carcinogens. The smoking-alcohol dominant cluster, while associated with elevated risks among unmarried individuals, those from smaller household sizes and high annual household income groups, is nonetheless strongly shaped by social contextual factors and behavioural inertia. Given the well-established national tobacco control and alcohol moderation policies, this cluster may be effectively addressed through the ongoing rural public health service infrastructure, obviating the need to establish additional targeted intervention strategies. Additionally, this study identified that overweight and obesity were highly prevalent across all latent classes, indicating that weight assessment and healthy lifestyle counselling ought to be incorporated into routine rural screening procedures to establish an integrated “screening–assessment–intervention” service model, with weight management adopted as a universal primary prevention measure for the entire study population. In conclusion, grounded in the local context of rural Jiangsu, and integrating RB clustering patterns alongside their associated SDOH profiles, a comprehensive, stratified, targeted and feasible prevention strategy that focuses on shared high-risk groups while accounting for the socioeconomic characteristics of each subgroup can significantly improve the efficiency and equity of primary prevention for UGC, providing evidence-based practical guidance for precision prevention and control in rural high-incidence regions.

In the present study, LCA identified four distinct UGC-related RB subgroups, confirming that RBs occur in clusters rather than in isolation. Unlike traditional univariate analysis, which overlooks clustering and masks true population risk profiles, LCA stratifies heterogeneous populations precisely, clarifies RB clustering patterns, and enables targeted UGC prevention. This underscores LCA’s methodological advantage and the study’s value in advancing UGC risk research from single-factor to clustered analysis. Despite its strengths, including a large multi-centre sample and standardised measurements that support population representativeness and data reliability, this study has limitations. First, self-reported RB data may introduce recall or reporting bias. Second, unmeasured confounding may exist, as Helicobacter pylori infection, physical activity, and intake of salt and spicy foods were not assessed in this study. Third, recruitment from screening-compliant participants may cause selection bias and limit generalisability. Fourth, excluding RB intensity and duration (e.g., former smoking, occasional drinking) may have reduced classification validity. Fifth, standard multinomial logistic regression treated latent class membership as definitive (ignoring classification uncertainty, which may bias association estimates), and the unconditional LCA did not adjust for age, sex, year, and so on (which may influence cluster identification). Sixth, the cross-sectional design precludes causal inferences about RB-outcome associations. Finally, the rural-only setting may limit the generalisability of findings to urban populations.

In conclusion, modifiable RBs for UGC commonly co-occur among rural residents aged 40–69 years in high-risk areas of Jiangsu Province. Four distinct behavioural clusters were identified and were differentially associated with SDOH. Men, older individuals, and those with lower educational attainment were consistently more likely to belong to high-risk clusters. Being married was associated with lower odds of belonging to the smoking-alcohol dominant cluster. Middle annual household income was associated with a lower risk of the nutrient insufficiency cluster but a higher risk of the unhealthy diet-dominant cluster, while high annual household income was associated with higher odds of the unhealthy diet-dominant and smoking-alcohol dominant clusters. Larger household size was associated with lower odds of the smoking-alcohol dominant cluster but higher odds of the nutrient insufficiency cluster. These findings provide empirical evidence for the development of precision intervention strategies to prevent UGC. Identifying higher-risk subgroups within the overall UGC high-risk population during cancer screening, based on the distribution characteristics of RBs, latent classes, and SDOH, is critical to implementing targeted health interventions and reducing the UGC burden.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization: X.F., J.H.Z., C.D., Z.L.H., P.F.L., and H.Y.T. Data curation: X.F., Z.L.H., and P.F.L. Formal analysis: X.F., J.H.Z., and Z.L.H. Funding acquisition: X.F., J.H.Z., C.D., and Z.L.H. Methodology: X.F., J.H.Z., and H.Y.T. Project administration: X.F., J.H.Z., C.D., and Z.L.H. Visualisation: X.F. Writing–original draught: X.F. All authors discussed the results and agreed on the final manuscript.

Funding

This study was supported by the China Early Gastrointestinal Cancer Physicians Growing Together Program (Grant No. GTCZ-2021-JS-32-0001), the Zhenjiang City Key Research and Development Plan (Grant No. SH2022051), and the General Program of Preventive Medicine Research Project of Jiangsu Province (Grant No. Ym2023031).

Data availability

The analysed data were sourced from a published population-based screening dataset, and the present work was a secondary analysis of the dataset. Raw data cannot be publicly shared due to ethical constraints and safeguards for participant privacy. Reasonable requests for de-identified analytical data may be directed to the corresponding author. All figures and tables are original, generated by the authors from the cleaned dataset using SPSS (version 27.0) and Mplus (version 8.3), with no copyrighted external material included.

Code availability

The standard Mplus syntax for LCA, supporting the findings of this study, has been deposited in the Zenodo repository and is available with a persistent, citable DOI: 10.5281/zenodo.20032517.

Declarations

Competing interests

The authors declare no competing interests.