A risk prediction method for Chinese lung cancer high-risk population based on multi-modal feature fusion

Abstract

The application discloses a kind of Chinese lung cancer high-risk population risk prediction method based on multi-modal feature fusion, comprising: step 1, data acquisition;Step 2, data preprocessing;Step 3, multi-source risk feature evaluation and screening: verify and evaluate the prediction efficiency of lung cancer epidemiology risk score, tumor marker and lung cancer related multi-gene risk score, and screen out the optimal single dimension score of each dimension from it;Step 4, multi-modal score construction: according to the smoking status of participants, the population is divided, the optimal single dimension score of each dimension screened out is standardized, then the multi-source features are fused using Logistic regression model and the weighted coefficients are optimized, and the multi-modal lung cancer risk prediction score suitable for smokers and non-smokers is constructed respectively.The present application has superior discrimination performance, can accurately identify high-risk individuals, guide subsequent imaging examination and early intervention, and improve the early diagnosis rate and treatment effect of lung cancer.

Description

Technical Field

[0001] This invention relates to the fields of epidemiological risk assessment, tumor marker detection, and genetic epidemiology, specifically to a risk prediction method for high-risk groups of lung cancer in China based on multimodal feature fusion, for individualized risk assessment and precise stratification of high-risk groups of lung cancer. Background Technology

[0002] Lung cancer is one of the leading causes of cancer-related deaths in my country, resulting in a severe disease burden. Numerous studies have shown that early screening and intervention for lung cancer can significantly reduce the risk of death. Currently, low-dose computed tomography (LDCT) of the chest is widely recognized as the only effective method for lung cancer screening. However, in large-scale screening practices, LDCT faces challenges such as a large screening population, low positive predictive value, and insufficient utilization of screening resources. Therefore, there is an urgent need for precise screening and stratified management of the screening population through scientifically sound risk prediction methods.

[0003] Lung cancer risk prediction scores quantify the probability of developing lung cancer within a certain period by comprehensively considering individual epidemiological characteristics and risk exposure factors. This is an important technical means to optimize screening strategies and improve screening effectiveness. Currently, several lung cancer risk prediction scores have been published internationally, such as Bach, Spitz, LLP, PLCOM2012, LCRAT, and HUNT models. These are mainly constructed based on macro-epidemiological factors such as age, sex, smoking history, and family history, and have been externally validated in some populations. Meanwhile, in recent years, some scores have also incorporated non-smokers into their assessment framework to cover risk factors such as passive smoking, occupational exposure, and chronic lung disease.

[0004] However, existing lung cancer risk prediction scores still rely primarily on traditional epidemiological variables, failing to adequately consider individual differences at the biological level. Currently, no research has effectively integrated traditional epidemiological factors, tumor markers, and genetic information. One study integrated tumor marker scores constructed using CA125, CEA, CYFRA 21-1, and Pro-SFTPB with PLCOM2012, improving the predictive efficacy of lung cancer incidence risk within one year. Meanwhile, polygenic risk scores (PRS) based on genome-wide association studies can quantify an individual's genetic susceptibility, providing a new technical means for lung cancer risk assessment; however, its application in the Chinese population and its integration with other risk factors require further investigation.

[0005] Furthermore, smokers and non-smokers differ significantly in the mechanisms of lung cancer development, risk factors, and genetic background. Using a single scoring system for risk prediction often fails to adequately account for the predictive performance of different subtypes within these populations. Existing research largely focuses on single risk sources or single population types, lacking a comprehensive risk prediction method that systematically integrates macro-epidemiological factors, tumor markers, and genetic risk information in high-risk lung cancer populations in China. Therefore, it is necessary to develop a lung cancer risk prediction method applicable to high-risk lung cancer populations in China, capable of distinguishing between smokers and non-smokers, and incorporating multimodal risk characteristics to improve the efficiency of population selection and the accuracy of risk assessment in lung cancer screening. Summary of the Invention

[0006] The purpose of this invention is to provide a risk prediction method for high-risk groups of lung cancer in China based on multimodal feature fusion. This method integrates multimodal information, including population epidemiological characteristics, tumor marker detection results, and multigene risk scores, aiming to improve the efficiency of early lung cancer screening and provide a scientific basis for personalized management and intervention of high-risk groups.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a risk prediction method for high-risk groups of lung cancer in China based on multimodal feature fusion, which includes the following steps:

[0008] Step 1, Data Collection: Based on the prospective lung cancer screening cohort, collect participants' demographic epidemiological information, tumor marker detection data, and raw sequencing data of single nucleotide polymorphisms (SNPs);

[0009] Step 2, Data Preprocessing: The collected population epidemiological characteristics, tumor marker detection results, and single nucleotide polymorphism detection results are cleaned and standardized.

[0010] Step 3, Multi-source risk feature assessment and screening: The system searches existing lung cancer risk assessment models, and based on the feature data obtained in Step 2, it verifies and evaluates the predictive efficacy of lung cancer epidemiological risk score, tumor markers and lung cancer-related multi-gene risk score, and screens out the best single-dimensional score for each dimension.

[0011] Step 4: Multimodal score construction: The participants are divided into groups based on their smoking status. The optimal single-dimensional scores of each dimension selected in Step 3 are standardized. Then, a Logistic regression model is used to fuse multi-source features and optimize the weighting coefficients to construct multimodal lung cancer risk prediction scores applicable to smokers and non-smokers.

[0012] Furthermore, in step 1, data acquisition includes:

[0013] The data came from an organized screening cohort for lung cancer. The inclusion criteria were: smokers must have smoked for ≥20 pack years, including those who had smoked for ≥20 pack years and had quit smoking for less than 15 years; non-smokers were considered high-risk if they met any of the following conditions: (1) living or working with a non-smoking woman who met the above conditions and was passively exposed to smoke for ≥20 years; (2) having a family history of lung cancer in a first-degree relative; (3) being exposed to occupational carcinogens for ≥1 year; and (4) having a history of chronic obstructive pulmonary disease.

[0014] Furthermore, in step 2, the data preprocessing process includes:

[0015] (1) Based on the epidemiological characteristics of the population: variables were set and structured for participants’ age, gender, height, body mass index (BMI), education level, smoking history, exposure to cooking fumes, history of chronic lung disease, history of cancer, history of diabetes and family history of cancer.

[0016] (2) Regarding tumor marker characteristics: obtain the detection results of carcinoembryonic antigen (CEA), cytokeratin 19 fragment (CYFRA 21-1), squamous cell carcinoma antigen (SCC-Ag) and nerve-specific enolase (NSE) based on chemiluminescent microparticle immunoassay, and perform logarithmic transformation on the detection results to conform to a normal distribution, or perform binarization based on clinical medical thresholds;

[0017] (3) Targeting the genetic characteristics of SNPs: Obtain the genetic information of lung cancer-related SNP sites based on gene chip detection, and perform gene quality control and genotype filling in sequence.

[0018] Furthermore, the quality control includes: removing low-quality samples with a genotyping rate ≤95%, sex inconsistency, or heterozygous inheritance; and removing low-quality SNP loci with a genotyping rate <95%, minor allele frequency (MAF) <0.001, located on sex chromosomes, or exhibiting linkage disequilibrium; subsequently, a two-stage strategy is used for genotyping imputation, and after imputation is completed, SNP loci with MAF <0.01 and imputation quality score <0.3 are further removed.

[0019] Furthermore, in step 3, the multi-source risk feature assessment and screening process specifically includes: systematically searching existing lung cancer risk assessment models, and based on the features processed in step 2, validating and evaluating publicly published and externally validated lung cancer epidemiological risk scores in a cohort, including 15 scores for smokers: including Bach, Spitz, and the LLP series (LLP, LLP... v2 LLP v3 Hoggart, PLCO M2012 Series (including PLCO) all2014), Pittsburgh, LLPi, LCRAT, HUNT, NCC-LC m2021 OWL and LCRS, ratings for non-smokers (6 in total): including PLCO all2014 , TNSF-SQNG, TNSF-NG, NCC-LC m2021 The predictive efficacy of tumor markers was evaluated by calculating the area under the receiver operating characteristic curve (AUC) of OWL and LCRS. The predictive efficacy of each tumor marker was evaluated by calculating the odds ratio (OR) and AUC for each standard deviation increase in the level of CEA, CYFRA21-1, SCC-Ag, and NSE. A total of 10 published lung cancer polygenic risk scores (PRS) (including the PRS constructed by Jia, Shi, Fritsche, Hung, Graff, Dai, Zhang, Wei, and others) were included for application evaluation, and the optimal PRS score was selected by calculating AUC.

[0020] Furthermore, for smokers, the optimal epidemiological risk score is selected from the LCRS score. The calculation method is to set the regression coefficients and score values ​​of each variable according to the preset reference group, and then sum the score results of all variables to obtain the total LCRS epidemiological risk score.

[0021] The optimal tumor marker combination score screening identified three markers—CEA, CYFRA21-1, and SCC-Ag—as having a significant risk association. The combined score calculation formula is: Score TM_smoker = -6.235+ 0.814×log(CEA + 1) +0.613×log(SCC + 1) + 0.832×log(Cyfra_21_1 + 1);

[0022] The optimal polygenic risk score was selected as PRS-33, which comes from the genotype data of each SNP locus processed in step 2. For each SNP locus, a value was assigned according to the chip detection results: wild homozygous type was recorded as 0, heterozygous type as 1, and variant homozygous type as 2. The genotype values ​​of each SNP were substituted into the PRS formula. Finally, the PRS value is equal to the sum of the products of the genotype values ​​of each locus and the corresponding weight coefficients.

[0023] The specific formula for calculating PRS is: PRS-33 = (0.1275 × rs71658797 rating) + (0.1398 × rs17038564 rating) + (0.0770 × rs3769821 rating + 0.0953 × rs2293607 rating) + (0.1222 × rs13314271 rating) + (0.2151 × rs13167280 rating) + (0.2223 × rs7705526 rating) + (0.1398 × rs4975616 rating) + (0.1044 × rs1056503 rating) + (0.1310 × rs2895680 rating) + (0.1484 × rs2517873 rating) + (0.2546) (ratings of rs3817963) + (ratings of rs1853837) + (ratings of rs5879422) + (ratings of rs6920364) + (ratings of rs11780471) + (ratings of rs4236709) + (ratings of rs10429489) + (ratings of rs62560775) + (ratings of rs1333040) + (ratings of rs4573350) + (ratings of rs1663689) + (ratings of rs12415204) + (ratings of rs1501) (r11591710's rating) + (0.1222 ×rs12265047's rating) + (0.0953 ×rs55768116's rating) + (0.0871 ×rs7953330's rating) + (0.4719 ×rs11571833's rating) + (0.1044 ×rs1200399's rating) + (0.0677 ×rs66759488's rating) + (0.0825 ×rs77468143's rating) + (0.2546 ×rs8034191's rating) + (0.1231 ×rs56113850's rating).

[0024] Furthermore, for non-smokers, the optimal epidemiological risk score is the TNSF-NG score, calculated as follows:

[0025] Score TNSF-NG = -20.072998610139 + 0.700644924363569 × Age - 0.00994485982851517 × Age 2+ 0.0000458183876060965×Age 3 − 0.471528052372111×Edu + 0.808595635603686×COPD + 0.740363567547149×FH − 0.214308004932562×HRT + 0.438451934792846×I(BMI<18.5)− 0.267204418144347×I(24≤BMI<27)−0.514908500745167×I(27≤BMI<30)− 0.327189560329112×I(30≤BMI)+0.00634436913270169×Fume + 0.0108447865674736×Fume_heavy

[0026] Where Age is age, Edu is education level, COPD is chronic obstructive pulmonary disease, FH is family history of lung cancer in first-degree relatives, HRT is hormone replacement therapy, Fume is cooking fume index, and Fume_heavy is heavy cooking fume index.

[0027] The optimal tumor marker score screening showed that CEA was significantly correlated, and the score calculation formula is: Score TM_nonsmoker = -5.533+ 1.057×log(CEA + 1);

[0028] The optimal polygenic risk score was selected from PRS-19, which is calculated as a weighted sum of scores from 19 loci: PRS-19 = (0.1020 × rs17038564 score) + (0.1170 × rs2293607 score) + (0.1900 × rs11375254 score) + (0.2650 × rs13167280 score) + (0.1450 × rs401681 score) + (0.1830 × rs2517873 score) + (0.0610 × rs3817963 score) + (0.1400 × rs1853837 score) + (0.0840 × rs5879422 score) + (0.1560 × rs1853837 ... (ratings of rs4236709) + (ratings of rs10429489) + (ratings of rs35201538) + (ratings of rs4573350) + (ratings of rs12265047) + (ratings of rs55768116) + (ratings of rs11610143) + (ratings of rs1200399) + (ratings of rs77468143) + (ratings of rs200595745).

[0029] Furthermore, in step 4, independent multimodal comprehensive scoring models are constructed for smokers and non-smokers respectively. The specific expressions for the multimodal scoring are as follows:

[0030] For smokers, the multimodal composite score is: Score multimodal_smoker = -3.977 + 0.413 × Score LCRS + 0.414×Score TM_smoker + 0.155×PRS-33;

[0031] For non-smokers, the multimodal composite score is: Score multimodal_nonsmoker = -4.440 + 0.146 × Score TNSF-NG + 0.432×Score TM_nonsmoker + 0.389×PRS-19;

[0032] Among them, Score LCRS With Score TNSF-NG The participants' epidemiological risk scores are respectively: Score TM_smoker With Score TM_nonsmokerThese are the combined scores of tumor markers for the corresponding population, and PRS-33 and PRS-19 are the multigene risk scores for lung cancer for the corresponding population.

[0033] Furthermore, after constructing the aforementioned multimodal score, the model's discriminative ability was evaluated using AUC as the primary indicator; the predictive gain of the multimodal score compared to the unimodal score was assessed by calculating the Net Reclassification Index (NRI) and the Integrated Discriminant Improvement Index (IDI); and the clinical net benefit of the multimodal score compared to the unimodal score was further evaluated through decision curve analysis (DCA).

[0034] An application of the risk prediction method described above in lung cancer risk prediction in high-risk populations, through the construction of a score, is used to assist clinical decision-making in actual lung cancer screening populations, to perform accurate risk stratification, and to optimize the early lung cancer screening process.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] The scoring system of this invention fully integrates validated demographic epidemiological characteristics, tumor marker features, and genetic information, demonstrating the advantages of its multimodal integration strategy. Using this scoring system for systematic screening of high-risk populations in China can more accurately identify potentially high-risk individuals, thereby effectively guiding subsequent imaging examinations and early intervention, significantly improving the early diagnosis rate and treatment outcomes of lung cancer. Specifically:

[0037] (1) Comprehensive multidimensional risk assessment: This invention overcomes the limitations of existing lung cancer risk prediction scores in terms of risk dimensions by comprehensively incorporating validated traditional epidemiological factors, tumor marker test results, and individual genetic susceptibility information. This method can comprehensively reflect the overall risk characteristics of lung cancer from multiple levels and perspectives, significantly improving the accuracy and discriminative ability of risk prediction;

[0038] (2) Differentiated risk assessment strategy: This invention adopts a differentiated assessment method to model the significant differences between smokers and non-smokers in terms of risk factor composition, disease pathogenesis, and genetic background. Through customized methods, the scoring is ensured to have good predictive performance in different population subtypes, thereby enhancing its applicability to high-risk populations in China;

[0039] (3) Comprehensive risk prediction score designed specifically for the Chinese population: The comprehensive risk prediction score provided by this invention is designed specifically for high-risk groups of lung cancer in China. Compared with many existing scores derived from non-Chinese populations, its predictive performance and applicability in the Chinese population are effectively improved. By systematically integrating multi-source risk factors based on Chinese characteristics, the practicality and effectiveness of the score are ensured.

[0040] In summary, this study focuses on lung cancer risk prediction in high-risk populations in China, based on a multimodal feature fusion approach. By validating and systematically integrating population epidemiological characteristics, tumor marker detection results, and multigene risk scores, the study achieved a comprehensive assessment of lung cancer risk. First, various risk scores were evaluated to select the optimal score. Then, standardization was performed to ensure data consistency and reliability. Subsequently, logistic regression was used to effectively fuse information from different data sources, constructing a lung cancer risk prediction score applicable to the Chinese population. This score integrates various validated risk scores, providing a scientific basis for accurately identifying high-risk individuals for lung cancer and developing personalized intervention strategies. The research findings not only overcome the shortcomings of existing scores in terms of risk dimensions and applicability but also provide key technical support for promoting the intelligent and precise assessment of lung cancer risk in my country, possessing significant clinical application value and public health significance. Attached Figure Description

[0041] Figure 1 This is a flowchart of a risk prediction method for high-risk groups of lung cancer according to an embodiment of the present invention.

[0042] Figure 2 It is the discriminative power of the lung cancer epidemiological risk score in people with different smoking statuses.

[0043] Figure 3 The figures show the five-part differential detection rate of the optimal lung cancer risk prediction score (LCRS) for smokers within one year, the optimal lung cancer risk prediction score (TNSF) for non-smokers, the optimal tumor marker combination score for smokers, the optimal tumor marker combination score for non-smokers, the optimal polygenic risk score (PRS-33) for smokers, and the optimal polygenic risk score (PRS-19) for non-smokers.

[0044] Figure 4 The analysis includes receiver operating characteristic (ROC) curves and decision curves for the comprehensive scoring system. Figure A shows the ROC curves for smokers comparing their multimodal comprehensive scores with their unimodal scores. Figure B shows the ROC curves for non-smokers. Figure C shows the decision curve analysis for smokers comparing their multimodal comprehensive scores with their unimodal scores. Figure D shows the decision curve analysis for non-smokers.

[0045] Figure 5 The figures show the five-category detection rate of the multimodal comprehensive score within one year. Figure A shows the five-category detection rate of the multimodal comprehensive score for smokers within one year, and Figure B shows the multimodal comprehensive score for non-smokers. Detailed Implementation

[0046] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0047] Figure 1 This application illustrates an embodiment of a risk prediction method for high-risk groups of lung cancer in China based on multimodal feature fusion. This method integrates population epidemiological characteristics, tumor marker detection results, and multigene risk scores based on a multimodal data fusion strategy. Its aim is to accurately identify potentially high-risk individuals, thereby guiding subsequent imaging examinations and early intervention, and improving the early diagnosis rate and treatment efficacy of lung cancer. Specifically, it includes the following:

[0048] I. Data Collection

[0049] The data for this invention comes from a lung cancer organizational screening cohort, which included 26,862 participants. Subject selection was based on the "Lung Cancer Screening and Early Diagnosis and Treatment Program (2024 Edition)" and national major public health service project standards, with strict inclusion and exclusion procedures developed in accordance with international NLST and NELSON studies. The specific criteria are as follows: Inclusion criteria: Residents aged 50-74 years with full civil capacity who meet any of the following conditions: (1) Smoking amount ≥ 20 pack-years (i.e., number of packs smoked per day × number of years of smoking ≥ 20), including current smokers or former smokers who quit smoking less than 15 years ago; (2) Passive smoking: Non-smoking women who live or work in the same room with smokers who meet the above smoking conditions, and whose passive smoking duration is ≥ 20 years; (3) Family history: Family history of lung cancer in first-degree relatives (parents, children and siblings); (4) Occupational exposure: History of exposure to occupational carcinogens ≥ 1 year, including asbestos, radon, beryllium, chromium, cadmium, silicon, soot and soot; (5) Past medical history: Previously diagnosed with chronic obstructive pulmonary disease.

[0050] II. Data Preprocessing

[0051] 1. Based on population epidemiological characteristics: Variable settings and structured processing were performed for participants’ age, gender, height, body mass index (BMI), education level, smoking history, exposure to cooking fumes, history of chronic lung disease, history of cancer, history of diabetes, and family history of cancer, and the parameters required for each lung cancer epidemiological risk score were defined.

[0052] 2. Tumor Marker Characteristics: Four specific markers were selected for detection, including: CEA, which is highly correlated with lung adenocarcinoma; CYFRA21-1, the preferred marker for lung squamous cell carcinoma screening; SCC-Ag, a specific marker; and NSE, a key marker for small cell lung cancer diagnosis. Quantitative analysis was performed using chemiluminescent microparticle immunoassay. CEA, CYFRA21-1, and SCC-Ag were detected using the Abbott ARCHITECT i2000 platform, while NSE was detected using the New Industries MAGLUMI 4000 Plus platform. The results were logarithmically transformed to conform to a normal distribution, or binarized based on positive values ​​of CEA ≥ 5.0 ng / mL, CYFRA21-1 ≥ 3.3 ng / mL, NSE ≥ 17.0 ng / mL, and SCC-Ag ≥ 1.5 ng / mL.

[0053] 3. Regarding the genetic characteristics of SNPs: Genetic information of lung cancer-related SNP loci based on gene chip detection was obtained, and gene quality control and genotyping were performed sequentially. The quality control included: sample quality control: (1) removing samples with a genotyping rate ≤95%, (2) removing samples with inconsistent sex, (3) removing samples with heterozygous inheritance, etc.; SNP quality control: (1) removing loci with a genotyping rate <95%, (2) removing loci with a minor allele frequency (MAF) <0.001, (3) removing loci on sex chromosomes, (4) removing loci with linkage disequilibrium, etc. Subsequently, a two-stage strategy was adopted for genotyping. The reference panel was composed of the third-stage reference panel of the 1000 Genomes Project (N=2,504) and the Nanjing Medical University Ommics Database (N=3,020). The genotyping process was completed using SHAPEIT4 and IMPUTE2. A total of 24,470,273 sites were filled, and SNPs with MAF < 0.01 and filling quality (INFO) < 0.3 were further excluded.

[0054] III. Verification and Evaluation of Multi-Source Risk Characteristics

[0055] A lung cancer epidemiological risk score was constructed based on demographic, behavioral, and medical history factors, including age, sex, height, body mass index, education level, smoking history, chronic lung disease, and family history of cancer. The score was constructed by substituting relevant parameters into a formula retrieved from the system and then weighting and summing them to obtain an individual's lung cancer risk score. Seventeen publicly published and externally validated scoring models were included for AUC and detection rate assessment, among which 15 models were for smokers: including Bach, Spitz, and the LLP series (LLP, LLP...). v2 LLP v3 Hoggart, PLCOM2012 Series (including PLCO) all2014 ), Pittsburgh, LLPi, LCRAT, HUNT, NCC-LC m2021 OWL and LCRS, including six ratings for non-smokers: PLCO all2014 The evaluation included TNSF-SQNG, TNSF-NG, NCC-LCm2021, OWL, and LCRS. The results showed that the optimal model for smokers was LCRS (AUC=0.663, 95% CI: 0.636-0.690), and the optimal model for non-smokers was TNSF-NG (AUC=0.585, 95% CI: 0.544-0.626). Figure 2 When the top 60% of the population based on risk scores were identified as high-risk screening subjects, the LCRS score in the smoking group detected 78.62% (114 / 145) of lung cancer cases, while the TNSF-NG score in the non-smoking group detected 68.75% (77 / 112) of lung cancer cases. Figure 3 A, B). For four common lung cancer-related tumor markers, the association between each standard deviation increase in expression level and the risk of lung cancer after binarization was assessed (Table 1). Analysis showed that in smokers, the expression levels of CEA, CYFRA21-1, and SCC-Ag were significantly positively correlated with lung cancer risk, with ORs of 2.03 (1.70-2.43), 1.61 (1.30-1.99), and 1.48 (1.18-1.83), respectively; while in non-smokers, only CEA showed a significant correlation (OR=1.81, 95% CI: 1.39-2.36). Based on the above significantly correlated markers, a combined score was constructed. Using the top 60% of the population as the screening cutoff, the tumor marker score detected 73.79% (109 / 145) of lung cancer cases in the smoking group and 69.64% (78 / 112) of lung cancer cases in the non-smoking group. Figure 3(C, D) A multi-gene risk score was constructed based on SNP loci associated with lung cancer risk, and a weighted summation formula was used for calculation. Locus genotypes were assigned values ​​as follows: wild-type homozygous = 0, heterozygous = 1, and variant homozygous = 2. A total of 10 lung cancer PRSs were included for application evaluation: including PRSs constructed by Jia, Shi, Fritsche, Hung, Graff, Dai, Zhang, Wei, and others. After AUC optimization (Table 2), the optimal score for smokers was PRS-33 (AUC = 0.546, 95% CI: 0.514-0.578), with specific SNP locus information shown in Table 3; the optimal score for non-smokers was PRS-19 (AUC = 0.627, 95% CI: 0.585-0.669), with specific SNP locus information shown in Table 4. When the top 60% of the population in terms of risk score were selected for screening, the PRS score could detect 60.69% (88 / 145) of lung cancer cases in the smoking group, and 80.36% (90 / 112) of lung cancer cases in the non-smoking group. Figure 3 E, F).

[0056] The calculation method for the LCRS epidemiological risk score for smokers is as follows: regression coefficients and score values ​​are assigned to each variable according to a preset reference group, and the score results of all variables are summed to obtain the total LCRS epidemiological risk score. The specific risk score assignment is shown in Table 5. The calculation formula for the TNSF-NG epidemiological risk score for non-smokers is as follows:

[0057] Score TNSF-NG = -20.072998610139 + 0.700644924363569 × Age - 0.00994485982851517 × Age 2 + 0.0000458183876060965×Age 3 − 0.471528052372111×Edu + 0.808595635603686×COPD + 0.740363567547149×FH − 0.214308004932562×HRT + 0.438451934792846×I(BMI<18.5)− 0.267204418144347×I(24≤BMI<27)−0.514908500745167×I(27≤BMI<30)− 0.327189560329112×I(30≤BMI)+0.00634436913270169×Fume + 0.0108447865674736×Fume_heavy

[0058] Where Age is age, Edu is education level, COPD is chronic obstructive pulmonary disease, FH is family history of lung cancer in first-degree relatives, HRT is hormone replacement therapy, Fume is the cooking fume index, and Fume_heavy is the heavy cooking fume index.

[0059] The formula for calculating tumor marker scores in smokers is: Score TM_smoker = -6.235 + 0.814 × log(CEA + 1) + 0.613 × log(SCC + 1) + 0.832 × log(Cyfra_21_1 + 1); The formula for calculating the tumor marker score in non-smokers is: Score TM_nonsmoker = -5.533+ 1.057×log(CEA + 1).

[0060] The formula for calculating the polygenic risk score for smokers is: PRS-33 = (0.1275 × rs71658797 score) + (0.1398 × rs17038564 score) + (0.0770 × rs3769821 score + 0.0953 × rs2293607 score) + (0.1222 × rs13314271 score) + (0.2151 × rs13167280 score) + (0.2223 × rs7705526 score) + (0.1398 × rs4975616 score) + (0.1044 × rs1056503 score) + (0.1310 × rs2895680 score) + (0.1484 × rs71658797 score) + (0.1310 × rs2895680 score) + (0.1484 × rs71658797 score) + (0.1398 ... (rr2517873 rating) + (0.2546 ×rs3817963 rating) + (0.1133 ×rs1853837 rating) + (0.0770 ×rs5879422 rating) + (0.0677 ×rs6920364 rating) + (0.1415 ×rs11780471 rating) + (0.1133 ×rs4236709 rating) + (0.1044 ×rs10429489 rating) + (0.1007 ×rs62560775 rating) + (0.0935 ×rs1333040 rating) + (0.1222 ×rs4573350 rating) + (0.1275 ×rs1663689 rating) + (0.0862) (ratings of rs12415204) + (ratings of rs11591710) + (ratings of rs12265047) + (ratings of rs55768116) + (ratings of rs7953330) + (ratings of rs11571833) + (ratings of rs1200399) + (ratings of rs66759488) + (ratings of rs77468143) + (ratings of rs8034191) + (ratings of rs56113850).The formula for calculating the polygenic risk score for non-smokers is: PRS-19 = (0.1020 × rs17038564 score) + (0.1170 × rs2293607 score) + (0.1900 × rs11375254 score) + (0.2650 × rs13167280 score) + (0.1450 × rs401681 score) + (0.1830 × rs2517873 score) + (0.0610 × rs3817963 score) + (0.1400 × rs1853837 score) + (0.0840 × rs5879422 score) + (0.1560 × rs4236709 score) + (0.1000 × rs1853837 score) + (0.1000 × rs1853837 score) + (0.1000 × rs1853837 score) + (0.1000 × rs18538564 ... (ratings of rs10429489) + (ratings of rs35201538) + (ratings of rs4573350) + (ratings of rs12265047) + (ratings of rs55768116) + (ratings of rs11610143) + (ratings of rs1200399) + (ratings of rs77468143) + (ratings of rs200595745).

[0061] Table 1. Association and discriminative power of single tumor markers with lung cancer incidence.

[0062]

[0063] Table 2. Association and discriminative power of polygenic genetic risk scores with lung cancer incidence.

[0064]

[0065] Table 3. Construction of PRS-33 related sites

[0066]

[0067] Table 4. Construction of PRS-19 related sites

[0068]

[0069] Table 5. LCRS Risk Score Assignment Table

[0070]

[0071]

[0072] IV. Construction of Multimodal Scoring

[0073] The optimal single-dimensional scores for each selected dimension were standardized, and then weighted using the coefficients output by Logistic regression to construct a multimodal score. In the smoking population, the lung cancer epidemiological risk score had the highest discriminative power, followed by the tumor marker score, and lastly the multigene risk score. In the non-smoking population, the multigene risk score had the highest discriminative power, followed by the lung cancer epidemiological risk score, and lastly the tumor marker CEA. Independent multimodal comprehensive scoring models were constructed for both smoking and non-smoking populations. The specific expression for the multimodal score is as follows: For the smoking population, the multimodal comprehensive score is: Score multimodal_smoker = -3.977 + 0.413 × Score LCRS + 0.414×Score TM_smoker + 0.155×PRS-33; For non-smokers, the multimodal composite score is: Score multimodal_nonsmoker = -4.440 + 0.146 × Score TNSF-NG + 0.432×Score TM_nonsmoker + 0.389×PRS-19.

[0074] Among them, Score LCRS With Score TNSF-NG The participants' epidemiological risk scores are respectively: Score TM_smoker With Score TM_nonsmoker These are the combined scores of tumor markers for the corresponding population, and PRS-33 and PRS-19 are the multigene risk scores for lung cancer for the corresponding population.

[0075] For smokers, the constructed multimodal score had an AUC of 0.690 (95% CI: 0.660–0.720), demonstrating good discriminative power; compared to the single-dimensional score, its net reclassification index (NRI) was 35.08%, and its integrated discrimination improvement index (IDI) was 1.11%. For non-smokers, the integrated score had an AUC of 0.662 (95% CI: 0.620–0.704), demonstrating moderate discriminative power; compared to the single best epidemiological score TNSF-NG (AUC = 0.586), the performance improvement of the integrated score was statistically significant, with an NRI of 39.00% and an IDI of 0.86% (Table 6). Decision curve analysis (DCA) results further confirmed the clinical applicability of the integrated score. For smokers, the integrated score curve was above other independent score curves in most threshold probability ranges, resulting in higher net clinical benefits. For non-smokers, DCA results showed that within a broad threshold probability range of 0.02 to 0.20, the composite score demonstrated superior net benefit compared to individual independent scores, suggesting its high potential application value in assisting clinical intervention decisions. Figure 4 In terms of actual screening efficacy, when the top 60% of the population based on risk scores were defined as high-risk screening subjects, the composite score in the smoking group could detect 82.07% (119 / 145) of lung cancer cases; the composite score in the non-smoking group also showed excellent sensitivity, detecting 82.14% (92 / 112) of lung cancer cases. Figure 5 The internal performance of the comprehensive score was verified by 10-fold cross-validation (Table 7). The results confirmed the robustness of the model: the average AUC of the model for smokers was 0.680, which was close to the fitting result at the time of construction, and there was no significant difference in AUC between layers, indicating that the model was robust; the average AUC of the model for non-smokers was 0.632, which was slightly lower than the result at the time of construction (0.662), but there was no statistically significant difference in AUC between layers, and the overall predictive performance remained stable.

[0076] Table 6 Performance improvement of the discrimination of the comprehensive lung cancer risk prediction model

[0077]

[0078] Table 7. Ten-fold cross-validation analysis of the overall score

[0079]

[0080] The specific applications of the above multimodal scoring in lung cancer screening are as follows:

[0081] This scoring system, used for systematic screening of high-risk populations in China, combines epidemiological risk, tumor marker levels, and multigene risk scores to more accurately identify potentially high-risk individuals. This guides subsequent imaging examinations and early intervention, improving the early diagnosis rate and treatment outcomes of lung cancer. Furthermore, this method allows for tailored follow-up plans based on individual risk, optimizing medical resource allocation and reducing false positive rates and unnecessary examinations.

[0082] This multimodal comprehensive score improves the accuracy of lung cancer risk prediction by integrating the three types of features mentioned above. In the dataset, the comprehensive score demonstrates superior discriminative performance and shows promising prospects for widespread application. It can more accurately identify potentially high-risk individuals, thereby effectively guiding subsequent imaging examinations and early intervention, and improving the early diagnosis rate and treatment outcomes of lung cancer.

[0083] The above descriptions are specific embodiments of the present invention. It should be understood that the above descriptions of the various embodiments of the present invention are intended to illustrate the technical principles of the present invention and are not intended to limit the present invention. For those skilled in the art, various equivalent substitutions, modifications, or improvements made to the present invention without departing from the spirit and essence of the present invention should be covered within the protection scope of the present invention. The terminology used herein is for illustrative purposes only and should not be construed as limiting the present invention.

Claims

1. A risk prediction method for high-risk groups of lung cancer in China based on multimodal feature fusion, characterized in that, Includes the following steps: Step 1, Data Collection: Based on a prospective lung cancer screening cohort, collect participants' demographic epidemiological information, tumor marker detection data, and raw single nucleotide polymorphism sequencing data; Step 2, Data Preprocessing: The collected population epidemiological characteristics, tumor marker detection results, and single nucleotide polymorphism detection results are cleaned and standardized. Step 3, Multi-source risk feature assessment and screening: The system searches existing lung cancer risk assessment models, and based on the feature data obtained in Step 2, it verifies and evaluates the predictive efficacy of lung cancer epidemiological risk score, tumor markers and lung cancer-related multi-gene risk score, and screens out the best single-dimensional score for each dimension. Step 4: Multimodal score construction: The participants are divided into groups based on their smoking status. The optimal single-dimensional scores of each dimension selected in Step 3 are standardized. Then, a Logistic regression model is used to fuse multi-source features and optimize the weighting coefficients to construct multimodal lung cancer risk prediction scores applicable to smokers and non-smokers.

2. The method for risk prediction of high-risk groups for lung cancer in China based on multimodal feature fusion according to claim 1, characterized in that, In step 1, data collection includes: The data came from an organized screening cohort for lung cancer. The inclusion criteria were: smokers must have smoked for ≥20 pack years, including those who had smoked for ≥20 pack years and had quit smoking for less than 15 years; non-smokers were considered high-risk if they met any of the following conditions: (1) living or working with a non-smoking woman who met the above conditions and was passively exposed to smoke for ≥20 years; (2) having a family history of lung cancer in a first-degree relative; (3) being exposed to occupational carcinogens for ≥1 year; and (4) having a history of chronic obstructive pulmonary disease.

3. The method for risk prediction of high-risk groups for lung cancer in China based on multimodal feature fusion according to claim 1, characterized in that, In step 2, the data preprocessing process includes: (1) Based on the epidemiological characteristics of the population: variables were set and structured for participants’ age, gender, height, body mass index, education level, smoking history, exposure to cooking fumes, history of chronic lung disease, history of cancer, history of diabetes and family history of cancer; (2) Regarding the characteristics of tumor markers: obtain the detection results of carcinoembryonic antigen, cytokeratin 19 fragment, squamous cell carcinoma antigen and nerve-specific enolase based on chemiluminescent microparticle immunoassay, and perform logarithmic transformation on the detection results to conform to a normal distribution, or perform binarization based on clinical medical thresholds; (3) Targeting the genetic characteristics of SNPs: Obtain the genetic information of lung cancer-related SNP sites based on gene chip detection, and perform gene quality control and genotype filling in sequence.

4. The method for risk prediction of high-risk groups for lung cancer in China based on multimodal feature fusion according to claim 3, characterized in that, The quality control includes: removing low-quality samples with a genotyping rate ≤95%, sex inconsistency, or heterozygous inheritance; and removing low-quality SNP loci with a genotyping rate <95%, minor allele frequency <0.001, located on sex chromosomes, or exhibiting linkage disequilibrium; subsequently, a two-stage strategy is used for genotyping imputation, and after imputation is completed, SNP loci with MAF <0.01 and imputation quality score <0.3 are further removed.

5. The method for risk prediction of high-risk groups for lung cancer in China based on multimodal feature fusion according to claim 1, characterized in that, In step 3, the multi-source risk feature assessment and screening process specifically includes: systematically searching existing lung cancer risk assessment models; based on the features processed in step 2, validating and evaluating publicly published and externally validated lung cancer epidemiological risk scores in a cohort; and including scores for smokers such as Bach, Spitz, LLP series, Hoggart, and PLCO. M2012 Series, Pittsburgh, LLPi, LCRAT, HUNT, NCC-LC m2021 OWL and LCRS, ratings for non-smokers: including PLCO all2014 , TNSF-SQNG, TNSF-NG, NCC-LC m2021 The predictive efficacy of tumor markers was evaluated by calculating the area under the receiver operating characteristic curve (ROC) and the LCRS, and the optimal score was selected. The predictive efficacy of each tumor marker was evaluated by calculating the odds ratio and AUC for each standard deviation increase in the level of CEA, CYFRA21-1, SCC-Ag, and NSE. The application evaluation of the published lung cancer polygenic risk score was carried out, and the optimal PRS score was selected by calculating the AUC.

6. The method for risk prediction of high-risk groups for lung cancer in China based on multimodal feature fusion according to claim 5, characterized in that, For smokers, the optimal epidemiological risk score is selected from the LCRS score. The calculation method is to set the regression coefficient and score value of each variable according to the preset reference group, and then sum the score results of all variables to obtain the total LCRS epidemiological risk score. The optimal tumor marker combination score screening identified three markers—CEA, CYFRA21-1, and SCC-Ag—as having a significant risk association. The formula for calculating the combination score is as follows: ； The optimal polygenic risk score was selected as PRS-33, which comes from the genotype data of each SNP locus processed in step 2. For each SNP locus, a value was assigned according to the chip detection results: wild homozygous type was recorded as 0, heterozygous type as 1, and variant homozygous type as 2. The genotype values ​​of each SNP were substituted into the PRS formula. Finally, the PRS value is equal to the sum of the products of the genotype values ​​of each locus and the corresponding weight coefficients. The specific formula for calculating PRS is as follows: 。 7. The method for risk prediction of high-risk groups for lung cancer in China based on multimodal feature fusion according to claim 5, characterized in that, For non-smokers, the optimal epidemiological risk score is the TNSF-NG score, and its calculation formula is as follows: Where Age is age, Edu is education level, COPD is chronic obstructive pulmonary disease, FH is family history of lung cancer in first-degree relatives, HRT is hormone replacement therapy, Fume is cooking fume index, and Fume_heavy is heavy cooking fume index. The optimal tumor marker score screening showed that CEA was significantly correlated, and the score calculation formula is as follows: ； The optimal polygenic risk score was selected from PRS-19, which is calculated as a weighted sum of scores from 19 loci: 。 8. The method for risk prediction of high-risk groups for lung cancer in China based on multimodal feature fusion according to claim 1, characterized in that, In step 4, independent multimodal comprehensive scoring models are constructed for smokers and non-smokers respectively. The specific expressions for the multimodal scoring are as follows: For smokers, the multimodal composite score is: ； For non-smokers, the multimodal composite score is: ； Among them, Score LCRS With Score TNSF-NG The participants' epidemiological risk scores are respectively: Score TM_smoker With Score TM_nonsmoker These are the combined scores of tumor markers for the corresponding population, and PRS-33 and PRS-19 are the multigene risk scores for lung cancer for the corresponding population.

9. The method for risk prediction of high-risk groups for lung cancer in China based on multimodal feature fusion according to claim 8, characterized in that, After constructing the above multimodal score, the discriminative ability of the model was evaluated using AUC as the main indicator; the predictive gain of the multimodal score compared with the single-modal score was assessed by calculating the net reclassification index and the comprehensive discriminative improvement index; and the clinical net benefit of the multimodal score compared with the single-modal score was further evaluated by decision curve analysis.

10. An application of the risk prediction method as described in any one of claims 1 to 9 in predicting the risk of lung cancer in high-risk populations, wherein the constructed score is used to assist clinical decision-making in actual lung cancer screening populations, to perform accurate risk stratification, and to optimize the early lung cancer screening process.

Images