A chronic obstructive pulmonary disease risk assessment device

Abstract

The application discloses a chronic obstructive pulmonary disease risk assessment device and belongs to the technical field of biomedical detection and risk assessment. The device can accurately predict the COPD risk in smokers and non-smokers and identify'susceptible' and 'tolerant' individuals under the same exposure level through a group of compact proteomic biomarkers. In addition, the device can also integrate comorbidity risk assessment and post-cessation vulnerability transition window assessment functions, thereby providing support for individualized follow-up and precise prevention.

Description

Technical Field

[0001] This invention relates to the field of biomedical detection and risk assessment technology, and in particular to a risk assessment device for chronic obstructive pulmonary disease. Background Technology

[0002] COPD is a major cause of increased morbidity and mortality worldwide, carrying a significant clinical and economic burden. Current diagnostic methods primarily rely on evidence of persistent airflow limitation in pulmonary function tests, typically only confirming the diagnosis after symptoms have appeared and irreversible damage has occurred, leading to missed diagnoses and delayed diagnosis. Particularly in some regions, insufficient pulmonary function testing after administering high-quality bronchodilators further limits early identification capabilities.

[0003] Smoking is one of the most significant modifiable risk factors for COPD, but the relationship between smoking and COPD is not entirely explicit; that is, not all smokers will develop COPD. Existing risk assessment methods are mostly based on exposure information such as smoking status, duration of smoking, or number of packs per year, which makes it difficult to distinguish between "susceptible" and "tolerant" individuals at the same exposure level, and also fails to reflect the molecular residual effects and short-term risk fluctuations after smoking cessation.

[0004] In recent years, high-throughput proteomics technologies have enabled the simultaneous detection of thousands of circulating proteins, making it possible to construct multi-protein biomarker panels for preclinical risk identification. However, there is still a lack of multi-protein risk assessment protocols that can be validated in large-scale populations and have clearly defined judgment rules, especially an integrated approach that takes into account COPD risk, smoking heterogeneity stratification, and risk window identification after smoking cessation.

[0005] Therefore, there is an urgent need for a risk assessment method that can be implemented in peripheral blood samples, has clear calculation and judgment steps, and can support individualized follow-up and intervention stratification. Summary of the Invention

[0006] The present invention aims to at least partially solve one of the technical problems existing in the prior art. To this end, the present invention provides a chronic obstructive pulmonary disease risk assessment device.

[0007] According to one aspect of the present invention, a chronic obstructive pulmonary disease (COPD) risk assessment device is provided, comprising: a data receiving module for receiving detection result data of a target protein in an in vitro plasma sample of a subject; a preprocessing module for preprocessing the detection result data, the preprocessing including quality control, missing value handling, and standardization; a risk score calculation module for inputting the preprocessed protein expression data into a preset risk assessment model; a risk stratification module for comparing the COPD risk score with at least one preset threshold and outputting a risk stratification result; and an output module for outputting the risk stratification result and / or the corresponding risk assessment report.

[0008] Preferably, the target protein consists of 11 proteins: ALPP, CXCL17, WFDC2, MMP12, GDF15, SCGB1A1, TNR, BCAN, REN, EGFR, and AGER. The chronic obstructive pulmonary disease (COPD) risk score (Score) is calculated using the following linear combination model: Score = + + + + + + + + + + Wherein, ALPP to AGER are the standardized expression values ​​of each target protein. to These are the preset model coefficients.

[0009] Preferably, the risk assessment model satisfies the proportional risk assumption, and the p-value of the interaction between protein risk score and time is greater than 0.05 according to the Schoenfeld residual test; and the variance inflation factor (VIF) of each predictor variable in the model is less than 5.

[0010] Preferably, the system further includes a molecular stratification module for smokers, which is used to: perform covariate correction on protein expression data of current smokers to obtain residual protein expression features; perform dimensionality reduction and clustering on the residual protein expression features to output susceptible or tolerant molecular subgroup labels; wherein, the susceptible subgroup is characterized by enrichment of pro-inflammatory signaling pathways and a significantly higher cumulative incidence of COPD than the tolerant subgroup.

[0011] Preferably, the covariate correction includes using a linear regression model to remove the influence of age and smoking duration on protein expression values; the dimensionality reduction uses principal component analysis (PCA) and / or unified manifold approximation and projection (UMAP); the clustering uses the k-means algorithm, and the number of clusters is determined to be 2 through silhouette coefficient analysis.

[0012] Preferably, the system also includes a vulnerability transition window assessment module for smoking cessation, which is used to: receive baseline variables of the subject, including at least one of age, body mass index (BMI), smoking duration, years since quitting smoking, and social deprivation index (TDI); invoke a preset protein trajectory prediction model to generate the expression trajectory of the target protein for the next 15 years under both a continuous smoking scenario and an immediate quitting smoking scenario; calculate the time-series risk value under the two scenarios based on a risk model; and within a preset short-term assessment window of 0-2 years, when at least n time points meet the RR (Risk-Responsiveness) threshold... quit (t)>RR continue When (t) + δ, a vulnerable transition window after smoking cessation is determined and an early warning message is output; where RR quit (t) represents the predicted risk at time point t under the scenario of immediate smoking cessation, RR continue (t) represents the predicted risk at time point t under the continuous smoking scenario, and δ is the preset risk difference threshold.

[0013] Preferably, the protein trajectory prediction model is a multilayer perceptron (MLP) regression model or a random forest regression model, and the risk model is a multivariate Cox proportional hazards model.

[0014] Preferably, the system further includes a comorbidity risk assessment module, which is used to: calculate a lung cancer risk score based on the expression level of CXCL17 in the subject's in vitro plasma sample; and / or calculate an emphysema risk score based on the standardized expression values ​​of four proteins, CXCL17, WFDC2, IL22 and TNR, according to a linear combination model; and output the corresponding comorbidity risk stratification results.

[0015] According to another aspect of the present invention, a method for assessing the risk of chronic obstructive pulmonary disease (COPD) is provided, comprising the following steps: receiving detection result data of target proteins in an in vitro plasma sample of a subject, wherein the target proteins are composed of 11 proteins including ALPP, CXCL17, WFDC2, MMP12, GDF15, SCGB1A1, TNR, BCAN, REN, EGFR, and AGER; preprocessing the detection result data, wherein the preprocessing includes quality control, missing value handling, and standardization; inputting the preprocessed protein expression data into a preset risk assessment model, and calculating a COPD risk score according to the linear combination model described above; comparing the COPD risk score with at least one preset threshold, and outputting a risk stratification result.

[0016] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0017] According to this invention, a compact set of proteomic biomarkers has been identified that can predict COPD risk in both smokers and non-smokers, and reveals a dual molecular profile of "susceptibility-tolerance" among smokers at the same exposure level. This molecular stratification provides a basis for early risk identification and precise prevention of smoking-related COPD. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0019] Figure 1 This is a schematic diagram illustrating the relationship between smoking intensity gradient and immune-metabolic axis imbalance according to embodiments of the present invention. (A) A volcano plot of smoking-related differentially expressed proteins in human samples. Significantly upregulated proteins are represented by red dots, and significantly downregulated proteins by blue dots. (B) A volcano plot of smoking-related differentially expressed proteins in mouse samples. (C) A scatter plot showing the consistency of smoking-related expression changes across species. Each point represents a comparison of the log2 fold change of a protein in humans with its corresponding change in a mouse model. The best-fit line with a 95% confidence interval is shown in the figure. The Pearson correlation coefficient r = 0.742 (p = 0.004). (D) Box plots of expression levels of selected proteins in humans and mice under smoking and non-smoking conditions. The upper panel shows the expression in current smokers (blue) and never-smokers (red); the lower panel shows the expression in smoke-exposed mice (blue) and control mice (red). Compared with the control, there was a statistically significant difference in expression levels in the smoking group (two-tailed t-test, p < 0.05). Statistical significance is indicated by an asterisk: ; ; ; (E) Gene Ontology (GO) enrichment bar chart of upregulated proteins in heavy smokers relative to non-smokers. Bar length represents the fold enrichment of each GO entry, and bar color represents significance (expressed as –log10 of the corrected p-value). (F) Heatmap of selected protein expression profiles from four smoking exposure groups (never, low, moderate, and heavy). Each row represents a protein (labeled with gene name), and each column represents a group. Rows are hierarchically clustered by expression pattern to highlight heterogeneous expression changes at different smoking exposure levels.

[0020] Figure 2This is a schematic diagram illustrating the identification of COPD-related protein biomarkers and the proteomic heterogeneity analysis in smokers according to embodiments of the present invention. (A) A scatter plot showing the adjusted relative risk (RR) of smokers compared to non-smokers in different human disease categories. Colored by disease system: respiratory system (green), digestive system (blue), malignant tumors (purple), and circulatory system (red). All associations shown in the figure are statistically significant (FDR-adjusted p < 0.0001). (B) Flowchart of the analysis from protein biomarker screening to population stratification. (C) A forest plot of COPD risk factor hazard ratios (HRs) obtained based on a multivariate Cox proportional hazards model, showing 95% confidence intervals (CIs) and p-values. A total of 15 features are displayed, including baseline variables (BMI, sex, age) and protein biomarkers (such as CXCL17, MMP12, EGFR). Points represent HR estimates, and horizontal lines represent 95% confidence intervals. (D) Receiver operating characteristic (ROC) curves comparing two COPD risk prediction models: a risk scoring model based on 11 protein characteristics (blue curve, AUC = 0.844) and a model based on smoking category (yellow curve, AUC = 0.773). (E) Violin plots of 11 protein risk score distribution in four smoking exposure groups (never smokers, low, moderate, and heavy smokers). The statistically determined optimal risk score cutoff (0.89) is indicated by a vertical dashed line. The superimposed kernel density curves and internal box plots represent the median (center line), interquartile range (box), and data range (whiskers), respectively. (F) Visualization of the residualized 11 protein expression profile in two-dimensional space using the Unified Manifold Approximation and Projection (UMAP). Subsequently, k-means clustering was applied to identify two distinct clusters (Cluster 1: n = 1,163, red; Cluster 2: n = 1,365, blue). The attached UMAP chart shows participants who developed COPD within 5, 10, or 15 years of follow-up (red dots; the number of COPD cases in each time interval is marked in the corresponding graph). (G) Kaplan–Meier curves comparing the cumulative COPD incidence rates of Cluster 1 and Cluster 2. The COPD incidence rate over time was significantly higher in Cluster 1 than in Cluster 2 (log-rank test). (H) Violin plot comparing the age and smoking duration distribution of participants in Cluster 1 and Cluster 2. No significant differences were observed between the two clusters in terms of age or smoking duration (p = ns, two-tailed t-test). (I) Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis of differentially expressed proteins between the two clusters. Bar plots show the upregulated (red) and downregulated (blue) pathways in Cluster 1 relative to Cluster 2.

[0021] Figure 3 This is a schematic diagram of a cross-sectional proteomic model related to smoking cessation status provided by an embodiment of the present invention. Specifically: (A) A line graph shows the average expression levels of prioritized protein biomarkers in four smoking groups: current smokers, smokers who quit for <10 years, smokers who quit for ≥10 years, and never smokers. (B) A box plot shows 408 differentially expressed proteins screened by LASSO regression, grouped into six clusters. The average expression levels of each cluster in different populations are: current smokers (red), smokers who quit for <10 years (blue), smokers who quit for ≥10 years (yellow), and never smokers (green). (C) A bubble chart of Gene Ontology (GO) biological process enrichment analysis. The intensity of the color indicates significance (FDR-corrected p-value), with darker colors (deeper red) indicating higher enrichment significance. (D) A violin plot shows the distribution of predicted risk scores at 3-year intervals after smoking cessation, with current smokers as the reference group. The left panel represents moderate smokers (smoking history 15–30 years), and the right panel represents heavy smokers (>30 years). (E) GSEA enrichment curves for selected inflammation and immune-related pathways. Each graph shows the running enrichment score for the corresponding pathway, with the normalized enrichment score (NES) and FDR-corrected p-value labeled. The core enriched proteins contributing the most to each pathway are listed below the corresponding graph. (F) Box plots compare the expression levels of representative protein biomarkers among current smokers, former smokers, and never smokers. Statistical significance is indicated by an asterisk (…). ; ; ; ).

[0022] Figure 4This is a schematic diagram illustrating a potential transitional window in heavy smokers after quitting smoking, based on proteomics-based MLP modeling according to an embodiment of the present invention. (A) A flowchart of the analysis from mechanism to risk, including cohort setup, model design, stochastic simulations of two scenarios (continuous smoking vs. quitting), and output steps quantifying the risk changes caused by quitting. (B) Simulated relative risk trajectories of light, moderate, and heavy smokers predicted by a multilayer perceptron (MLP) model trained on cross-sectional biomarker data. Solid lines represent continuous smoking, and dashed lines represent quitting; heavy (red), moderate (blue), and light (green). (C) Simulated percentage changes in disease risk relative to continuous smoking after quitting for light, moderate, and heavy smokers. (D) A violin plot (nested box plot) showing the distribution of the percentage change in risk after quitting: light (blue), moderate (yellow), and heavy (red). A horizontal dashed line represents 0% (no change). The upper half of the violin plot represents individuals with decreasing risk, and the lower half represents individuals with increasing risk. Box plot elements represent the median (center line), interquartile range (box), and overall range (whiskers). (E) The proportion of individuals with increased and decreased COPD risk predicted by the model at years 0, 1, and 2 after quitting smoking (group colors are shown in the legend). (F) Violin plots compare baseline characteristics of different subgroups. Left: BMI and age distribution of individuals predicted to experience a rebound and decrease in risk upon quitting smoking. Right: BMI and age distribution of current smokers and individuals who quit smoking 1 year ago. Box plot elements represent the median (center line), interquartile range (box), and overall range (whiskers) (two-tailed t-test). (G) Volcano plots show differential protein expression between individuals classified as high-risk and low-risk based on model predictions. Significantly upregulated genes are indicated by red dots, and significantly downregulated genes are indicated by blue dots. (H) Scatter plots compare changes in biomarkers in heavy and moderate smokers during the early stages of quitting smoking. Each point represents the log2 fold change of a single biomarker from baseline (current smoking) to early quitting (0–2 years) in heavy smokers (x-axis) and moderate smokers (y-axis). The diagonal line indicates that the magnitude of change is equal in both groups. The left figure shows NMR-based metabolic biomarkers, and the right figure shows blood biomarkers; some representative biomarkers are labeled for illustrative purposes.

[0023] Figure 5This is a schematic diagram illustrating how a multi-protein biomarker panel, provided according to embodiments of the present invention, can predict the risk of lung cancer and emphysema in COPD and non-COPD cohorts. Specifically: (A) A forest plot shows the hazard ratios (HRs) and 95% confidence intervals (CIs) for variables associated with lung cancer progression in COPD patients, including CXCL17, sex, body mass index (BMI), years since smoking cessation, and other variables. CXCL17 is significantly associated with lung cancer progression (HR = 3.21, 95% CI: 1.49–6.91; p = 0.003). (B) Kaplan–Meier curves show the cumulative incidence of lung cancer in COPD patients with high CXCL17 expression (red) and low expression (blue). The incidence is significantly higher in the high expression group (log-rank p < 0.0001). (C) A forest plot shows the HRs (95% CIs) for predictors of emphysema progression in COPD patients, including CXCL17, WFDC2, IL-22, TNR, and other variables. WFDC2, IL-22, and TNR were significantly associated with emphysema progression. (D) Cumulative incidence of emphysema in COPD patients stratified by the combined expression levels of the four key proteins. Patients with higher combined expression levels had a significantly higher incidence of emphysema progression (log-rank p < 0.0001). (E) Cumulative incidence of lung cancer stratified by CXCL17 expression levels in non-COPD individuals. Patients were divided into quartiles based on expression levels from low to high. (F) Cumulative incidence of emphysema stratified by the combined expression levels of the four proteins in non-COPD individuals. Patients were divided into quartiles based on expression levels from low to high. Detailed Implementation

[0024] While the invention can be embodied in many different forms, what is disclosed herein are specific illustrative embodiments that demonstrate the principles of the invention. It should be emphasized that the invention is not limited to the specific embodiments illustrated herein. Furthermore, any section headings used herein are for organizational purposes only and are not to be construed as limiting the subject matter.

[0025] To avoid ambiguity, the relevant terms in this invention are defined as follows:

[0026] Subject / Individual: refers to a human individual who undergoes testing and / or risk assessment using the methods of this invention. Adults are preferred.

[0027] Current smoker: refers to an individual who was still smoking at baseline.

[0028] Former smokers: refers to individuals who have stopped smoking at baseline.

[0029] Never smokers: Individuals with no clear smoking history at baseline or cumulative smoking exposure below a preset threshold.

[0030] Light / moderate / severe smokers: Stratified according to cumulative smoking duration, pack-years, or other smoking exposure metrics. In a preferred embodiment, classified by smoking duration as: light (<15 years), moderate (15–30 years), severe (>30 years).

[0031] COPD event: Preferably defined according to the ICD-10 code J44 (other chronic obstructive pulmonary diseases) in hospital diagnosis records.

[0032] Emphysema event: Preferably defined according to the ICD-10 code J43.

[0033] Lung cancer event: Preferably defined according to the ICD-10 code C34.

[0034] Protein expression value: Can be the output value of a high-throughput protein detection platform (such as the NPX value of the Olink platform), or an absolute or relative concentration value converted by a standard curve. Different detection platforms can be mapped through a calibration model.

[0035] Normalized expression value : Represents the value of the i-th subject and the j-th target protein after pretreatment and normalization. Preferably, z-score normalization is used.

[0036] / COPD comprehensive protein risk score: A risk score calculated based on the expression values of multiple target proteins according to preset model parameters, used to evaluate the risk of COPD.

[0037] The present invention provides a method for risk assessment of respiratory diseases based on plasma proteomics, at least including the following steps: Step S101: Collect peripheral blood samples from subjects and separate plasma; Step S102: Detect the expression levels of at least one group of target proteins; Step S103: Perform quality control, missing value processing, and normalization on the detection data; Step S104: Input the normalized protein expression values into a preset risk assessment model to calculate the COPD risk score; Step S105: Output the COPD risk stratification result according to the threshold; Optional step S106: Perform "susceptible–tolerant" molecular stratification on smokers; Optional step S107: Evaluate the vulnerable transition window for individuals who have quit smoking; Optional step S108: Output the comorbidity risk score of lung cancer and / or emphysema. The above steps can be executed by manual operation, a semi-automatic system, or a computer program, or can be integrated into a software system supporting a kit.

[0038] Correspondingly, the present invention provides a risk assessment device for respiratory diseases based on plasma proteomics. The device at least includes: a data receiving module, a data preprocessing module, a risk score calculation module, a risk stratification module, and an output module. The data receiving module is used to receive the detection result data of at least one group of target proteins in the plasma sample of the subject in vitro; the data preprocessing module is used to perform quality control, missing value processing, and standardization on the detection result data; the risk score calculation module is used to input the standardized protein expression value into a preset risk assessment model to calculate the COPD risk score; the risk stratification module is used to output the COPD risk stratification result according to a threshold; the output module is used to output the risk stratification result and / or the risk assessment report. Optionally, the device further includes a molecular stratification module for performing "susceptible - tolerant" molecular stratification on smokers; optionally, the device further includes a vulnerable transition window assessment module for assessing the vulnerable transition window of quitters; optionally, the device further includes a comorbidity risk assessment module for outputting the comorbidity risk score of lung cancer and / or emphysema. The above modules can be implemented by hardware circuits, software programs, or a combination of both, or can be integrated into a kit supporting software system.

[0039] Example 1: COPD risk assessment method based on an 11-protein panel

[0040] 1. Sample collection and processing

[0041] Collect the peripheral venous blood sample of the subject, preferably using a blood collection tube containing an anticoagulant (EDTA anticoagulant tube). The blood collection volume is 5 mL. Centrifuge the collected blood sample within a preset time to separate the plasma. Perform a second centrifugation if necessary to reduce cell residue.分装 the separated plasma and store it, preferably stored at -80°C to avoid repeated freezing and thawing.

[0042] 2. Detection of target proteins

[0043] In a preferred embodiment, the Olink Explore platform (based on proximity extension assay, PEA) is used to detect plasma proteins. The platform can output the NPX value of the protein. In another embodiment, ELISA, Luminex, MSD, multiplex immunoassay platform, targeted mass spectrometry, etc. can be used to detect the target proteins of the present invention. If an alternative platform is used, the detection values should be mapped to a unified scoring system through calibration samples or conversion models.

[0044] Detect the following 11 target proteins: ALPP, CXCL17, WFDC2, MMP12, GDF15, SCGB1A1, TNR, BCAN, REN, EGFR, and AGER.

[0045] 3. Data Preprocessing

[0046] Quality control is performed on the obtained protein expression matrix. Preferably, samples with a missing rate greater than a preset threshold are deleted; proteins with a missing rate greater than the preset threshold are also deleted. In one embodiment, the missing rate threshold is 30%. Missing values ​​remaining after quality control are imputed. KNN imputation is preferably used, with k being 5. In a preferred embodiment, if model training and validation data are to be separated, the imputator parameters are only fitted to the training set and applied to the validation / test set to avoid information leakage. Protein expression values ​​are standardized, preferably using z-score standardization, calculated as follows:

[0047] ,

[0048] in, This represents the original expression value of the j-th protein in the i-th subject. and denoted as the mean and standard deviation of the j-th protein in the training set, respectively.

[0049] 4. Calculation of COPD Comprehensive Protein Risk Score

[0050] In this embodiment, a linear combination model is used to calculate the COPD comprehensive protein risk score for each subject. It can be in the following form:

[0051]

[0052] in, Let j be the standardized expression value of the target protein for the i-th subject. The corresponding protein coefficients are as follows. Preferably, the order of the above 11 proteins is fixed as follows: ALPP, CXCL17, WFDC2, MMP12, GDF15, SCGB1A1, TNR, BCAN, REN, EGFR, AGER.

[0053] Score = 0.21457393 CXCL17 + 0.73792488 WFDC2 + 0.18887395 GDF15 +0.12830138 ALPP + 0.25248641 MMP12 - 0.47919873 TNR - 0.22955520 BCAN -0.40041569 SCGB1A1 + 0.18104341 REN - 0.75348940 EGFR - 0.48531675 AGER.

[0054] 5. Risk Stratification and Determination

[0055] Will Compare with the preset threshold T: when When T ≥ T, it is considered a high risk of COPD; when When the risk level is <T, the patient is considered to be at low risk for COPD. In some implementations, a medium-risk level, a recommended follow-up examination cycle, and / or a recommended pulmonary function test may be output.

[0056] Example 2: Training method for constructing an 11-protein COPD risk model

[0057] 1. Data Sources and Queue Construction

[0058] In one embodiment, a population cohort containing plasma proteomic data, baseline demographic information, smoking history, and longitudinal health records is used as the training data source. Preferably, UK Biobank proteomic data resources are used. Subjects must have at least the following information: baseline protein expression data; baseline age, sex, BMI, TDI, and other covariates; smoking status and smoking exposure information; and COPD outcome information during follow-up.

[0059] 2. Data partitioning

[0060] The eligible samples are randomly divided into a training set and a validation set (or a training set / validation set / test set). A stratified random partitioning method is preferred to maintain a stable proportion of COPD events, with a partitioning ratio of 7:3. Preferably, the training set is used for feature selection and model building, the validation set is used for threshold determination and performance evaluation, and the test set is used for final performance confirmation.

[0061] 3. Screening of smoking-related candidate proteins

[0062] First, differential protein analysis is performed based on smoking status or smoking intensity stratification to screen for differentially expressed proteins (DEPs) that are significantly associated with smoking. Preferred statistical methods may include those used by limma et al., with a significance threshold of FDR < 0.05. Smoking-related DEPs are then used as the candidate feature set for subsequent COPD risk modeling.

[0063] 4. Stable Feature Selection and Model Building

[0064] In the training set, penalized regression modeling is performed on candidate proteins. LASSO or LASSO-Cox models are preferred. In one implementation, proteins are first ranked according to their association strength in a one-way Cox model; the top N proteins are selected for the LASSO-Cox model (N can be one or more of 200, 500, or 1000 for sensitivity analysis). Cross-validation is used to determine the regularization parameter λ, preferably using the one-standard-error rule. ) or minimum error criterion ( To improve feature stability, repeated resampling fitting was performed (200 times), and the inclusion frequency of each protein in the model was recorded. Proteins with frequencies exceeding a preset threshold (70%) were selected as core features. Ultimately, 11 stable protein features were obtained, and the final scoring model was fitted to obtain coefficients β1 to β11.

[0065] 5. Model Evaluation

[0066] The model performance was evaluated on the validation set (and / or test set), and its COPD discrimination ability was assessed using ROC curves and AUC metrics. The model was compared with control models based solely on smoking status, smoking duration, or traditional clinical variables to demonstrate the technical effectiveness of the multi-protein score of this invention.

[0067] Example 3: Molecular Stratification Method for Smokers' "Susceptibility-Tolerance"

[0068] This embodiment illustrates how, under the same or similar levels of smoking exposure, molecularly susceptible and tolerant smokers can be identified based on protein expression characteristics, thereby supporting more refined risk stratification.

[0069] 1. Target audience screening

[0070] Current smokers were selected, preferably heavy current smokers (smoking duration > 30 years), and those with pre-existing COPD or lung cancer were excluded to reduce the impact of existing diseases on protein expression.

[0071] 2. Residualization (covariate correction)

[0072] Covariate adjustments were made to target protein expression values ​​to reduce the influence of factors such as age and duration of smoking. Linear regression was preferred, fitted separately for each protein.

[0073]

[0074] For each subject i and each protein j, take the residual. This serves as the corrected expression characteristic of the subject on the j-th protein; a residual protein expression matrix is ​​then constructed for subsequent dimensionality reduction and clustering analysis.ij Let be the expression value of the i-th subject on the j-th protein. Through the above residualization process, while preserving inter-individual differences in protein expression, the linear effect of age and smoking duration on protein expression can be reduced, thereby improving the sensitivity of subsequent molecular stratification to biological differences beyond exposure dose.

[0075] 3. Dimensionality reduction

[0076] PCA dimensionality reduction is performed on the residual protein expression matrix, retaining the top principal components whose cumulative explained variance reaches a preset threshold (90%). UMAP is then used to embed the samples into a two-dimensional or three-dimensional space; example parameters include: =30、 =0.1.

[0077] 4. Clustering Hierarchy

[0078] Cluster analysis is performed in a reduced-dimensional space, preferably using k-means clustering. In one implementation, K=2. The number of clusters K can be selected using the silhouette coefficient, Calinski-Harabasz index, or Gap statistic. Cluster labels (Cluster 1, Cluster 2) for each individual are output, and their COPD risk differences are assessed in conjunction with subsequent follow-up events.

[0079] 5. Interpretation and Application of Stratification Results

[0080] In one implementation, Cluster 1, characterized by enriched pro-inflammatory signals and a higher risk of subsequent COPD, is defined as "susceptible"; Cluster 2, characterized by enhanced metabolic / detoxification pathways and a lower risk of subsequent COPD, is defined as "tolerant". The system outputs "molecular subgroup label + risk predisposition + follow-up recommendations". For example, for "susceptible" current heavy smokers, it suggests prioritizing smoking cessation intervention, close monitoring, and further examination.

[0081] Example 4: A method for identifying vulnerable transition windows after smoking cessation

[0082] 1. Obtaining Input Variables

[0083] Baseline information of the participants should be obtained, including at least: age, body mass index (BMI), smoking duration, years since quitting smoking (current smokers are counted as 0), and Townsend Deprivation Index (TDI).

[0084] 2. Target Protein and Prediction Task

[0085] In one implementation, each protein in the 11-protein COPD risk panel is used as the prediction target, that is, a regression model is trained for each protein to predict its expression trajectory on the future time axis.

[0086] 3. Machine Learning Model Building (MLP Preferred)

[0087] A multilayer perceptron (MLP) regression model was employed. For each target protein, the preferred model structure included three hidden layers, with examples of 128, 64, and 32 neurons, and the activation function was ReLU. Preferred training settings included: optimizer: Adam; learning rate: Loss function: Mean Squared Error (MSE); Maximum number of training epochs: 1000; Early stopping strategy: Monitor validation set loss. Input variables are preferably z-score standardized first. If a training / validation set partition is used, the standardization parameters are preferably calculated from the training set and fixed for use in the inference phase.

[0088] 4. Dual-Scenario Counterfactual Simulation

[0089] For the same subject, two counterfactual scenarios are constructed: Scenario A (Continuous Smoking): It is assumed that the subject continues to smoke within the prediction window, and the smoking cessation period remains at 0 or is set according to the continuous smoking rule; Scenario B (Immediate Smoking Cessation): It is assumed that the subject quits smoking immediately from the current point in time, and the smoking cessation period increases with the prediction time. Under each scenario, the predicted trajectory of each target protein is generated within a preset prediction duration (e.g., 15 years) according to a time step (e.g., 1 year).

[0090] 5. Risk Curve Calculation

[0091] The protein trajectories obtained under the two scenarios are input into a preset COPD risk model (e.g., a Cox proportional hazards model or an equivalent risk model) to calculate the risk value or relative risk (RR) at different time points. Preferably, the risk model can also incorporate BMI, TDI, smoking duration, years since quitting smoking, and interaction terms for correction.

[0092] 6. Fragile Transition Window and "Rebounder" Identification

[0093] To identify individuals whose risk may not monotonically decrease in the short term after quitting smoking, this invention defines a "vulnerable transition window" and uses it to determine "rebounders." The vulnerable transition window refers to a situation where, within a pre-defined short-term assessment window after quitting smoking, the subject's predicted risk under an immediate quitting scenario is higher than the predicted risk under a continuing smoking scenario at multiple time points, exceeding a pre-defined risk difference threshold. This suggests that the subject may experience a brief period of increased risk before achieving long-term benefits from quitting. This determination is used for individualized follow-up stratification and monitoring strategy development, and is not used to deny the overall long-term benefits of quitting smoking.

[0094] In one implementation, a short-term evaluation window W=[0,T] after smoking cessation is first defined. w ], where T w The preset window endpoint is preferably 2 or 3 years. Further, within the short-term evaluation window, a set of discrete time points {tk} is generated according to a preset time step Δt, where tk∈[0, T]. w The time step Δt can be 1 year, 0.5 years, 0.25 years, or other preset values, preferably consistent with the output resolution of the protein trajectory prediction model. It should be understood that a smaller time step usually results in higher decision sensitivity, but is more sensitive to model fluctuations; a larger time step results in more robust decision-making, but may reduce the ability to capture short-term fluctuations.

[0095] In one implementation, for each subject, an "immediate smoking cessation scenario" and a "continuous smoking scenario" are constructed, and corresponding predicted risk trajectories are generated at the same time point set {tk}. The predicted risk can be obtained by concatenating a protein trajectory prediction model and a risk model. For example, future protein expression values ​​are generated through a protein trajectory prediction model, and then input into a risk model (such as a Cox proportional hazards model or other preset risk models) to obtain risk values ​​at each time point. To avoid ambiguity, this document refers to RR. quit (t): Represents the predicted risk or relative risk at time point t under the scenario of immediate smoking cessation; RR continue (t): Represents the predicted risk value or relative risk at time point t under the continuous smoking scenario.

[0096] Preferably, the input conditions used in the two scenarios, except for the smoking status variable, are kept consistent to reduce the impact of differences in non-target variables on the judgment results. Optionally, covariates such as body mass index (BMI), social deprivation index (TDI), and age can be included simultaneously in both scenarios to maintain consistency in the prediction framework.

[0097] In one implementation, if at least n discrete time points tk exist within the short-term evaluation window and satisfy the following condition: RR quit (tk)>RR continue (tk)+δ

[0098] The subject is then classified as a "rebounder," meaning they exhibit a vulnerable transition window after quitting smoking. Here, tk represents the k-th predicted time point within the short-term assessment window; RR quit (tk) represents the predicted risk value at time point tk under the immediate smoking cessation scenario; RR continue(tk) represents the predicted risk value at time point tk under the continuous smoking scenario; δ is the preset risk difference threshold, used to suppress misjudgments caused by model noise, numerical errors, or slight fluctuations; n is the minimum number of time points that meet the above judgment conditions, used to improve the stability of "rebounder" identification.

[0099] In one embodiment, δ is a fixed constant; in another embodiment, δ can be determined based on the distribution of risk values ​​in the model output, the risk difference distribution statistic in the validation sample, or the minimum acceptable difference in the target application scenario. For example, δ can be determined in any of the following ways: (1) based on the validation sample... (1) The quantile is determined; (2) The stability is determined based on the principle of optimal determination under cross-validation; (3) The sensitivity / specificity is determined based on the preset target; (4) The threshold is preset by the clinical user as an empirical threshold. The present invention does not limit the specific method of determining δ.

[0100] In one implementation, n is a fixed positive integer; preferably, n is determined according to the short-term evaluation window length T. w It is set in conjunction with the time step Δt to avoid occasional fluctuations at a single point in time triggering the "rebounder" determination. For example, when T... w When Δt = 1 year and the number of discrete time points is relatively small, n can be 1 or 2. When Δt = 0.5 years or less and the number of discrete time points increases, n can be increased accordingly. Furthermore, n can also be expressed as a threshold proportion of the total number of time points within the window (e.g., the proportion of time points that meet the conditions is not less than a preset proportion), and this invention does not limit this.

[0101] To improve the robustness of the judgment results, in one embodiment, when comparing RR... quit (tk) and RR continue Before (tk), the risk trajectory can be smoothed, outlier truncation removed, or numerical stabilization performed. The smoothing process may include moving averages, local regression smoothing, or other preset smoothing methods; the outlier truncation may include cropping risk values ​​that exceed a preset range. Preferably, the risk trajectories for both scenarios use the same smoothing and truncation rules to avoid introducing systematic bias. In an alternative implementation, in addition to the above-mentioned "at least n time points exceeding the threshold" determination rule, any one or a combination of the following can also be used to determine the "rebounder":

[0102] (1) Within the short-term assessment window, (1) The maximum value exceeds the preset peak threshold; (2) Within the short-term evaluation window, (3) Within the short-term evaluation window, the length of a continuous time period that is positive and exceeds δ reaches the preset duration; The integral area (or discrete summation area) exceeds a preset area threshold. The above alternative criteria can be used alone or in combination with the aforementioned criteria to adapt to different data resolutions and application requirements.

[0103] In one embodiment, after determining that a subject is a "rebounder," the device or system can output a warning message stating "a vulnerable transition window exists after smoking cessation," and further output auxiliary information related to this determination, including but not limited to: the determination window length, time step, number of time points meeting the criteria, maximum risk difference, time point when the threshold is first exceeded, and recommended reassessment time. This output helps in the development of subsequent individualized follow-up, metabolic management, or supportive intervention strategies.

[0104] 7. Output and Intervention Recommendations

[0105] For individuals identified as "rebounders," the system can output a "vulnerable transition window warning after smoking cessation" and recommend enhanced follow-up, such as shortening the interval between follow-up visits / tests; strengthening metabolic management (e.g., weight management); and strengthening monitoring of lung function, blood indicators, or imaging. Preferably, higher priority monitoring is suggested for individuals who are heavy smokers, older, and have a higher BMI.

[0106] Example 5: A multi-protein combined early warning method for the risk of lung cancer and emphysema

[0107] This embodiment illustrates the multi-disease respiratory system risk assessment function of the present invention, which is particularly suitable for the joint assessment of lung cancer / emphysema risk in COPD patients and smokers.

[0108] 1. Lung Cancer Risk Assessment

[0109] In one implementation, CXCL17 is used as a core protein biomarker for lung cancer risk. The expression level of CXCL17 in the plasma of subjects is measured and compared with a preset threshold to output a high / low risk stratification for lung cancer. In another implementation, a multi-protein model is used to assess lung cancer risk, and the scoring formula can be a linear combination.

[0110]

[0111] Where m is the number of proteins used for lung cancer risk assessment, λk is the corresponding coefficient, and Z ik This represents the value of the k-th protein in the i-th subject.

[0112] 2. Emphysema Risk Assessment

[0113] In a preferred embodiment, an emphysema risk panel composed of the following four proteins is used: CXCL17, WFDC2, IL22, and TNR. After detecting and standardizing these four proteins, a comprehensive emphysema risk score is calculated.

[0114] Here, γ0 to γ4 are preset coefficients, corresponding to the model coefficients (weights) of the four proteins respectively. The ScoreEmphysema is compared with the threshold TEmphysema to output the emphysema risk stratification.

[0115] Joint report output

[0116] The system can output the following in the same report: COPD risk score and grade; lung cancer risk score and grade (or CXCL17 early warning result); and emphysema risk score and grade. For individuals with multiple high-risk indications, the report can indicate "Comprehensive High Risk for Respiratory System". Specific Implementation

[0117] This invention utilized plasma proteomic data from over 50,000 participants in the UK Biobank (Resource ID: 30900), comprising 2,923 proteins measured at enrollment using the Olink Explore platform. After quality control, 2,920 proteins and 38,133 participants were retained for analysis, matched against baseline characteristics, questionnaire data, and health records spanning up to 20 years.

[0118] This invention aims to characterize the proteomic features of smoking exposure, smoking cessation status, and COPD stage, and to identify biomarkers with potential preclinical risk prediction value. This invention also analyzes proteomic changes in specific subgroups after smoking cessation to provide a basis for individualized intervention strategies. Before model construction, the data were preprocessed and randomly divided into training and test sets in equal proportions.

[0119] The data for this invention were obtained from the UK Biobank (Resource ID: 30900). Plasma protein levels were measured using the OlinkExplore platform, which employs Proximity Extension Assay (PEA) technology, offering high sensitivity and specificity, and capable of simultaneously quantifying thousands of proteins. Plasma protein data were acquired as NPX-normalized values ​​and further quality-controlled. Proteins and participants with a missing value rate exceeding 30% were excluded. After quality control, residual missing values ​​were imputed using k-nearest neighbor (KNN, k = 5). Participants were required to have complete baseline demographic information, smoking history, and longitudinal outcome data. Participants whose smoking status was not categorized or whose smoking duration could not be calculated were excluded. The final analysis cohort included 38,133 participants.

[0120] Diagnostic data preprocessing

[0121] Clinical diagnostic information was derived from ICD-10 codes in hospital records, including baseline and follow-up data. COPD was defined as ICD-10 code J44 (other chronic obstructive pulmonary diseases), emphysema as J43, and lung cancer as C34. Only hospital-recorded clinical diagnoses were used, excluding self-reported illnesses. The J44 code distinguishes COPD from chronic bronchitis (J40–J42), which was not considered a COPD outcome in this study. Although some participants had lung function data (e.g., FEV1 / FVC), these were not used in the COPD definition in the primary analysis due to a high rate of missing data.

[0122] Follow-up data preprocessing

[0123] The invention compiled key longitudinal variables, including gender, baseline age, age at enrollment, body mass index (BMI), Townsend deprivation index (TDI), and time to first visit, time to smoking initiation, and time to cessation. These variables were used to construct individual follow-up timelines, determine key time points (such as time to cessation), and calculate follow-up duration. Based on cumulative smoking duration, participants were divided into four groups: never smoked, light smokers (<15 years), moderate smokers (15–30 years), and heavy smokers (>30 years). Smoking cessation status was further divided into current smokers, those who quit for <10 years, those who quit for ≥10 years, and never smoked. Baseline characteristics of different smoking stratification groups were compared. Smoking history was primarily self-reported and cross-validated with clinical records. Individuals with missing or internally inconsistent follow-up information were excluded.

[0124] Animal models of smoking exposure

[0125] To validate the results of the population study, this invention established a chronic tobacco smoke exposure model in A / J mice using the DSIBuxco Smoke Generator (DSI, USA). A / J mice were randomly assigned to a filtered air control group and a smoke exposure group (n = 8 per group). The exposure group inhaled the smoke from 12 cigarettes daily, 5 days a week, for 60 consecutive days. Each exposure lasted 3 minutes, followed by 15 minutes of fresh air. Plasma samples were collected on day 60. The control group was raised under the same conditions but without smoke exposure. A / J mice were chosen because they are highly susceptible to smoke-related lung injury and exhibit significant inflammation and emphysema-like changes. The 60-day exposure duration was considered to induce detectable COPD-like pathological changes. This model is not intended to directly correspond to human smoking intensity or pack-years, but rather to serve as a standardized experimental perturbation for cross-species comparison of smoking-related proteomic changes.

[0126] Identification of COPD risk-related stable protein biomarkers

[0127] Feature selection and model building were performed only on the training set to avoid information leakage. First, a univariate Cox proportional hazards model was built for each candidate protein in the training set, and they were ranked according to their association strength with newly diagnosed COPD. The top N proteins (N = 200, 500, or 1000, used for sensitivity analysis) were selected for subsequent penalty modeling. Then, a LASSO-Cox model (glmnet, α = 1) was constructed, using 10-fold cross-validation, and the one-standard-error rule was applied. Determine the regularization parameters. To evaluate the stability of feature selection, repeat the resampling 200 times, each time randomly selecting 80% of the training set samples to refit the model. Proteins with non-zero selection coefficients were recorded, and their selection frequencies were calculated. After multiple iterations, core proteins with a selection frequency >70% were identified. In a more rigorous resampling sensitivity analysis, three additional proteins were stably included in the model. Ultimately, a total of 11 proteins were retained for subsequent modeling.

[0128] Constructing a comprehensive protein risk score:

[0129]

[0130] In the independent validation set, the score was assessed using a multivariate Cox model (adjusted for age, sex, BMI, and TDI). The proportional hazards hypothesis was tested using the Schoenfeld residuals (R survival package cox.zph). Sensitivity analysis was performed by further adjusting the package year variable in a sub-cohort with complete package year information (N = 6,323). The results showed that the protein score remained significantly associated with COPD risk (HR = 1.98, 95% CI 1.85–2.13). ).

[0131] ProteinScore = 0.21457393 CXCL17+0.73792488 WFDC2+0.18887395 GDF15+0.12830138 ALPP+0.25248641 MMP12−0.47919873 TNR−0.22955520 BCAN−0.40041569 SCGB1A1+0.18104341 REN−0.75348940 EGFR−0.48531675 AGER.

[0132] After the model was constructed, this invention performed a proportional hazards hypothesis test on the constructed multivariate Cox proportional hazards model to assess the reasonableness of the model assumptions. The Schoenfeld residual method was used to test the protein risk score and each covariate (BMI, gender, Townsend deprivation index, and age at enrollment) included in the model, and the global test statistic was calculated. The results showed that the chi-square value of the protein risk score was 0.288 (degrees of freedom = 1, p = 0.59), the chi-square value of BMI was 0.168 (p = 0.68), the chi-square value of gender was 0.287 (p = 0.59), the chi-square value of Townsend deprivation index was 2.003 (p = 0.16), and the chi-square value of age at enrollment was 1.308 (p = 0.25). The p-values ​​for all individual variables were greater than 0.05, indicating that there was no significant interaction between the variables and time, i.e., the hazard ratio remained constant during the follow-up period. The chi-square value of the global test was 4.108 (degrees of freedom = 5, p = 0.53), which was also not significant, further confirming that the overall model satisfies the proportional hazards assumption. Therefore, the COPD risk assessment model of this invention has a robust statistical basis and can be used for subsequent survival analysis and risk prediction.

[0133] To assess whether multicollinearity exists among the predictor variables in a model, this invention calculates the Variance Inflation Factor (VIF). This index quantifies the degree of variance inflation in coefficient estimates caused by the correlation between independent variables in a regression model. Generally, a VIF value greater than 5 or 10 indicates severe multicollinearity, which may affect the stability of the model estimation.

[0134] In this model, the VIF of all included variables was calculated, including clinical covariates (BMI, sex, Townsend deprivation index, and age at enrollment) and 11 target proteins (CXCL17, WFDC2, GDF15, ALPP, MMP12, TNR, BCAN, SCGB1A1, REN, EGFR, and AGER). The results showed that the VIF values ​​of all variables were at low levels: BMI had a VIF of 1.25, sex 1.22, Townsend deprivation index 1.13, and age at enrollment 1.35; the VIF values ​​of each protein ranged from 1.21 (EGFR) to 3.18 (CXCL17), all significantly lower than the commonly used threshold of 5. This result indicates that there is no significant multicollinearity among the predictor variables in this model, the model parameter estimates are stable and reliable, and the interpretation of the protein risk scores and their coefficients is statistically effective.

[0135] To evaluate the stability of the model during feature selection, this invention conducted sensitivity analyses on candidate protein sets of different sizes. In the training set, this invention selected the top 200, 500, and 1000 proteins most strongly associated with COPD as candidate features, respectively. A 10-fold cross-validation was performed using the LASSO-Cox model, and the optimal regularization parameter for each candidate set was determined according to the one-standard-error rule (i.e., selecting the largest regularization parameter within one standard error range of the minimum cross-validation error). By repeating random subsampling fitting 200 times, we calculated the results for different candidate set sizes. The mean and standard deviation. The results showed that when the number of candidate proteins was 200, The mean was 0.0167 and the standard deviation was 0.0036; when the number of candidate proteins increased to 500, The mean was 0.0153 and the standard deviation was 0.0033; when the number of candidate proteins further increased to 1000, The mean was 0.0148, and the standard deviation was 0.0033. It can be seen that as the number of candidate proteins increases, The mean showed a slight downward trend, but remained at the same order of magnitude overall, with a small and stable standard deviation. This result indicates that the regularization strength selected by the model is insensitive to the size of the candidate protein set, the feature selection process has good stability, and the 11 core proteins ultimately selected do not depend on a specific candidate set selection, but rather reflect robust molecular signals related to COPD risk.

[0136] Unsupervised embedding and clustering of residual proteomes

[0137] To explore molecular heterogeneity at similar exposure levels, this invention focuses on heavy current smokers without baseline COPD or lung cancer. Linear regression residuals were performed on 11 proteins (adjusted for age and smoking duration) to preserve inter-individual protein variability. PCA was then performed, retaining the top 6 principal components that cumulatively explained approximately 90% of the variance. UMAP embedding (n...) was performed using the R package uwot. neighbors = 30, min dist= 0.1, Euclidean distance, fixed random seed). k-means clustering was performed in a two-dimensional embedding space (25 random initializations). The optimal number of clusters was determined to be 2 by silhouette coefficient analysis (K = 2–8). Cluster stability was assessed by bootstrap resampling (100 times, 80% subsampling) and quantified using the adjusted Rand index (ARI), showing high stability (median ARI = 0.996). Further validation of method consistency was performed using Ward.D2 hierarchical clustering, with ARI = 0.76 in UMAP space and ARI = 0.63 in PCA space, indicating that the structure is not method-specific. It should be emphasized that these clusters are analytical partitions on the principal variation axis, rather than explicit biological subtypes.

[0138] Smoking cessation risk model construction

[0139] A multilayer perceptron (MLP) model was built using scikit-learn and trained solely on the current smoker. Each MLP contains three hidden layers (128, 64, and 32 neurons respectively), a ReLU activation function, and an Adam optimizer (learning rate...). The model was initialized with Xavier, using the MSE loss function, and trained for a maximum of 1000 epochs with an early stopping strategy. Input variables included age, BMI, smoking duration, years since quitting (current smokers were set to 0), and TDI, all standardized using z-scores. The model simulated a 15-year protein trajectory and incorporated it into a Cox model (including the interaction term of BMI, TDI, protein level, and smoking status × intensity) to predict COPD risk. To verify robustness, a sensitivity analysis was performed using random forest regression, and the consistency of risk trends simulated by the MLP and random forest was compared.

[0140] Statistical analysis

[0141] Differential protein analysis was performed using limma (FDR < 0.05). GO / KEGG enrichment was performed using clusterProfiler (FDR < 0.05). GSEA was performed using fgsea (MSigDB v7.5.1). Protein heterogeneity was analyzed using UMAP and hierarchical clustering. ROC curves were used to evaluate the model's discriminative ability (AUC). All analyses were performed in R (v4.3.0) and Python (v3.13.5). The main R packages used included limma, clusterProfiler, ggplot2, ComplexHeatmap, and umap; the Python package used was scikit-learn (v1.6.1).

[0142] result:

[0143] 1. Smoking intensity gradient is associated with immune-metabolic axis imbalance.

[0144] To characterize dose-dependent proteomic alterations, this invention analyzed plasma proteomic data from UK Biobank participants measured using the Olink Proximity Extension Analysis (PEA) platform and assessed how cumulative smoking exposure remodels the circulating proteome.

[0145] Participants were divided into four groups based on smoking duration: never smokers, light smokers (<15 years), moderate smokers (15–30 years), and heavy smokers (>30 years), with never smokers serving as the reference group. Differential expression analysis showed that 1,958 proteins were significantly altered in heavy smokers compared to never smokers (Figure 1A).

[0146] To validate the findings experimentally, a mouse model of chronic cigarette smoke exposure was established (Figure 1B), and a consistent pattern of proteomic changes was observed, including elevated levels of Il6, Notch3, Tnfrsf11b, and Il1b. Cross-species correlation analysis was limited to proteins significantly affected by smoking in both humans and mice, and the results showed a high degree of consistency in the direction of changes in smoking-related proteins between humans and mice (Pearson r = 0.742, P = 0.004; Figures 1C, D).

[0147] Gene ontology enrichment analysis of smoking-related differentially expressed proteins revealed that they are primarily involved in host defense responses and immune cell chemotaxis (Figure 1E). Pathways related to antimicrobial defense and myeloid and lymphoid cell chemotaxis were significantly enriched, accompanied by systemic activation of chemokine and interleukin-1 signaling pathways, suggesting enhanced inflammation. These results indicate that cigarette smoke is associated with abnormal immune cell recruitment and tissue homeostasis disturbances, potentially promoting persistent inflammation.

[0148] To further elucidate the co-expression patterns of proteins associated with smoking exposure, this invention performed unsupervised hierarchical clustering analysis on smoking-related differentially expressed proteins (Figure 1F). KEGG enrichment analysis identified six protein clusters with different exposure-dependent patterns.

[0149] Cluster 2 was significantly enriched in immune-related pathways—including cytokine-receptor interactions, viral protein-cytokine-receptor interactions, chemokine signaling pathways, and complement cascades—characteristic proteins including TNF, IL6, and CXCL8. These proteins remained at low levels in light and moderate smokers, but were significantly upregulated in heavy smokers, showing a trend of increasing inflammatory signaling with increasing exposure.

[0150] Cluster 3 primarily involves metabolic regulation, including general metabolic processes, pyruvate metabolism, and peroxisome function. This module contains various detoxification and mitochondrial-associated enzymes (SOD1, SOD2, ADH1B, ADH4, MVK, MECR), which are elevated in the mild exposure group and decreased in heavy smokers, suggesting an early compensatory response to smoke-induced oxidative stress, followed by potential functional exhaustion.

[0151] Cluster 4 encompasses cardiovascular and neuroendocrine regulatory pathways—including cardiomyopathy, myocyte cytoskeleton dynamics, hormone signaling, and cell adhesion—and its key proteins (CACNB1, AGT, TNNI3, and LEP) are consistently downregulated in smokers, suggesting that cardiovascular-endocrine homeostasis may be impaired in the long term.

[0152] Although cluster 5 has limited pathway enrichment, it contains a variety of proteins associated with airway remodeling and chronic respiratory inflammation (ADAM9, ALDH3A1, IL1B, CEACAM6, MMP9, SCGB3A1), consistent with extracellular matrix degradation, mucosal damage and immune cell recruitment observed in long-term tobacco exposure.

[0153] Overall, these results indicate that with increasing smoking intensity, significant immune-metabolic remodeling occurs in the body, characterized by enhanced inflammatory signaling, disordered immune homeostasis, and weakened tissue repair processes, exhibiting a clear dose-dependent trend.

[0154] 2. Proteomic biomarkers can predict COPD risk in different smoking characteristics and molecularly defined susceptibility-tolerance phenotypes.

[0155] To explore the association between smoking and disease, we mapped ICD-10 phenotypes to smoking status and performed correlation analysis. Multiple diseases were significantly associated with smoking, with chronic obstructive pulmonary disease (COPD) showing the most significant association (Figure 2A). Consistent with epidemiological observations, the cumulative incidence of COPD differed among different smoking status groups in this cohort, with a lower incidence in former smokers than in current smokers. Based on these observations, this invention conducted proteomics-based risk modeling analysis to explore potential mechanisms (Figure 2B).

[0156] Given the known role of smoking in the pathogenesis of COPD, this invention applied LASSO regression to 1,958 smoking-related differentially expressed proteins (DEPs) in a training cohort to screen for COPD risk-related protein biomarkers. Using the minimum λ criterion in cross-validation, LASSO screened 330 protein features from the 1,958 DEPs. Gene ontology enrichment analysis of these candidate proteins revealed that they primarily involve processes consistent with smoking biology, including immune cell migration and infiltration, extracellular matrix remodeling, and membrane / receptor-related activities. These pathways are consistent with the core chronic inflammation and airway remodeling in COPD pathogenesis.

[0157] To obtain a stable predictive model, this invention employs a one-standard-error rule combined with repeated resampling LASSO analysis, ultimately identifying 11 proteins associated with stable COPD status: ALPP, CXCL17, WFDC2, MMP12, GDF15, SCGB1A1, TNR, BCAN, REN, EGFR, and AGER (Figure 2C). In the independent validation cohort, this combination of 11 protein features outperformed the model based solely on smoking category in COPD discrimination (AUC = 0.844 vs 0.773). Figure 2 D). The comprehensive protein score constructed based on these biomarkers was negatively correlated with the age of COPD onset (Pearson r = −0.12, ).

[0158] Notably, the 11-protein score still retains predictive value in never-smokers and can identify individuals at high risk for emphysema and lung cancer, suggesting that the protein panel may reflect shared molecular perturbations unrelated to exposure in these respiratory diseases. This invention compares the score distributions among four smoking status groups and classifies individuals into high-risk or low-risk groups based on data-driven thresholds (Figure 2E). Some never-smokers were classified as high-risk, while low-risk individuals were present at all exposure levels, highlighting the significant heterogeneity of COPD susceptibility.

[0159] To further analyze this heterogeneity and minimize the confounding effects of environmental, occupational exposure, and smoking cessation behaviors, this invention focuses on heavy smokers who have not quit. After residualizing protein expression based on age and smoking duration, this invention uses UMAP for dimensionality reduction and combines it with k-means clustering to identify two subgroups (Figure 2F). Individuals in Cluster 1 had significantly higher cumulative COPD incidence rates during 5, 10, and 15 years of follow-up, and the differences persisted after adjusting for age and smoking duration (Figure 2G–H). Since the exposure burden was comparable between the two groups, this difference in risk suggests that molecular characteristics rather than simply exposure dose may play a role.

[0160] Differential expression analysis and KEGG enrichment results (Figure 2I) showed that the low-risk Cluster 2 was characterized by enhanced metabolic and detoxification programs—including enhanced cytochrome P450-mediated detoxification pathways—suggesting a higher clearance capacity for tobacco-derived toxins; conversely, Cluster 1 was enriched with pro-inflammatory signaling pathways, suggesting a weaker capacity for inflammation control.

[0161] In summary, these results identified a compact set of proteomic biomarkers in this cohort that can predict COPD risk in both smokers and non-smokers, and revealed a dual molecular profile of "susceptibility-tolerance" among smokers at the same exposure level. This molecular stratification provides a foundation for early risk identification and precise prevention of smoking-related COPD.

[0162] 3. Molecular residual effects after smoking cessation: persistent changes and withdrawal-related shifts

[0163] Smoking cessation is the most effective intervention to slow the progression of COPD and reduce its incidence. However, the molecular characteristics of former smokers are not the same as those of non-smokers, and differences remain in immune, metabolic, and structural pathways. Analysis of 11 previously identified protein biomarkers in former smokers showed that these proteins did not consistently revert to low-risk levels (Figure 3A), suggesting that the physiological changes are not a simple risk reversal, but rather a more complex molecular remodeling.

[0164] To systematically characterize this dynamic change, this invention analyzed the peripheral blood proteome of four population groups: current smokers, those who quit smoking less than 10 years ago, those who quit smoking ≥10 years ago, and never smokers. Among all 2,920 proteins, this invention used the LASSO (Least λ Criterion) to screen for 408 smoking cessation-related proteins that showed significant changes at different stages of smoking cessation. Unsupervised clustering analysis based on these proteins identified three distinct response patterns (Figure 3B), and functional enrichment support was obtained for each (Figure 3C): a near-normalization pattern, a persistent alteration pattern, and a post-smoking cessation shift pattern.

[0165] Proteins in clusters 1 and 5 are aberrantly expressed in current smokers, but gradually approach baseline levels in former smokers, particularly in long-term smokers. Enrichment analysis suggests that these proteins are involved in immune cell proliferation, migration, activation, and T cell signaling and adhesion-related pathways, consistent with the characteristics of partial adaptive immune function recovery over time.

[0166] In contrast, clusters 2 and 3 exhibited persistent alterations: their expression remained abnormal even after long-term smoking cessation, and in some cases deviated even further from that of never-smokers. Enrichment analysis suggested involvement in immune tolerance (negative regulation of exogenous stimuli), leukocyte migration, and lipopolysaccharide (LPS) responses, indicating long-term alterations in innate immune signaling pathways. Cluster 3 was also enriched in bone resorption and systemic metabolic pathways, consistent with inflammation-related metabolic disorders and bone loss, suggesting that these changes are difficult to completely reverse even after smoking cessation.

[0167] It is worth noting that although both cluster 3 and cluster 5 are involved in immune cell recruitment, only cluster 5 contains T cell-specific pathways, suggesting that T cell function may be more easily restored than other immune components, while innate inflammatory signals and metabolic abnormalities tend to persist.

[0168] The third pattern emerged in clusters 4 and 6. In these clusters, protein levels in current smokers were similar to those in never-smokers, but significantly dysregulated after quitting, deviating from baseline levels. Cluster 6 enriched TNF and LPS response pathways and T cell activation pathways, and overlapped with cluster 3 in humoral regulation, suggesting a post-quittal inflammation-related shift characterized by a specific enhancement of airway secretory activity, an alteration not observed in continuing smokers. Although cluster 4 did not reach statistical significance in GO enrichment analysis, it contained several well-known COPD-related proteins (AGER, COL1A1, CTSB, IL5) involved in epithelial maintenance, extracellular matrix remodeling, and airway inflammation. Contrary to normalization, these proteins remained abnormally suppressed, consistent with the characteristics of immune exhaustion or impaired tissue repair capacity inherited from chronic smoke exposure. Overall, clusters 4 and 6 revealed a unique post-quittal molecular profile characterized by persistent inflammation and impaired tissue repair function.

[0169] Based on the aforementioned 11 protein biomarkers, this invention calculated a comprehensive risk score in former smokers, finding that those who quit smoking within 3 years had significantly lower scores than current smokers (Figure 3D). To explore the mechanism behind this early risk transition, this invention performed GSEA analysis on heavy current smokers and heavy smokers who quit smoking within 3 years. The results showed that many pathways remained significantly altered in recent smokers (Figure 3E). Further focusing on the core proteins driving these signals revealed two opposing trends (Figure 3F): the levels of IDO1, KYNU, BCAT1, LEP, and NOS2 in recent smokers were comparable to those in current smokers, suggesting that the state of immune metabolic activation persisted; while the levels of structural damage-related biomarkers (MMP8, MMP9, SFTPD) in recent smokers were lower than those in current smokers, suggesting that these proteins showed a partial recovery trend. This combination of characteristics suggests that in the early stages after heavy smokers quit smoking, there exists an atypical transitional phenotype: improved airway integrity coexisting with systemic immune-metabolic instability, which may represent a transient window of vulnerability.

[0170] Overall, these results challenge the simplistic paradigm that "quitting smoking restores molecular homeostasis." Instead, the smoking cessation process reveals a complex remodeling: some systems (such as T-cell-related processes) show reversible recovery; others (such as impaired immune tolerance, enhanced LPS inflammatory signaling, and metabolic disorders) exhibit persistent damage; and still others show specific dysregulation after withdrawal, manifesting as inflammatory rebound and abnormal tissue remodeling. During the most pronounced phase of risk fluctuations, beneficial adaptations and detrimental changes coexist, suggesting that smoking cessation is more of a multi-layered biological recalibration process than a simple reversal.

[0171] 4. MLP modeling based on proteomics suggests a fragile transition window related to lipid dysregulation after smoking cessation in heavy smokers.

[0172] Given the paradoxical biological changes observed after smoking cessation, this invention further explores whether this period of rapid change may introduce additional risks and attempts to go beyond simple cross-sectional inferences. This invention trains a multilayer perceptron (MLP) regression model based on plasma proteomic data from current smokers to simulate long-term protein trajectories under two counterfactual scenarios: continued smoking and immediate cessation, stratifying them according to baseline smoking intensity (Figure 4A).

[0173] For each of the 11 key proteins, we constructed an independent MLP model containing three hidden layers (128, 64, and 32 neurons, ReLU activation function) and employed the Adam optimizer (learning rate...). The loss function is the mean squared error. The model input features include age at enrollment, body mass index (BMI), Townsend deprivation index (TDI), total smoking duration, and years since quitting smoking (current smokers are set to 0).

[0174] The trained MLP model generated 15-year protein expression trajectories for each participant under two scenarios. Baseline smoking intensity was categorized as mild (≤15 years), moderate (16–30 years), and severe (>30 years) based on total smoking duration. The simulated protein levels were then incorporated into a Cox proportional hazards model to estimate COPD risk, adjusted for BMI, TDI, smoking duration, and years since quitting.

[0175] The model predictions showed a risk curve consistent with the smoking dose: even after quitting smoking, the predicted COPD risk in all groups increased over time, and the higher the previous smoking intensity, the greater the increase (Figure 4B). At the time of quitting smoking, all groups predicted an immediate relative risk decrease: approximately 60% for light smokers, approximately 43% for moderate smokers, and approximately 20% for heavy smokers (Figure 4C). The risk decrease was most significant in the first 3 years after quitting smoking, with the most prominent differences between different intensity groups, suggesting that the early benefits for heavy smokers are relatively limited and the risk decrease is slower, which is consistent with epidemiological observations.

[0176] Risk simulations further suggest that approximately 10% of heavy smokers do not exhibit a sustained decline in risk after quitting. Instead, for these individuals, the model predicts a higher COPD risk after quitting compared to a scenario where continued smoking would have resulted in a potentially increased risk after quitting. Within the 0–2 year window after quitting, these "rebound" events primarily occur in the year heavy smokers quit (Figures 4D–E).

[0177] Within this high-risk subgroup identified by the model, older individuals and those with higher BMIs were more prevalent, and the increase in BMI after quitting smoking was only observed in heavy smokers (Figure 4F). Therefore, heavy smokers predicted to experience an early resurgence in risk are often older and overweight, suggesting a possible interaction between age-related comorbidities and metabolic factors. These results suggest that a susceptible subgroup may exist among recent smokers who may experience a brief period of increased risk before reaping the long-term benefits of quitting.

[0178] To investigate why the risk of rebound is primarily limited to heavy smokers, this invention compared the proteomic characteristics of "rebounders" and "non-rebounders" and analyzed specific changes in heavy former smokers (Figure 4G). Compared to non-rebounders, rebounders exhibited more pronounced metabolic dysregulation during smoking, particularly higher leptin (LEP) levels and lower IGFBP1 levels. This suggests that LEP may continue to rise after smoking cessation, thereby exacerbating immuno-metabolic disorders. Previous studies have shown that leptin signaling can exacerbate neutrophilic airway inflammation by promoting the activation of pro-inflammatory (M1) macrophages and chemokine-mediated neutrophil recruitment.

[0179] In a short-term follow-up study of moderate and heavy former smokers, this invention observed a decrease in neutrophils and basophils in both groups. However, adverse changes in liver function indicators and complex changes in multiple lipoprotein subtypes were observed only in heavy former smokers (Figure 4H), while similar phenomena were not observed in moderate former smokers. These results suggest that quitting smoking in heavy smokers may be accompanied by liver function disturbances and lipid metabolism instability.

[0180] Lipid-centric metabolic changes are consistent with the higher baseline fat accumulation in heavy smokers and may be further amplified by age and post-quittal weight gain, which may explain the early risk increase predicted by the MLP model. In older and overweight heavy smokers, persistent inflammatory signals and metabolic disturbances may still exist in the early stages after quitting.

[0181] Although smoking cessation generally reduces risk, the above results suggest that individualized strategies (such as close clinical monitoring, metabolic management, and supportive care) may be needed for high-risk individuals in the first few years after quitting to ensure they fully realize the long-term benefits of smoking cessation.

[0182] 5. Multiprotein biomarker panel for predicting the risk of respiratory comorbidities

[0183] In COPD patients, this invention further explores plasma protein biomarkers associated with the risk of lung cancer and emphysema. Using the LASSO-Cox modeling framework, which combines high-dimensional feature screening with survival analysis, this invention analyzed a cohort of 916 COPD patients without baseline emphysema. This method identified the chemokine CXCL17 as the most robust predictor of subsequent lung cancer development.

[0184] In the multivariate Cox regression model, after adjusting for demographic and clinical covariates, higher levels of CXCL17 were independently associated with a significantly increased risk of lung cancer (HR = 3.21, 95% CI: 1.49–6.91, p = 0.003) (Figure 5A). Kaplan–Meier curves further confirmed this association, with the cumulative incidence rate in the high CXCL17 group being significantly higher than that in the low CXCL17 group (log-rank p < 0.0001; Figure 5B). This result indicates that CXCL17 is overexpressed in lung adenocarcinoma and promotes tumor cell migration, invasion, and recruitment of tumor macrophages.

[0185] In the COPD cohort, this invention subsequently modeled and analyzed new-onset emphysema. Multivariate Cox regression identified a predictive panel consisting of four proteins—CXCL17, WFDC2, IL22, and TNR. WFDC2 and IL22 were associated with increased risk (WFDC2: HR = 2.42, p < 0.001; IL22: HR = 1.37, p = 0.04), while TNR showed a protective association (HR = 0.44, p = 0.003) (Figure 5C). Kaplan–Meier analysis based on a comprehensive protein risk score clearly distinguished between high-risk and low-risk individuals, with the high-risk group developing emphysema more rapidly (Figure 5D).

[0186] IL-22 can promote emphysema by participating in airway remodeling; WFDC2 is a lung epithelial secretory protein involved in various lung diseases; CXCL17 can recruit macrophages and regulate mucosal inflammation; TNR, as an extracellular matrix glycoprotein, may participate in tissue remodeling processes. This evidence collectively supports the biological rationale for this warning panel. Subsequently, this invention...

[0187] The generalization ability of this proteomic feature was assessed in COPD individuals. The comprehensive risk score constructed based on the above biomarkers retained strong predictive power for both lung cancer and emphysema in the non-COPD cohort (Figure 5E–F), indicating that the captured risk signals are not limited to the diagnosed COPD population, but also reflect a broader range of smoking-related lung susceptibility.

[0188] Overall, these results construct an integrated risk assessment framework based on proteomics, covering a variety of smoking-related respiratory diseases. By stratifying COPD and non-COPD populations, this multi-protein panel overcomes the limitations of single-disease prediction, supporting integrated early warning and personalized screening strategies, thereby providing a basis for precision prevention of lung cancer, emphysema, and COPD in high-risk smokers.

[0189] It should be understood that although the above description of the sample processing, data detection, model calculation, and result interpretation procedures involved in this invention is in the form of steps, these steps can be implemented through a combination of manual operation and semi-automatic systems, or through integrated computer program products or dedicated devices. Accordingly, this invention provides a respiratory disease risk assessment device based on plasma proteomics, which includes at least: a data receiving module for acquiring the detection results of target proteins in in vitro plasma samples of subjects; a preprocessing module for performing quality control, missing value processing, and standardization on the detection results; a risk score calculation module for inputting the standardized protein expression values ​​into a preset risk assessment model to calculate a COPD risk score; a risk stratification module for outputting risk stratification results according to a preset threshold; and an output module for presenting a risk stratification report. In a preferred embodiment, the device may further include a smoker molecular stratification module, a post-smoking cessation vulnerable transition window assessment module, and / or a comorbidity risk assessment module to achieve a refined assessment of smoker heterogeneity, short-term risk fluctuations after smoking cessation, and the comorbidity risk of lung cancer / emphysema. The above modules can be implemented by hardware circuits, software programs, or a combination thereof, and are preferably integrated into the reagent kit's accompanying software system or a standalone risk assessment workstation.

[0190] Comparative Example 1: Sensitivity analysis of random substitution for protein panel robustness

[0191] To verify the robustness of the 11 protein combinations defined in this invention (hereinafter referred to as "Panel A") as COPD risk prediction features and to rule out the possibility that their predictive performance may stem from random combinations, this invention designed and implemented a sensitivity analysis based on a random substitution strategy. This analysis aimed to compare the predictive performance differences between Panel A and a large number of randomly generated alternative panels containing different protein combinations.

[0192] 1. Research Design and Methodology

[0193] 1.1 Data Preparation and Population Screening

[0194] This invention uses the same data source as the main analysis (Final_Merged_Data.csv). To ensure the accuracy of the analysis, the data was first cleaned, including only participants who did not have the target disease at baseline. Individuals with lung cancer (baseline_C34 = "Yes"), emphysema (baseline_J43 = "Yes"), or COPD (baseline_J44 = "Yes") at baseline were excluded, and samples with a follow-up time ≤ 0 were also removed. The final outcome for the analyzed population was a new COPD event during the follow-up period (event_J44, 0 / 1), with a follow-up time of follow-up_time_J44 (in days). The adjusted clinical covariates in the model included body mass index (BMI), sex, Townsend deprivation index (TDI), and recruitment age.

[0195] 1.2 Partitioning of Training and Test Sets

[0196] To maintain consistency across all comparisons, this invention uses a fixed random seed (seed_split=123) to randomly divide the cleaned queue into training and test sets in a 1:1 ratio. All subsequent sensitivity analysis iterations are performed on this fixed-split dataset to ensure the comparability of comparison results between different protein panels.

[0197] 1.3 Construction of the candidate protein pool

[0198] The "candidate protein pool" used for random replacement is mainly derived from the list of differentially expressed smoking-related proteins listed in the Smoke.csv file. This invention takes the intersection of this list with the protein column names in the main dataset to ensure that all candidate proteins are present in the detection data of this invention. If the Smoke.csv file is unavailable or the number of candidate proteins is insufficient, the process degenerates into extracting all numerical protein variables from the main dataset, excluding follow-up time, outcome, clinical covariates, and baseline exclusion variables. To avoid confusion caused by introducing proteins identical to those in Panel A, this invention explicitly removes 11 proteins (CXCL17, WFDC2, GDF15, ALPP, MMP12, TNR, BCAN, SCGB1A1, REN, EGFR, AGER) from the candidate protein pool in Panel A.

[0199] 1.4 Random Replacement Strategy and Iterative Process

[0200] This invention sets B = 100 independent random iterations (seed_random = 2026). In each iteration, the following operations are performed:

[0201] Protein Removal: Randomly select k_out proteins from Panel A as "proteins to be removed", where the value of k_out is randomly determined between 5 and 6.

[0202] Adding proteins: randomly select k_in proteins from the candidate protein pool as "proteins to be added". To simulate possible panel shrinkage, the value of k_in is determined by the following rules: it is equal to k_out with a 75% probability, and equal to k_out-1 with a 25% probability (minimum is 1).

[0203] Generate a random panel: Construct the random panel (Panel_rand) for this iteration, which consists of: (all proteins in Panel A) - (proteins removed in this iteration) + (proteins added in this iteration).

[0204] 1.5 “Fair Subset” and Missing Control

[0205] To ensure fair comparison between Panel A and each generated random panel (Panel_rand) on the exact same sample set, this invention constructs a "fair subset" within both the training and test sets in each iteration. This subset requires that the samples simultaneously meet the following criteria: no missing data; ① outcomes and follow-up time; ② all clinical covariates; ③ all 11 proteins of Panel A; ④ all proteins of Panel_rand in this iteration. If the sample size of the constructed training or test set is less than 100, or the number of COPD events in the test set is less than 30, then that iteration is considered insufficient or unstable, and its results will be marked as missing (NA) and not included in the final performance comparison.

[0206] 1.6 Model Fitting and Performance Evaluation

[0207] In each iteration of the training set "fair subset", this invention fits two multivariate Cox proportional hazards models respectively:

[0208] Baseline model (Panel A): Surv(time, event) ~ BMI + gender + TDI + age + PanelA protein

[0209] Stochastic model (Panel_rand): Surv(time, event) ~ BMI + gender + TDI + age + Panel_rand protein

[0210] In the corresponding test set "fair subset", this invention calculates the linear predictor (lp) value of each model as a risk score, and uses the following metrics to evaluate its predictive performance:

[0211] Harrell's C-index: Evaluates the overall discriminative power of the model.

[0212] Area under the time-dependent ROC curve (AUC): Using the timeROC package and the marginal weighting method, the AUC values ​​for the 5th, 10th, and 15th years (i.e., 5×365, 10×365, and 15×365 days) are calculated to evaluate the predictive accuracy of the model at specific time points.

[0213] 1.7 Quantitative Comparison Indicators

[0214] To quantify the performance advantage of Panel A over a random panel, this invention calculates the following difference:

[0215] dC = C-index_PanelA - C-index_Panel_rand

[0216] dAUC_t = AUC_PanelA(t) - AUC_Panel_rand(t), where t = 5, 10, 15 years.

[0217] Meanwhile, this invention statistically analyzes the “win rate” of Panel A in 100 iterations (i.e., the percentage of iterations where dC>0 or dAUC_t>0), and calculates the mean, median, standard deviation, and key quantiles (5%, 25%, 75%, 95%) of the distribution of each difference (dC, dAUC_5y, dAUC_10y, dAUC_15y).

[0218] 2. Results

[0219] After 100 random iterations and rigorous fair subset selection, this invention obtained stable and reliable comparison results. Panel A exhibits a significant and consistent performance advantage over randomly generated alternative panels.

[0220] Throughout all valid iterations, the Panel A-based model consistently outperformed the randomized panel-based model on the test set in terms of C-index. Panel A achieved a "win rate" of 99% (i.e., dC > 0 in 99 out of 100 iterations). The mean C-index difference was +0.052, the median was +0.048, and its 95% confidence interval lower bound was significantly greater than 0. This indicates that Panel A systematically outperforms the randomly combined protein panels in predicting overall COPD risk.

[0221] Panel A demonstrated overwhelming performance superiority at the 5, 10, and 15-year time points. The mean and median of dAUC were positive at each time point, and Panel A's win rate reached or exceeded 98%. Particularly noteworthy was its superiority in early risk identification, with Panel A's AUC value averaging 0.067 higher than the random panel in the 5-year prediction.

[0222] This comparative example, through rigorous random substitution sensitivity analysis, confirms that the 11 protein combinations (Panel A) defined in this invention have significant and robust performance advantages in predicting COPD risk. Its discriminative power (C-index) and time-dependent prediction accuracy (AUC) systematically outperform a large number of randomly generated alternative protein panels, with a win rate approaching 100%. This result eliminates the possibility that the predictive effectiveness of the protein combinations in this invention is due to chance, and conversely verifies that this specific protein combination is an optimized and screened core feature set capable of stably capturing COPD risk signals.

[0223] Comparative Example 2: COPD Risk Assessment Model Based on 10 Alternative Proteins

[0224] To verify whether the predictive performance of the 11-protein combination (Panel A) defined in this invention stems from the selection of specific proteins, an alternative model (Panel B) consisting of 10 proteins was constructed as a comparative example. Panel B contains 6 proteins common to Panel A (CXCL17, WFDC2, GDF15, SCGB1A1, TNR, and AGER), and 4 other proteins reported in the literature to be associated with respiratory diseases (CA14, IGDCC4, ADM, and TNFRSF10B), for a total of 10 proteins. This comparative example aims to evaluate the COPD risk prediction ability of this 10-protein model in the same population and to directly compare it with Panel A of this invention.

[0225] 1. Materials and Methods

[0226] 1.1 Data Sources and Population Screening

[0227] The exact same dataset as this invention (a subset of the UK Biobank proteomics dataset) was used, with inclusion criteria and data cleaning procedures consistent with those used in this invention: individuals with baseline COPD, lung cancer, or emphysema were excluded, and samples with a follow-up time ≤ 0 were removed. The final analysis cohort included 38,133 participants, who were randomly divided into training and test sets in a 1:1 ratio using the same seed (seed_split=123), ensuring complete consistency with the data partitioning used for Panel A model training.

[0228] 1.2 Protein Detection and Preprocessing

[0229] The expression values ​​of the 10 proteins (CXCL17, WFDC2, GDF15, TNR, AGER, SCGB1A1, CA14, IGDCC4, ADM, and TNFRSF10B) in Panel B were all obtained from the detection results of the same Olink Explore platform. The data preprocessing steps were exactly the same as those in this invention: samples and proteins with a missing value rate >30% were deleted, K-nearest neighbor imputation (k=5) was used to fill in the remaining missing values, and the protein expression values ​​were z-score standardized based on the mean and standard deviation of the training set.

[0230] 1.3 Model Construction and Evaluation

[0231] On the training set, the expression values ​​of 10 proteins in Panel B were used as predictor variables, while adjusting for age, gender, BMI, and the Townsend deprivation index, to construct a multivariate Cox proportional hazards model. The model form is as follows: h(t|X) = h0(t) exp(β1 ALPP + β2 CXCL17 + ... + β10 CDCP1 + γ Covariates

[0232] In the test set, a linear predictive value was calculated for each participant as a risk score. The same assessment metrics as Panel A were used.

[0233] Harrell's C-index: Evaluates the overall discriminative power of the model.

[0234] Area under the time-dependent ROC curve (AUC): The AUC values ​​for year 5 (1825 days), year 10 (3650 days), and year 15 (5475 days) are calculated using the timeROC package (weighted method).

[0235] The predictive performance of Panel B was directly compared with the performance of Panel A (11 proteins) of the present invention on the test set, and the difference (Panel A – Panel B) was calculated.

[0236] 2. Results

[0237] Panel A's C-index was 0.8447, higher than Panel B's 0.7439. Panel A consistently showed an advantage in AUC for years 5, 10, and 15, with differences of +0.11099, +0.18036, and +0.10276, respectively.

[0238] Although Panel B already includes six core proteins common to this invention and additionally introduces four candidate biomarkers supported by literature, its predictive performance is lower than that of the 11-protein combination of this invention across all evaluation metrics. This consistent difference indicates that the specific combination of this invention carries additional molecular information that is independently related to COPD risk, information that cannot be fully captured by other common alternative proteins. Particularly noteworthy is that Panel A maintained its advantage in both short-term (5-year) and long-term (15-year) predictions, suggesting that this combination has more stable predictive ability.

[0239] This comparative example directly compared the 11-protein model of this invention with an alternative model consisting of 10 proteins. The results showed that the protein combination of this invention exhibited slightly better predictive performance in both C-index and AUC at multiple time points. This was statistically significant in a large population and consistent in direction, confirming the superiority and irreplaceability of the 11 proteins screened in this invention as biomarkers for COPD risk assessment. These results further support the inventiveness and practicality of the technical solution of this invention, namely, that more accurate COPD risk stratification can be achieved through specific multi-protein combinations.

[0240] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the claims.

Claims

1. A chronic obstructive pulmonary disease risk assessment device, characterized in that, include: The data receiving module is used to receive the detection results data of the target protein in the in vitro plasma sample of the subject; The preprocessing module is used to preprocess the detection result data, and the preprocessing includes quality control, missing value handling and standardization. The risk scoring calculation module is used to input the preprocessed protein expression data into the preset risk assessment model; The risk stratification module is used to compare the chronic obstructive pulmonary disease risk score with at least one preset threshold and output the risk stratification result; The output module is used to output the risk stratification results and / or the corresponding risk assessment report.

2. The apparatus according to claim 1, characterized in that, The target protein consists of 11 proteins: ALPP, CXCL17, WFDC2, MMP12, GDF15, SCGB1A1, TNR, BCAN, REN, EGFR, and AGER. The chronic obstructive pulmonary disease risk score (Score) is calculated using the following linear combination model. Score = + + + + + + + + + + ; Wherein, ALPP to AGER are the standardized expression values ​​of each target protein. to These are the preset model coefficients.

3. The apparatus according to claim 1, characterized in that, The risk assessment model satisfies the proportional risk assumption. According to the Schoenfeld residual test, the p-value of the interaction between protein risk score and time is greater than 0.05; and the variance inflation factor (VIF) of each predictor variable in the model is less than 5.

4. The apparatus according to claim 1, characterized in that, It also includes a smoker molecular stratification module, which is used for: Covariate correction was performed on protein expression data from current smokers to obtain residual protein expression characteristics. The residual protein expression characteristics are reduced in dimensionality and clustered to output susceptible or tolerant molecular subpopulation tags; Among them, the susceptible subgroup showed enrichment of pro-inflammatory signaling pathways and a significantly higher cumulative incidence of COPD than the tolerant subgroup.

5. The apparatus according to claim 4, characterized in that, The covariate correction includes using a linear regression model to remove the effects of age and smoking duration on protein expression values; The dimensionality reduction employs principal component analysis (PCA) and / or unified manifold approximation and projection (UMAP). The clustering was performed using the k-means algorithm, and the number of clusters was determined to be 2 through silhouette coefficient analysis.

6. The apparatus according to claim 1, characterized in that, It also includes a vulnerability transition window assessment module for smoking cessation, which is used for: Baseline variables of the subjects were received, including at least one of age, body mass index (BMI), duration of smoking, years since quitting smoking, and the Total Deprivation Index (TDI). A preset protein trajectory prediction model is invoked to generate the expression trajectory of the target protein for the next 15 years under both continuous smoking and immediate smoking cessation scenarios. Calculate time series risk values ​​under two scenarios based on risk models; Within the pre-set 0-2 year short-term evaluation window, when at least n time points satisfy RR... quit (t)>RR continue When (t) + δ, a vulnerable transition window after smoking cessation is determined and an early warning message is output; where RR quit (t) represents the predicted risk at time point t under the scenario of immediate smoking cessation, RR continue (t) represents the predicted risk at time point t under the continuous smoking scenario, and δ is the preset risk difference threshold.

7. The apparatus according to claim 6, characterized in that, The protein trajectory prediction model is a multilayer perceptron (MLP) regression model or a random forest regression model, and the risk model is a multivariate Cox proportional hazards model.

8. The apparatus according to claim 1, characterized in that, It also includes a comorbidity risk assessment module, which is used for: Lung cancer risk score is calculated based on the expression level of CXCL17 in in vitro plasma samples of subjects; and / or, emphysema risk score is calculated according to a linear combination model based on the standardized expression values ​​of four proteins: CXCL17, WFDC2, IL22 and TNR. It also outputs the corresponding comorbidity risk stratification results.

9. A method for assessing the risk of chronic obstructive pulmonary disease, characterized in that, Includes the following steps: The system receives detection results of target proteins in in vitro plasma samples from subjects. The target proteins consist of 11 proteins: ALPP, CXCL17, WFDC2, MMP12, GDF15, SCGB1A1, TNR, BCAN, REN, EGFR, and AGER. The detection result data is preprocessed, including quality control, missing value handling, and standardization. The preprocessed protein expression data is input into a preset risk assessment model, and the risk score of chronic obstructive pulmonary disease is calculated according to the linear combination model described in claim 1. The chronic obstructive pulmonary disease risk score is compared with at least one preset threshold, and the risk stratification result is output.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method of claim 9.

Images