Data mining-based lung cancer diagnostic models and high-risk warning for occupational group

Abstract

Screening is the key to improving the five-year survival rate of lung cancer patients. We aimed to investigate the risk factors for lung cancer and to achieve screening and early warning of individuals at high risk of lung cancer in the coke oven worker population. According to the inclusion and exclusion criteria, 185 lung cancer patients and 163 normal controls were included. To screen for biomarkers, we established a mouse lung cancer model and a cellular malignant transformation model. Proteomics was utilised to screen for differentially expressed proteins. Data mining techniques were applied to combine candidate molecular markers of lung cancer with traditional tumor markers to construct lung cancer screening models. The population of coke oven workers was screened for individuals at high risk of lung cancer. CDH1, CLEC3B, CLU, sCD146 and VIM were investigated as candidate markers. Among all the screening models, support vector machine and C5.0 models performed better based on eight protein markers and epidemiological information. The models were applied to predict individuals at high risk of lung cancer among coke oven workers, and a total of 13 high-risk individuals were screened. Dysregulated expression of CDH1, CLU, CLEC3B, sCD146 and VIM was associated with lung cancer. Support vector machine and C5.0 models proved to be more effective and could be applied to screen high-risk individuals among coke oven workers.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-46450-9.

Introduction

Lung cancer is the leading cause of morbidity and mortality worldwide^1–4. Although smoking is the primary risk factor for lung cancer, there are potentially preventable occupational exposures that are risk factors for developing the disease. International estimates indicated that approximately 10–30% of lung cancer cases were due to hazardous occupational exposures⁵. Coal tar pitch (CTP) is a major environmental exposure for coke oven workers in the foundry, coking and steel making industries. Studies have confirmed the carcinogenicity of CTP, which can cause occupational lung cancer^6,7. In 1985, CTP was classified as 'carcinogenic to humans’ (Group I) by the International Agency for Research on Cancer, and can cause tumor such as lung cancer^8,9. Long-term or acute high-dose exposure to CTP extracts (CTPE) may cause various adverse effects such as oxidative stress, telomere damage, chromosomal aberrations, and ultimately tumorigenesis^10,11.

Due to a lack of effective screening and early diagnosis methods, about 70% of lung cancer patients are diagnosed at an advanced stage, missing the optimal time for treatment. As a result, the overall five-year survival rate for lung cancer patients is less than 20%¹². The current gold standard for lung cancer diagnosis is histopathological biopsy, but it is not suitable for lung cancer screening due to its high invasiveness and poor compliance¹³. Low-dose computed tomography (LDCT) is an internationally accepted means of screening for lung cancer, but LDCT has the drawbacks of high false-positive rates, radioactive hazards and overdiagnosis, which severely limit its uptake in the population^14,15.

Currently, lung cancer screening remains an urgent challenge, and the key is to identify early molecular markers with high sensitivity and specificity. Circulating tumor DNA, circulating tumor cells and non-coding RNA, as emerging molecules, have been a hot research topic in recent years. In view of the immaturity of their isolation and extraction and detection technologies, there are challenges in translating them to the clinic, and they are not yet suitable for clinical practice¹³. Protein molecules, as the carriers of life, are the bearers and regulators of all life activities, and have attracted much attention in the field of tumor marker research. The most commonly used lung cancer markers in clinical practice are also mainly protein molecules, such as carcino-embryonic antigen (CEA), neuron-specific enolase (NSE), cytokeratin 19 fragment (CYFRA21-1), etc. However, these tumor markers have the drawbacks of low sensitivity, poor specificity, and difficult to achieve early diagnosis. Currently, these lung cancer markers are mainly used for clinical auxiliary diagnosis, and most of them are used in advanced lung cancer, which have certain limitations in the diagnosis of early lung cancer.

In recent years, especially with the rapid development of proteomics and other technologies, the amount of information has increased dramatically in a geometric progression, and it has become an urgent task to extract useful knowledge or rules from the huge amount of information to be used in the diagnosis of lung cancer.

Data mining technology emerges to address this challenge, integrating database management, artificial intelligence, statistics, and pattern recognition into a unified information-processing framework. It holds unique advantages in solving multi-parameter problems caused by large numbers of variables, particularly in discerning complex correlations and hidden patterns within high-dimensional datasets. The medical community is gradually applying this technology for medical diagnosis and statistical analysis to discover the medical diagnostic rules and patterns in them, and to assist doctors in diagnosing diseases¹⁶. Intelligent methods commonly used for classification and prediction include genetic algorithms, artificial neural networks (ANN), C5.0 decision trees and support vector machines (SVM). These techniques have been widely used in the field of lung cancer screening models^11,17,18. The development of screening models can help to avoid over-diagnosis, maximize the use of data information and provide a broad perspective on lung cancer screening^19,20.

This study aims to develop an integrated lung cancer screening model by synergizing data mining algorithms with molecular biomarkers. The model’s clinical utility will be validated through a prospective cohort study targeting coke oven workers—a high-risk population with occupational exposure to carcinogens. The ultimate objectives are to establish an early detection framework for preclinical malignancies and implement risk stratification protocols, thereby mitigating lung cancer incidence and mortality through timely intervention. Our precision prevention strategy bridges population epidemiology with translational biomarkers through three synergistic components: longitudinal cohort monitoring of occupational exposures, murine carcinogenesis models replicating workplace pollutants, and organoid-based toxicity assessments. This cross-scale validation paradigm synergizes machine learning-enhanced biomarker panels with clinical risk predictors, creating an adaptive screening algorithm. Implementation in high-risk industries demonstrates dual benefits: reducing screening-related costs while maintaining high sensitivity for pre-neoplastic detection (Fig. 1).

Fig. 1.

Flowchart of experimental design.

Materials and methods

Study design

Plasma samples from patients diagnosed with lung cancer in the Department of Respiratory Medicine were collected from September 2018 to September 2019. Exclusion and inclusion criteria are shown in Supplementary File. Signed informed consent forms were also obtained from the patients and individuals used in this study. The normal population from the health examination center in June 2019 was selected as the healthy control. Coking workers in the coking workshop of Henan Iron and Steel Group Co., Ltd. were selected as the occupational population in July 2018.

Screening of candidate molecular markers for lung cancer

CTPE was used as toxic substance to induce the malignant transformation of human bronchial epithelial (BEAS-2B) cells (BeNa Culture Collection). BEAS-2B cells were initially exposed to 15.0 µg/ml of CTPE (5 times), and then cultured to passage 40 (Fig. S1). All experiments were performed with mycoplasma-free cells. 180 C57BL/6 mice (Beijing Vital River Laboratory Animal Technology Co., Ltd., No.11400700362613) were randomly divided into normal control, vehicle control and CTPE groups. Mice in the CTPE group were intratracheally instilled 4 times with 1 mg/mice CTPE (once a week). At month 12, the mice underwent weighing and subsequent humane euthanasia via intraperitoneal injection of 1% pentobarbital sodium at a dose of 120 mg/kg. Following euthanasia, a necropsy was performed to access the main abdominal vein for blood collection (Fig. S2). Differentially expressed proteins were screened by the whole protein tandem mass tagging quantitative proteomic technique (Jingjie PTM BioLab Co. Inc.). Differential expression analysis was performed using MaxQuant (v.1.5.2.8) for proteomic quantification and MoMo (V5.0.2) for Motif analysis. The study combined the proteomic results and applied bioinformatics analysis to screen out candidate molecular markers.

Bioinformatics analysis

Screening criteria for differential proteins were P < 0.05 and fold change greater than 1.3 or less than 1/1.3. The biological functions of the differential proteins were investigated by Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG, Kanehisa Laboratories, Japan; www.kegg.jp) analysis²¹. The Protein–protein Interaction (PPI) network was constructed using the STRING database (v.10.5) and the Cytoscape software (version 3.9.1, https://cytoscape.org/). Meanwhile, the top 11 hub proteins were screened by the degree using the CytoHubba plug-in.

Validation of differential protein expression levels in population samples

The baseline data and plasma samples were collected from 185 lung cancer patients (LC), 163 normal controls (NC) and 163 coke oven workers (CO), and the expression levels of relevant differential proteins in plasma were measured by Enzyme linked immunosorbent assay (ELISA).

Construction of models for lung cancer screening

Artificial neural network algorithms are sensitive to the scale of input features, and unnormalized data with large magnitudes can lead to issues such as vanishing, exploding gradients or slow convergence²². To address this, input variables are typically normalized to the range [0, 1] or standardized to have a mean of 0 and standard deviation of 1. In this study, the min–max normalization method was used to normalize the data, with the formula: X_norm = (X − X_min)/(X_max − X_min), where X is the original data, X_norm is the normalized value, and X_min and X_max are the minimum and maximum values of the original data respectively²³. This normalization ensures numerical stability and enhances the training efficiency of neural networks²⁴.

Using the partition node in IBM SPSS Modeler 18.0, samples were divided into training and prediction sets at a 3:1 ratio with a random seed of 11,111,111. This software provides robust functionality for data partitioning and model training in machine learning workflows. A decision tree C5.0 model, an artificial neural network (ANN) model, a support vector machine (SVM) model, and a Fisher discriminant analysis model were constructed. The operation interface is shown in Fig. 2.

Fig. 2.

Flow chart of data mining. Note: The Date.sav node was the input data source; the Type node described the type of variable; the Feature Selection node can filter the variables and rank the importance of the variables; the Partition node can set the partition ratio of the data, and the study divided the samples into training set and testing set at a ratio of 3:1, and the number of random seeds was set to 1,111,111 (For statistical analyses in the model-building process, normality was tested using the Shapiro–Wilk test (n ≤ 50) or Kolmogorov–Smirnov test (n > 50), and homogeneity of variance was assessed via Levene’s test, as described in Section "Statistical methods").

The lung cancer screening models were developed to compare their predictive performance. The area under the receiver operating characteristic (ROC) curve was used to assess the predictive power of the model, and the selected superior model was used to screen individuals at high risk of lung cancer in the coke oven worker population.

Sample size justification

To ensure adequate statistical power and model stability, the sample size was determined based on the Events Per Variable (EPV) principle, a widely accepted heuristic for multivariable prediction model development. According to common methodological guidelines, an EPV ≥ 10 is generally required to avoid overfitting and ensure reliable estimation of regression coefficients.

A total of 348 participants were enrolled, including 185 patients with lung cancer and 163 healthy controls. The cohort was randomly split into training and validation sets at a 3:1 ratio using stratified sampling to maintain the original case–control distribution. The training set included 139 cases and 122 controls (n = 261). Given that 10 predictor variables were included in the model, the EPV in the training set was 13.9, which exceeded the recommended threshold of 10. These results confirm that the sample size was sufficient for developing a stable prediction model with acceptable generalizability.

Statistical methods

SPSS 21.0 software was used for statistical analysis, and Normality of data distribution was assessed using the Kolmogorov–Smirnov test and Shapiro–Wilk test, while homogeneity of variance was evaluated using Levene’s test. For datasets conforming to normal distribution and homogeneous variance, data were expressed as mean ± standard deviation, and statistical comparisons between two groups were performed using the two-independent-samples t-test; for three or more groups, one-way ANOVA was applied. For non-normally distributed data or those with heterogeneous variance, data were expressed as median (interquartile range, M (P₂₅, P₇₅)), and the Mann–Whitney U test or Kruskal–Wallis H test was used for group comparisons, respectively. Comparison between groups for qualitative data was performed by χ² test, with the test level set at α = 0.05.

Institutional review and board statement

This study was conducted in accordance with the code of ethics of the world medical association with the approval by the “Zhengzhou University Life Science Ethics Review Committee” (code: ZZURIB 2021-106). All experiments were performed in accordance with relevant guidelines and regulations. And methods were reported in accordance with ARRIVE guidelines.

Results

Screening of differentially expressed protein

The grey area in Fig. 3 represented the differentially expressed proteins without statistical significance. The red showed the upregulated proteins and the green area was the downregulated proteins. A total of 6185 proteins were identified in the cellular experiments, among which 157 were upregulated and 40 were downregulated (Fig. 3a). In animal experiments, 976 proteins were detected, with 39 differentially expressed (27 upregulated and 12 downregulated) (Fig. 3b). GO and KEGG analyses of the differentially expressed proteins identified enrichment in pathways critical for tumor progression (Figs. S3 and S4). The results of cellular experiments showed that the differentially expressed proteins were involved in biological processes such as regulation of cell migration and movement, exerting biological functions such as calcium binding, and participating in signalling pathways such as MAPK, etc. The results of animal experiments showed that the differentially expressed proteins were involved in biological processes such as regulation of polypeptide secretion, constituting cellular components such as membrane protein complexes, and participating in signalling pathways such as galactose metabolism).

Fig. 3.

Screening candidate protein markers in vivo and in vitro models of lung cancer. (a) Volcano plots of the differentially expressed proteins in cellular experiments. (b) Volcano plots of the differentially expressed proteins in animal experiments. (c) Hub proteins for cellular experiment. (d) Hub proteins for animal experiment.

PPI networks were constructed using the STRING database (Figs. S5 and S6), and hub proteins were identified using the CytoHubba plug-in in Cytoscape by ranking nodes based on their degree centrality (a measure of the number of interactions per protein). The top 11 hub proteins with the highest degree scores were selected from both cellular and animal experiment datasets (Fig. 3c,d). Among these, CDH1, VIM, CLEC3B, and CLU were consistently ranked in the top, indicating their critical roles in protein interaction networks. sCD146 was added based on clinical expertise and previous literature linking its soluble form to NSCLC metastasis^25,26. These five markers were combined with three traditional tumor markers (CEA, CYFRA21-1, NSE) to form an 8-protein panel for subsequent model construction.

Baseline information of the study participants

A total of 511 participants were enrolled in this study and their characteristics were collected (Table 1). Quantitative information not conforming to normal distribution was expressed as M (P₂₅, P₇₅). The qualitative data were expressed by frequencies and corresponding rates. The histological pathological types and Tumor Node and Metastasis (TNM) stages of the included lung cancer patients are shown in Table S1.

Table 1.

Description and comparison of baseline information of the study participants.

Variables	NC (n = 163)	LC (n = 185)	CO (n = 163)	χ2/F	P
Age (years)	70 (76, 81)	63 (55, 69)	44 (41, 46)	370.31	< 0.0001
Sex				62.62	< 0.001
Male	71	134	136
Female	92	51	27
History of smoking	17	78	80	62.14	< 0.001
History of alcohol consumption	13	48	66	102.39	< 0.001

NC = normal controls; LC = lung cancer patients; CO = coke oven workers; Age is presented as median (interquartile range, M (P₂₅, P₇₅)); Male/Female/smoking/alcohol history is presented as frequency.

Comparison of the expression levels of eight proteins in plasma

As shown in Fig. 4, the expression levels of CEA, CYFRA21-1, CLU, VIM and sCD146 were increased (P < 0.05), while the level of CDH1 was decreased in LC compared to NC (P < 0.05). There were no significant differences in the levels of NSE and CLEC3B between LC and NC (P > 0.05). In addition, the expression levels of CDH1 and sCD146 were decreased in CO (P < 0.05), while the levels of CYFRA21-1 and VIM were increased in CO compared to the NC (P < 0.05). There were no significant differences in the levels of CLU, CLEC3B, CEA and NSE between the two groups (P > 0.05).

Fig. 4.

Comparison of plasma expression levels of 8 proteins in LC, CO and NC. Note: *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001; ns, the difference is not statistically significant.

Comparison of the expression levels of 8 proteins in lung cancer patients with different TNM stages

The expression levels of sCD146 and CEA in early-stage (stage I and II) lung cancer patients were lower than those in advanced stage (stage III and IV) (P < 0.05), while the level of CLU in early-stage lung cancer patients was increased compared to advanced stage lung cancer patients (P < 0.05). There were no correlations between the levels of CDH1, CLEC3B, VIM, CYFRA21-1, NSE and the stages of lung cancer (P > 0.05) (Table 2). The levels of CEA, CLU and VIM were higher in patients with early-stage lung cancer than in healthy controls (P < 0.05), while the levels of CDH1 and CLEC3B were decreased in patients with early-stage lung cancer compared with that of healthy controls (P < 0.05) (Table S2). The expression levels of CEA, CLU, CYFRA21-1, sCD146 and VIM were increased in the advanced lung cancer patients compared with healthy controls (P < 0.05), but the level of CDH1 was lower in advanced lung cancer patients than in healthy controls (P < 0.05) (Table S3).

Table 2.

Comparison of plasma expression levels of eight proteins in lung cancer patients with different TNM stages.

Protein name	TNM staging		Z	P
	I and II (n = 42)	III and IV (n = 104)
CDH1 (ng/ml)	68.68 (46.33, 108.78)	98.22 (57.70, 133.74)	− 1.716	0.086
CLU (µg/ml)	259.37 (207.10, 318.27)	218.05 (197.31, 279.68)	− 2.252	0.024
sCD146 (ng/ml)	4.72 (3.36, 6.57)	6.21 (4.27, 8.77)	− 2.395	0.017
VIM (ng/ml)	521.11 (414.88, 625.23)	535.28 (415.76, 613.38)	− 0.238	0.812
CLEC3B (µg/ml)	0.89 (0.50, 1.02)	0.95 (0.69, 1.19)	− 1.478	0.139
CEA (ng/ml)	1.83 (1.38, 3.27)	3.98 (1.51, 16.00)	− 2.041	0.041
CYFRA21-1 (ng/ml)	1.34 (1.18, 2.28)	1.11 (0.86, 1.71)	− 0.778	0.436
NSE (ng/ml)	6.59 (3.99, 10.41)	6.43 (4.60, 10.89)	− 0.904	0.366

Construction and evaluation of data mining models for lung cancer screening

Using lung cancer status (present/absent) as the response variable, model performance was evaluated on the independent test set. In the first-layer subsystem, predictive models for lung cancer diagnosis were developed based on three traditional plasm biomarkers (CEA, CYFRA21-1, NSE). Among these models, the SVM demonstrated superior performance, achieving an AUC of 0.760 on the test set (Table S4 and Fig. 5A). Based on the five candidate molecular markers (CDH1, CLEC3B, CLU, sCD146 and VIM) for lung cancer, models in the second-layer subsystem were developed, of which the C5.0 model constructed with the better performance, with AUC of 0.842 on the test set (Table S5 and Fig. 5B). The models were constructed in the third-layer subsystem based on the candidate molecular markers and traditional tumor markers, among which the C5.0 model had the better prediction efficiency, with an AUC of 0.907 on the test set (Table S6 and Fig. 5C). Combined with eight protein molecules and epidemiological data, the SVM model took on the better effect in the fourth-layer subsystem, with an AUC of 0.949 on the test set (Table S7 and Fig. 5D).

Fig. 5.

ROC curves of the classification results of the prediction set by the four models. (a) Based on 3 traditional lung cancer markers (CEA, CYFRA21-1, NSE). (b) Based on 5 candidate molecular markers (CDH1, CLEC3B, CLU, sCD146 and VIM). (c) Based on 8 protein molecules (CEA, CYFRA21-1, NSE, CDH1, CLEC3B, CLU, sCD146 and VIM). (d) Based on 8 protein molecules and epidemiological information (CEA, CYFRA21-1, NSE, CDH1, CLEC3B, CLU, sCD146, VIM, History of smoking and History of alcohol consumption).

Model comparison and optimization

To screen out the optimal models, the models constructed in the four-layered system were compared. The comparison of the models showed that SVM⁸⁺ [0.949 (0.884–0.983)] and C5.0⁸⁺ [0.937 (0.868–0.976)] models exhibited superior predictive efficiency on the test set (Table 3). Preferred models will be further used to screen of high-risk individuals.

Table 3.

Evaluation of the classification effect of the five models on the prediction set.

Model	Accuracy (%)	Sensitivity (%)	Specificity (%)	Positive predictive value (%)	Negative predictive value (%)	AUC (95% CI)
SVM3	78.35	60.98	91.07	83.33	76.12	0.760 (0.663–0.841)
C5.05	85.57	75.61	92.86	88.57	83.87	0.842 (0.754–0.908)
C5.08	90.72	90.24	91.07	88.09	92.73	0.907 (0.830–0.956)
SVM8+	94.85	95.12	94.64	92.86	96.36	0.949 (0.884–0.983)
C5.08+	93.81	92.68	94.64	92.68	94.64	0.937 (0.868–0.976)

SVM³, Based on 3 traditional lung cancer markers (CEA, CYFRA21-1, NSE); C5.0⁵, Based on 5 candidate molecular markers (CDH1, CLEC3B, CLU, sCD146 and VIM); C5.0⁸, Based on 8 protein molecules (CEA, CYFRA21-1, NSE, CDH1, CLEC3B, CLU, sCD146 and VIM); SVM⁸⁺ and C5.0⁸⁺, Based on 8 protein molecules and epidemiological information (CEA, CYFRA21-1, NSE, CDH1, CLEC3B, CLU, sCD146, VIM, History of smoking and History of alcohol consumption).

Prediction analysis of high-risk individuals for lung cancer in the coke oven workers population

The models (C5.0⁸⁺ and SVM⁸⁺) were applied to predict the high-risk individuals of lung cancer among occupational group, and 13 high-risk individuals were screened. In addition, 11 out of the 13 high-risk individuals had a single job, and only 2 of them had worked in 2 types of jobs (Table 4).

Table 4.

Basic information of 13 high-risk individuals screened from the coke oven workers.

No	Gender	Age (years)	Smoking history	Type of work	Occupational exposure (years)
79	Male	45	25 years, 40/day	Uplift pipefitter	22
230	Male	41	No	Furnace cover worker	22
239	Male	44	No	Coal distributor	23
259	Male	40	15 years, 10/day	Coal distributor	15
263	Male	27	8 years, 12/day	Oilermaker/coal truck driver	7
295	Male	43	24 years, 20/day	Coal distributor	21
329	Male	46	No	Boilermaker/coal distributor	20
375	Female	45	No	Coal unloading truck driver	28
379	Male	52	39 years, 20/day	Coal distributor	21
397	Male	41	14 years, 14/day	Furnace coverman	20
524	Male	41	16 years, 10/day	Cokemaker	20
740	Male	47	22 years, 10/day	Coal loader	28
1007	Male	43	27 years, 20/day	Coke carrier	24

Discussion

To enhance the prognosis of at-risk populations, it is imperative to implement impactful education, public health interventions and prompt screening strategies for the prevention, diagnosis and timely management of occupational populations. Proteomic has played a major role in the identification of biomarkers in human diseases^27,28. Large-scale and systematic analysis of proteins can provide a complete identification and unique characterization of biomarker in human disease research²⁹. Furthermore, the integration of biomarker findings from different lung diseases and multiple sample types can serve as a valuable resource for future clinical validation studies^30,31. In recent decades, diagnostics have mainly focused on proteins (CYFRA21-1, CA125, CA19-9 and NSE), which represent only a small fraction of the total human genome^32,33. However, none of them can currently be used as independent biomarkers for the diagnosis of lung cancer due to their high false positive rate and low sensitivity^33,34. Therefore, the search for biomarkers of lung cancer and the continuous development of lung cancer screening tools are extremely important to reduce lung cancer mortality.

The results showed that the expression of CDHI, CLU, VIM, CLEC3B and sCD146 proteins was also significantly altered at protein levels. A reduction in CDH1 expression was associated with an increase in epithelial-mesenchymal transition (EMT)^35,36, and further induced EMT in lung cancer by increasing cell viability, decreasing apoptosis and losing tight junctions with neighboring cells, which is an important initial step in the process of tumor invasion and metastasis^37,38. Downregulated expression of CLEC3B has been demonstrated in several tumors^39,40. A study has shown that CLEC3B was significantly downregulated in stage IA lung cancer patients and has a high diagnostic accuracy⁴¹. CLU has been shown to be involved in many stages of NSCLC tumorigenesis, including cancer cells survival, apoptosis, tumor angiogenesis and metastasis⁴². Studies have shown that high CD146 expression was associated with lymph node metastasis, poor survival and was an independent prognostic factor in NSCLC patients^43–45. Vimentin (VIM) is used as a classical marker of EMT reprogramming⁴⁶. During EMT, the VIM protein mediates signaling pathways and directes cell migration by forming cell protrusions, reducing cell adhesion and increasing the ability of cells to migrate⁴⁷.

Data mining is a method of data processing that is well suited to extracting the information contained in data, and when applied to medical research has its own advantages over traditional statistical methods¹⁶. The ANN, C5.0 decision trees and SVM to correctly predict the prediction set, which performed better than the prediction accuracy of fisher discriminant analysis. Therefore, the advantages of data mining over traditional statistical methods were also reflected here.

In order to optimize the model, we developed a four-layer diagnostic system for lung cancer. For classification prediction models, the best way to evaluate their prediction results is to compare their classification accuracy and AUC for the prediction set¹⁶. The AUC of five novel protein molecules for lung cancer was 0.842, superior to that of CEA, CYFRA21-1 and NSE. Candidate molecular markers combined with CEA, CYFRA21-1 and NSE showed the AUC of 0.907. It has been shown that combining the traditional biomarkers with one novel marker to construct a lung cancer diagnostic model can significantly improve the AUC and sensitivity⁴⁸. Based on the 8 molecular markers and epidemiological data, the SVM model had the highest performance. Medical data including clinical epidemiological information, laboratory tests and imaging studies, etc. These information resources are very valuable for disease diagnosis and can help doctors make correct diagnoses for patients.

The joint optimal models were used to predict lung cancer among 163 coke oven workers classified as being at high risk of lung cancer, 13 high-risk individuals were identified through systematic screening, and 11 of them have an accumulated occupational exposure duration of more than 20 years. The intuitive data reveals that the proportion of those with an occupational exposure duration of over 20 years among the high-risk individuals reaches 84.6% (11/13), which is significantly higher than their proportion of 77.3% (126/163) among the total 163 study subjects. It was confirmed that the relative risk of lung cancer was found to be much higher in coke oven workers with 15 or more years of exposure than in non-coke oven workers^49,50. In addition to occupational exposure, 9 workers had a long history of smoking. Smoking is a major risk factor for lung cancer⁵¹, and exposure to coke oven emissions and smoking may have a synergistic effect on the development of lung disease. Only two workers changed the type of job, and others had a long history of a single type of job. When the body encounters exogenous environmental toxic effects with endogenous damage, the key regulatory pathway is DNA repair. But when the body is exposed to environmental toxins for a long period of time, the DNA repair capacity of the lung may be disrupted to the extent that it is unable to repair DNA damage caused by environmental toxins, which in turn may lead to the development of lung cancer⁵². In conclusion, the process of lung carcinogenesis is also one in which damage and repair of the organism go hand in hand, so a regular change of job type may be beneficial in repairing oxidative stress damage to the lung and reducing the risk of lung cancer. The study’s results significantly demonstrate that workers with ≥ 20 years of occupational exposure, a smoking history, job-specific duties, and elevated model scores exhibit a substantially higher risk of lung cancer than that predicted by single-factor assessments. This conclusion underscores the pressing need to implement a dual-dimensional "exposure history-molecular profiling" surveillance strategy in occupational health management. Such a strategy would optimize early risk identification among at-risk populations and furnish a scientific foundation for the development of targeted interventions.

Although this study has had some notable successes and significant predictive results in the early diagnosis of lung cancer, it has some limitations. While the study screened for candidate molecular markers associated with lung cancer, it lacked functional and mechanistic studies of related molecules in lung carcinogenesis. In addition, the study collected a limited number of early lung cancer samples, which may reduce the effectiveness of the models in early lung cancer screening. Finally, our predictive models were developed and internally validated using an independent test set from a single-center occupational cohort. The lack of external validation across geographically or demographically different populations restricts the generalizability of our findings. Further large-scale, multi-center studies are needed to externally validate the models and verify their clinical utility.

These screened coke oven workers will be the focus of follow-up visits, and further screening will be recommended if necessary. The early warning results of lung cancer screening in the coke oven worker cohort will form the basis of a follow-up study, and the group will continue to follow these high-risk individuals to further evaluate the effectiveness of the models.

Conclusion

Based on in vivo and in vitro lung cancer models induced by CTPE, the changes in CDHI, CLU, VIM and CLEC3B proteins were demonstrated as early molecular events through observation and verification. In addition, clinical research confirmed that plasma expression levels of CDHI, CLU, VIM, CLEC3B and sCD146 were associated with lung cancer. Using data mining technology, lung cancer screening models were constructed based on the candidate molecular marker, traditional tumor markers and epidemiological information, of which SVM and C5.0 proved to be more effective in predicting the results. The models could be applied to the screening of high-risk individuals among coke oven workers.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

**Conceptualization:** Yaru Chai, Yongjun Wu; **data curation:** Yaru Chai, Shuyin Duan, Yongjun Wu; **formal analysis:** Yaru Chai, Shuyin Duan, Penghui Ji; **funding acquisition:** Yongjun Wu, Sitian He **; investigation:** Yaru Chai, Shuyin Duan; **methodology:** Yaru Chai, Shuyin Duan, Penghui Ji; **project administration:** Yongjun Wu, Lijun Miao, Lihua Ding; **resources:** Lijun Miao, Hong Liu, Sitian He, Yongjun Wu; **supervision:** Lijun Miao, Hong Liu, Lihua Ding, Yongjun Wu, Sitian He; **validation:** Yaru Chai, Shuyin Duan, Penghui Ji, Lihua Ding, Yongjun Wu; **visualization:** Yaru Chai, Shuyin Duan, Penghui Ji; **writing—original draft preparation:** Yaru Chai, Shuyin Duan, Penghui Ji; **writing—review and editing:** Yaru Chai, Shuyin Duan, Huijie Yuan, Ziqi Wang, Lihua Ding, Yongjun Wu, Sitian He. All authors have read and agreed to the published version of the manuscript.

Funding

This study was supported by grants from the National Natural Science Foundation of China (No. 82273685), the Scientific Research Project of Colleges and Universities in Hainan Province (No. Hnky2025ZC-7), the Academic Promotion Program of Hainan Medical University (No. XSTS2025031), the Hainan Province Clinical Medical Center, and the First Affiliated Hospital of Zhengzhou University. Hainan Province Postdoctoral Research Project, Grant No. BSHYJ002, awarded to Sitian He.

Data availability

The data underlying this article are available in the article and in its online supplementary material. The Supplementary Information with this article can be found in the online version.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval and informed consent

Approval of the research protocol by an institutional review board: This study was conducted in accordance with the code of ethics of the world medical association with the approval by the “Zhengzhou University Life Science Ethics Review Committee” (code: ZZURIB 2021-106). Signed informed consent forms were also obtained from the patients and individuals used in this study.