Diabetes-related pancreatic cancer risk prediction method based on machine learning model and biological age

Abstract

The invention relates to the technical field of machine learning medical prediction, in particular to a diabetes-related pancreatic cancer risk prediction method based on a machine learning model and biological age, and the method comprises the steps: constructing a health reference population queue and a type 2 diabetes application verification queue; determining a core index panel through an automatic machine learning process, and training by adopting a regularization survival analysis model to obtain biological age and age acceleration; based on multi-modal features such as age acceleration, predicting a future pancreatic cancer absolute risk probability by using a machine learning competitive risk model, performing risk grade division, calculating an equal-risk age and supporting risk trajectory simulation according to the future pancreatic cancer absolute risk probability; and packaging the model into a risk prediction toolkit with an adaptive calibration function, and outputting a comprehensive risk assessment report. According to the method, the biological age is trained by adopting the pure health queue, so that the interference of the disease state on aging measurement is avoided, and the prediction precision of the pancreatic cancer risk of the type 2 diabetes mellitus population is improved.

Description

Technical Field

[0001] This application relates to the field of machine learning medical prediction technology, and in particular to a method for predicting the risk of diabetes-related pancreatic cancer based on machine learning models and biological age. Background Technology

[0002] In the field of machine learning-based medical risk prediction, building a pancreatic cancer early warning model for people with type 2 diabetes is a significant challenge. The high malignancy and difficulty in early diagnosis of pancreatic cancer make accurate risk stratification of this high-risk population an urgent clinical need. Although existing technologies have attempted to apply various machine learning models, their predictive performance is limited by two core bottlenecks: insufficient purity of training data and ambiguity in the biological meaning of feature engineering.

[0003] A key flaw in existing methods stems from the paradigm of training set construction. Currently, most machine learning models used to derive biological age or directly predict risk are trained on a mixed cohort of healthy individuals and patients. This paradigm leads to the learned features and generated biological age indicators being contaminated by known disease pathology information. When this indicator is applied to diabetic populations, the model cannot effectively distinguish whether the risk signal originates from underlying, universally accelerated physiological aging or merely reflects an existing diabetic state. This background noise severely limits the model's generalization ability and predictive specificity in high-risk populations.

[0004] Existing methods also have shortcomings in feature selection and model interpretability. Some models rely on conventional indicators that are abnormally late, leading to delayed warnings; while other models employ complex feature combinations, their impure training base makes it difficult to elucidate the relationship between biological age and pancreatic cancer risk, resulting in conclusions that cannot effectively guide clinical practice. More importantly, existing approaches generally fail to characterize the metabolic heterogeneity of diabetic patients and lack a robust indicator that can be extracted from multidimensional physiological data using machine learning, corresponding to the core processes of aging and uniformly quantifying this heterogeneity.

[0005] Therefore, there is an urgent need in this field for an innovative machine learning technology solution that can generate a biological age index that is not affected by specific diseases by constructing a pure health benchmark cohort and performing automated feature screening with clear biological guidance, thereby breaking through the bottlenecks of existing prediction models in terms of accuracy and generalization ability. Summary of the Invention

[0006] The purpose of this application is to provide a method for predicting the risk of diabetes-related pancreatic cancer based on machine learning models and biological age, in order to solve the problems of low prediction accuracy, delayed early warning, and inaccurate risk stratification caused by existing machine learning models when applied to specific prediction scenarios due to training data contamination, unclear biological significance of features, and insufficient generalization ability.

[0007] According to an embodiment of this application, a method for predicting the risk of diabetes-related pancreatic cancer based on a machine learning model and biological age is proposed, including the following steps: S1: Based on preset health screening criteria, construct a healthy reference population cohort and a type 2 diabetes application validation cohort; perform standardized preprocessing on the baseline clinical physiological indicators and biomarker data of the healthy reference population cohort to obtain a standardized dataset; extract the semantic embedding features of structured indicators and unstructured text from the electronic health records of the type 2 diabetes application validation cohort. S2: On the standardized dataset, using all-cause mortality, major adverse cardiovascular events, and serious renal events as joint supervised endpoints, a panel of core indicators was determined through an automated machine learning process; a machine learning regularized survival analysis model incorporating adversarial regularization was trained to obtain biological age (BA); S3: Calculate the age acceleration of individuals in the type 2 diabetes application validation cohort, use a machine learning competitive risk model to predict the absolute risk probability of pancreatic cancer in the next 3-5 years, classify the risk level accordingly, and calculate the age of pancreatic cancer and other risks; simulate the risk trajectory based on the biological age (BA) and the machine learning competitive risk model. S4: The validated machine learning competitive risk model, risk trajectory simulation algorithm and related parameters are packaged into a risk prediction toolkit containing an executable script; the toolkit has an adaptive calibration function and outputs a risk assessment report.

[0008] Furthermore, individuals who did not have diabetes, any malignant tumors, systemic inflammatory diseases, grade III-IV heart failure, grade B or higher liver failure, or stage 3 or higher chronic kidney disease at baseline were electronically screened from a large-scale prospective cohort database to form a healthy reference population cohort; and baseline clinical physiological indicators and biomarker data of the healthy reference population cohort were obtained.

[0009] A four-layer automated screening process is used to determine the core indicator panel for constructing biological age and the simplified sub-panels derived from it from the indicators of the standardized dataset, including: Usability and stability were screened, and indicators with a missing rate higher than 30% or an intragroup correlation coefficient lower than 0.6 were removed; Perform age-related trend screening and retain indicators whose correlation with actual age, after error detection rate correction, has a q value of less than 0.05; Multi-endpoint outcome-oriented stable selection was conducted, guided by joint supervised endpoints. A Cox-Elastic Net machine learning model with shared coefficients was used in conjunction with Bootstrap resampling technology to screen out indicators that are robustly related to disease risk and form a candidate indicator set. Feature dimensionality reduction and hierarchical classification are performed. The indicators are sorted according to the absolute value of the coefficients of each indicator in the candidate indicator set. The top K indicators are selected and classified into the basic physiological layer, metabolic regulation layer or cell senescence layer according to the three-level classification rules to form the core indicator panel and simplified sub-panels with different numbers of indicators, where K is a preset positive integer.

[0010] Furthermore, a Cox-Elastic Net model incorporating adversarial regularization terms is constructed and trained as a machine learning regularized survival analysis model. With the objective of minimizing the combination of the negative log-partial likelihood sum of the joint supervised endpoint and the adversarial loss, the data matrix of the core indicator panel or simplified sub-panel is trained to obtain the weight coefficient vector w, and a 1-dimensional aging axis Z is calculated based on this vector. The calculation formula is as follows: Where X represents the data matrix of the core indicator panel or the simplified sub-panel; w represents the weight coefficient vector obtained by Cox-Elastic Net regularization through the fusion of adversarial regularization; By fitting the Gompertz mortality law between the one-dimensional aging axis Z and mortality risk, calibration parameters are obtained. and The one-dimensional aging axis Z is calibrated as biological age BA, and the calculation formula is as follows: in, and The calibration parameters are obtained by fitting the Gompertz mortality law between the 1D aging axis Z and the risk of death. The formula for calculating age-related acceleration is: Where AgeAccel represents age acceleration; BA represents biological age; sex represents sex; CA represents actual age; and E(BA|CA,sex) represents the expected value of biological age BA predicted by a linear regression model based on actual age CA and sex.

[0011] When calculating age acceleration, two versions are calculated in parallel: one is the age acceleration that includes glycated hemoglobin when constructing the core indicator panel; the other is the age acceleration that does not include glycated hemoglobin when constructing the core indicator panel.

[0012] The machine learning-based competitive risk model is the Fine-Gray machine learning competitive risk model; the age acceleration is used as the core predictor variable, and is included in the model along with the new-onset diabetes status, body mass index, and estimated glomerular filtration rate as covariates; if the age acceleration used is a version calculated based on the core indicator panel that does not include glycated hemoglobin, then the glycated hemoglobin level is included in the model as a supplementary covariate.

[0013] Furthermore, when deploying the risk prediction toolkit, age acceleration that does not include glycated hemoglobin is preferentially used for calculation to improve generalization ability.

[0014] In summary, the beneficial technical effects of this application are as follows: By adopting a two-stage framework of modeling healthy individuals and applying it to diabetic individuals, we introduce biological age and age acceleration, which characterize the overall physiological aging rate, as core variables into the pancreatic cancer competitive risk model for type 2 diabetic individuals. This integrates long-term physiological decline information and significantly improves the ability to identify early risks.

[0015] The core indicators were determined through a four-layer automated screening process. The CoxElasticNet model with fusion adversarial regularization was used for training, and biological age was calibrated according to the Gompertz mortality law to ensure the reliability of the indicators and the stability of the model, so that the aging acceleration has a clear biological explanation basis.

[0016] The model integrates structured clinical indicators, biological age-derived variables, and semantic features extracted from electronic health record texts. Through risk trajectory simulation, it maps intervention measures into indicator adjustments, dynamically predicts future risks, and provides decision support for personalized health management.

[0017] The validated model and algorithm are packaged into an integrated toolkit with adaptive calibration capabilities. It can be optimized based on local data to ensure generalization performance. The output risk assessment report intuitively presents the age of equal risk, risk level and management recommendations, which is convenient for clinical promotion and application. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of a method for predicting the risk of diabetes-related pancreatic cancer based on a machine learning model and biological age, as provided in an embodiment of this application. Detailed Implementation

[0019] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0020] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0021] In the embodiments of this application, the same reference numerals denote the same components, and for the sake of brevity, detailed descriptions of the same components are omitted in different embodiments. It should be understood that the thickness, length, width, and other dimensions of various components in the embodiments of this application shown in the accompanying drawings, as well as the overall thickness, length, width, and other dimensions of the integrated device, are merely illustrative and should not constitute any limitation on this application.

[0022] In this application, "multiple" means two or more (including two).

[0023] The following is combined with Figure 1 The method for predicting the risk of diabetes-related pancreatic cancer based on machine learning models and biological age, according to embodiments of this application, will be described in detail below.

[0024] See appendix Figure 1 The following is a flowchart illustrating the steps of a method for predicting the risk of diabetes-related pancreatic cancer based on a machine learning model and biological age, as provided in this application embodiment.

[0025] Example This embodiment uses data from a specific evaluation period in the UK Biobank database. The specific operation procedure is as follows: The UK Biobank is a large biobank containing genetic, lifestyle, and health information of approximately 500,000 participants, and the scale and quality of its data provided an ideal basis for the implementation of this application.

[0026] S1: Queue Construction and Data Standardization Preprocessing Based on pre-defined health screening criteria, a healthy reference population cohort and a type 2 diabetes application validation cohort were constructed. The healthy reference population cohort was constructed using electronic screening from data of a specific evaluation period in the UK Biobank database. Specific screening criteria included: No self-reported history of diabetes and glycated hemoglobin <6.5%; All histories of malignant tumors were excluded based on ICD-10 codes (C00-C97); Exclude major systemic inflammatory or autoimmune diseases such as rheumatoid arthritis and systemic lupus erythematosus; Exclude heart failure classified as NYHA functional class III or IV; No severe liver failure: Exclude liver failure of Child-Pugh grade B or C; No severe kidney disease: Exclusion of stage 3 or higher chronic kidney disease, i.e., estimated glomerular filtration rate <60 mL / min / 1.73 mL / min. 2 Or, the urine albumin / creatinine ratio is ≥30mg / g.

[0027] It should be noted that the NYHA classification is a heart function classification standard developed by the New York Heart Association. Class III indicates significant limitation of physical activity, while Class IV indicates inability to perform any physical activity.

[0028] The Child-Pugh classification is a standard method for assessing liver reserve function in patients with cirrhosis. Grade B and Grade C represent moderate and severe liver dysfunction, respectively.

[0029] Following the rigorous screening process described above, 302,144 individuals were ultimately selected from 500,000 participants in the UK Biobank database to form a healthy reference population cohort.

[0030] A validation cohort for type 2 diabetes was constructed from the same database, and individuals meeting the following criteria were selected: At baseline, the patient had been diagnosed with type 2 diabetes by a clinician. No history of pancreatic cancer (excluded by ICD-10 code C25); It has complete baseline clinical physiological parameters data; Ultimately, 25,817 individuals were selected to form a validation cohort for the application of the treatment in type 2 diabetes.

[0031] The baseline clinical physiological parameters and biomarker data of the healthy reference population cohort were standardized and preprocessed to obtain a standardized dataset.

[0032] The baseline clinical physiological indicators and biomarker data of the healthy reference population cohort were systematically standardized and preprocessed, including imputation of missing values ​​using the random forest algorithm, Winsorize tail reduction of continuous variables, and transformation of non-normally distributed variables to make their distribution normal.

[0033] The Random Forest algorithm was used to perform multiple imputation of missing values ​​for continuous variables. RandomForestRegressor for continuous variables and RandomForestClassifier for categorical variables were used from the scikit-learn library. The number of decision trees (n_estimators) was set to 100, the maximum depth (max_depth) to 10, and other parameters were kept at their default values. Five complete datasets were generated, and the average of the five imputation results was used for subsequent analysis.

[0034] All continuous clinical physiological parameters were subjected to Winsorize reduction, which raised values ​​below the 1st percentile to the 1st percentile and lowered values ​​above the 99th percentile to the 99th percentile. For example, if the 1st percentile of systolic blood pressure is 90 mmHg and the 99th percentile is 180 mmHg, then all values ​​below 90 mmHg were adjusted to 90 mmHg, and all values ​​above 180 mmHg were adjusted to 180 mmHg.

[0035] The Shapiro-Wilk normality test was performed on all continuous variables at a significance level of α=0.05. Variables with p-values ​​less than 0.05 were determined to be skewed. Different transformation methods were used according to the numerical characteristics of the variables. For positive variables that have a zero value, the log1p transformation is used; For positive variables without zero values, use the natural logarithm transformation; For variables containing negative numbers, the Yeo-Johnson transform is used, and its parameter λ is determined by maximum likelihood estimation.

[0036] All numerical indicators after the above processing are stratified by gender and Z-score standardized to have a mean of 0 and a standard deviation of 1.

[0037] It should be noted that this series of data preprocessing steps is to ensure the quality of subsequent machine learning model training. Random forest imputation can handle complex nonlinear relationships between variables; Winsorize processing can suppress the influence of extreme values ​​while preserving the data distribution characteristics; and normalization transformation ensures that the data meets the premise assumptions of the statistical model.

[0038] Semantic embedding features of structured indicators and unstructured text were extracted from the electronic health records of the type 2 diabetes application validation cohort.

[0039] Pre-defined structured clinical indicators were extracted from electronic health records associated with the type 2 diabetes application validation cohort; and unstructured clinical text data were extracted, and semantic embedding features related to pancreatic cancer risk were extracted using a natural language processing model.

[0040] S2: On a standardized dataset, using all-cause mortality, major adverse cardiovascular events, and severe renal events as joint supervised endpoints, a panel of core indicators was determined through an automated machine learning process. A machine learning-regularized survival analysis model incorporating adversarial regularization was trained to obtain the biological age (BA). Through a four-layer automated screening process, the core indicator panel for constructing biological age and the simplified sub-panels derived from it were determined from the indicators of the standardized dataset: Usability and stability screening was performed by first removing indicators with a missing rate higher than 30% from all 42 baseline physiological indicators. For the remaining indicators, the within-group correlation coefficient was calculated, and an inverse random-effects model was used to remove indicators with an ICC lower than 0.6.

[0041] Age-related trend screening was performed, and the Spearman rank correlation coefficient between each indicator and the actual age was calculated. The Benjamini-Hochberg procedure was used to correct for the false detection rate, and indicators with a corrected q value of less than 0.05 were retained.

[0042] Multi-endpoint outcome-oriented stability selection used all-cause mortality, major adverse cardiovascular events (including myocardial infarction and stroke), and severe renal events (eGFR decrease ≥50% or progression to end-stage renal disease) as combined monitoring endpoints. A Cox-Elastic Net model with shared coefficients was employed, with the Elastic Net mixture parameter α set to 0.5, and stability selection was performed using 1000 bootstrap resampling cycles. The specific implementation process is as follows: Construct the joint likelihood function: Where L1, L2, and L3 are the Cox model partial likelihood functions for the three endpoints, respectively; α is the Elastic Net mixing parameter; β is the weight coefficient vector; and λ is the total regularization strength. The coordinate descent method is used for iterative optimization until the model coefficients converge. 1000 samplings with replacement were conducted in the healthy reference population cohort, with the sample size of each sampling being the same as that of the original cohort. Record the frequency with which each metric is selected (with a non-zero coefficient) in 1000 Bootstrap samples; Metrics that are selected more than 80% of the time in all Bootstrap samples are identified as core metrics.

[0043] Feature dimensionality reduction and hierarchical classification were performed. Candidate indicators were sorted based on the absolute value of their coefficients, and the top K=8 indicators were selected. Each indicator was then assigned to a basic physiological layer, a metabolic regulation layer, or a cellular senescence layer according to a three-tiered classification rule, resulting in a core indicator panel and simplified sub-panels with varying numbers of indicators. The indicators in the core indicator panel were clustered into three levels based on their physiological significance: a basic physiological layer reflecting basic organ function, a metabolic regulation layer reflecting glucose and lipid metabolism, and a cellular senescence layer reflecting cell damage and repair levels. The final eight core indicators were: systolic blood pressure, diastolic blood pressure, low-density lipoprotein cholesterol, glycated hemoglobin (for the glucose-containing version), eGFR, serum albumin, telomere length, and the inflammatory factor IL-6. Two types of core indicator panels were constructed: Sugar-containing panel: Includes all 8 indicators mentioned above The sugar-free panel includes 7 indicators excluding glycated hemoglobin. It should be noted that the intragroup correlation coefficient is an indicator that measures the stability and repeatability of data. An ICC > 0.6 indicates that the indicator has good stability.

[0044] Error detection rate correction is a multiple hypothesis testing correction method, and the q-value can be understood as the p-value after multiple test correction.

[0045] On a standardized dataset, using all-cause mortality, major adverse cardiovascular events, and severe renal events as joint supervised endpoints, a Cox-Elastic Net model incorporating adversarial regularization terms was constructed and trained as a machine learning regularized survival analysis model.

[0046] Using the data matrix X of the core metric panel, based on 12 years of follow-up data, the Elastic Net regularization strength parameter λ=0.032 was determined through 10-fold cross-validation. The data matrix of the core metric panel or a simplified sub-panel was trained with the objective of minimizing the combination of the negative log-partial likelihood sum of the joint supervision endpoint and the adversarial loss. The resulting weight coefficient vector w was then used to calculate the 1D aging axis Z according to the formula: Where X represents the data matrix of the core metrics panel or the simplified sub-panel; w represents the weight coefficient vector obtained by Cox-Elastic Net regularization.

[0047] Based on data from a healthy reference population cohort, the Gompertz mortality law was fitted between the 1-dimensional aging axis Z and mortality risk to obtain calibration parameters c0 and c1, and the 1-dimensional aging axis Z was calibrated as biological age BA.

[0048] In a healthy reference cohort, the correlation between the one-dimensional aging axis Z and all-cause mortality risk was validated to conform to Gompertz's mortality law. Calibration parameters were obtained by fitting the data using maximum likelihood estimation. =45.2, =8.1, the one-dimensional aging axis Z is calibrated as biological age BA, and the calculation formula is: in, and The calibration parameters are obtained by fitting the Gompertz mortality law between the one-dimensional aging axis Z and the risk of death. It should be noted that Gompertz's law of mortality is a classic model describing the exponential increase in mortality rates of organisms with age.

[0049] The calibration parameters were obtained by fitting the data using the maximum likelihood estimation method. Specifically, based on follow-up data from a healthy reference population cohort, the one-dimensional aging axis Z-value of each individual, along with their follow-up time and mortality status (0 = censored, 1 = deceased), were substituted into the Gompertz mortality function. The calibration parameters were then obtained by jointly solving the likelihood function. =45.2 and =8.1. This set of parameters makes the mortality risk predicted based on the aging axis Z best match the actual mortality distribution observed in the cohort.

[0050] This method employs 10-fold cross-validation, searching on a predefined λ-value grid (ranging from 0.01 to 0.1 with a step size of 0.002) to minimize the sum of the negative log-partial likelihoods of the joint supervision endpoint, ultimately determining the optimal ElasticNet regularization strength parameter λ = 0.032.

[0051] Based on data from a healthy reference population cohort, a linear regression model was established to predict biological age (BA) using chronological age (CA) and sex as independent variables, yielding the expected value of BA, E(BA|CA, sex); age acceleration was calculated. Where AgeAccel represents age acceleration; BA represents biological age; sex represents sex; CA represents actual age; and E(BA|CA,sex) represents the expected value of biological age BA predicted by a linear regression model based on actual age CA and sex.

[0052] Parallel computation of two versions of age acceleration: The sugar-containing version of age acceleration, when constructing a 1D aging axis Z-axis, the core indicator panel includes glycated hemoglobin; The deglycosylated version of age acceleration, when constructing a 1D aging axis Z, the core indicator panel does not include glycated hemoglobin.

[0053] It should be noted that the aging acceleration reflects the degree to which an individual's physiological aging deviates from the average level of their peers of the same age and sex; a positive value indicates accelerated aging, while a negative value indicates delayed aging. Two versions of parallel computing are provided to adapt to different application scenarios.

[0054] S3: Calculate the age acceleration of individuals in the type 2 diabetes application validation cohort, use a machine learning competitive risk model to predict the absolute risk probability of pancreatic cancer in the next 3-5 years, classify the risk level accordingly and calculate the age of pancreatic cancer and other risks; realize risk trajectory simulation based on the bio-age BA and machine learning competitive risk model.

[0055] Obtain the corresponding data of the core indicator panel for individuals in the type 2 diabetes application validation cohort and extract the semantic embedding features.

[0056] Based on core indicator panel, weight coefficient vector w, and bio-age calibration parameters and We calculated the age acceleration of individuals in the type 2 diabetes application validation cohort.

[0057] By using age acceleration, semantic embedding features, new-onset diabetes status, body mass index, and estimated glomerular filtration rate as input variables, the data are fed into a machine learning competitive risk model and executed to predict the absolute risk probability of pancreatic cancer in the next 3-5 years.

[0058] The endpoint was the first occurrence of pancreatic cancer (ICD-10 code C25), with death as a competing risk. The machine learning-based competing risk model was the Fine-Gray machine learning competing risk model; age acceleration was the core predictor variable, and it was included in the model along with new-onset diabetes status, body mass index, and estimated glomerular filtration rate as covariates. If the age acceleration used was a version calculated based on a panel of core indicators excluding glycated hemoglobin, then glycated hemoglobin level was included as a supplementary covariate. In this embodiment, forward stepwise regression (significance level α=0.05) was used to screen predictor variables from all candidate variables, and the variables ultimately included in the model were consistent with those described above.

[0059] The cumulative CIF form of the constructed Fine-Gray competitive risk model is as follows: Where Λ0(t) represents the baseline cumulative risk, U is the predictor variable vector, and θ is the corresponding coefficient vector. Taking the sugar-free version of the model as an example, by fitting the training set data, the hazard ratios (HRs) and coefficients of some predictor variables are obtained as follows: Age acceleration (per standard deviation increase): HR = 1.58, θ ≈ 0.46 Newly diagnosed diabetes status (yes / no): HR=2.35, θ≈0.85 Body Mass Index (for every 5 kg / m² increase) 2 ): HR=1.18, θ≈0.17 Estimate glomerular filtration rate (per 15 mL / min / 1.73 m) 2 ): HR=1.25, θ≈0.22 Glycated hemoglobin (for every 1% increase): HR = 1.21, θ ≈ 0.19 These coefficients and hazard ratios suggest that, after adjusting for other covariates, age acceleration and new-onset diabetes status are strong predictors of pancreatic cancer risk in individuals with type 2 diabetes.

[0060] Based on the preset risk classification algorithm, individuals are classified into low-risk, medium-risk, or high-risk levels according to their absolute risk probability.

[0061] Based on the clinical decision thresholds determined by decision curve analysis, the population is divided into three risk levels: Low risk: 5-year absolute risk probability <0.5% Medium risk: 5-year absolute risk probability 0.5%-2.0% High risk: 5-year absolute risk probability > 2.0% It should be noted that the Fine-Gray competing risk model can correctly handle the competing event of death, avoiding overestimation of the risk of developing pancreatic cancer.

[0062] Including newly diagnosed diabetes status helps to identify secondary diabetes that may be caused by early-stage pancreatic cancer. The risk level thresholds mentioned above were determined based on clinical practice consensus and supplemented by decision curve analysis, weighing the clinical costs of false positives and false negatives, and aiming to provide a clear basis for risk stratification management.

[0063] The risk age for conditions such as pancreatic cancer is calculated by matching an individual's current absolute risk probability with an age-risk distribution curve based on a healthy reference population cohort.

[0064] The process involves matching the current absolute risk probability of individuals with type 2 diabetes with the age-risk distribution curve based on a healthy reference population cohort to calculate and output the risk age for conditions such as pancreatic cancer.

[0065] Risk trajectory simulation is achieved based on a risk model that combines biological age analysis (BA) with machine learning. According to the preset intervention-indicator mapping rules, the simulated intervention value for an individual is converted into the simulated adjustment amount of the corresponding indicator in the core indicator panel. The value is then substituted into steps S32 and S33 to recalculate the age acceleration and absolute risk probability of pancreatic cancer after the intervention. This process is iterated to generate risk probability trajectory maps for multiple future time points.

[0066] S4: The validated machine learning competitive risk model, risk trajectory simulation algorithm and related parameters are packaged into a risk prediction toolkit containing executable scripts; the toolkit has an adaptive calibration function and outputs a risk assessment report.

[0067] A validation set was partitioned from the type 2 diabetes application validation cohort, and time extrapolation and geographic extrapolation validation procedures were executed to evaluate the generalization performance of the machine learning competitive risk model and risk trajectory simulation function.

[0068] The training and validation sets were randomly divided in a 7:3 ratio from the type 2 diabetes application validation cohort. Simultaneously, time and geographic extrapolation validation was performed using a completely independent US community atherosclerosis risk study cohort.

[0069] The verification results show: Internal validation set: C-index=0.81 ARIC external validation set: C-index=0.78 The slope of the calibration curve is 0.95. It should be noted that the C-index is an indicator for evaluating the model's discriminative ability, with a value ranging from 0.5 to 1.0. The closer it is to 1, the stronger the discriminative ability. The slope of the calibration curve evaluates the consistency between the predicted probability and the actual risk, with an ideal value of 1.

[0070] The validated machine learning competitive risk model and its parameters, semantic embedding feature extraction model parameters, risk trajectory simulation algorithm, and adaptive calibration function based on incremental learning strategy are encapsulated into a risk prediction toolkit.

[0071] The validated competitive risk model and its parameters are packaged into an engineerable risk prediction toolkit, specifically including: Parameter file: Weighting coefficients for the core indicator panel, biological age calibration parameters, and all coefficients of the Fine-Gray competitive risk model. Executable Script: The executable script is written in Python 3.8+ and relies on scientific computing libraries such as scikit-survival, pandas, and numpy. The main functional modules of the script include data preprocessing, biological age calculation, risk prediction, and results output. The visual risk scoring card generated by the results output module should include at least the following: Individual basic information: Anonymous ID, actual age, gender.

[0072] Key predictive results: Age at risk of pancreatic cancer, absolute risk probability over the next n years, and risk level. Clinical management recommendations: Automatically match and display preset follow-up or examination recommendations based on risk level. The toolkit includes definitions for core indicator panels, weight coefficient vectors, biological age calibration parameters, coefficients for machine learning competitive risk models, and executable scripts that can be loaded and executed by computing devices.

[0073] When deploying the risk prediction toolkit, age acceleration, which does not include glycated hemoglobin, is used for calculation first. The risk prediction toolkit supports receiving updated data from individuals after a six-month follow-up examination and re-estimates their pancreatic cancer risk probability based on changes in core indicators. The toolkit's adaptive calibration function, when deployed in new medical institutions, uses local healthy population sample data provided by the medical institutions to calibrate the linear regression model for calculating E(BA|CA,sex).

[0074] Deployment configuration: When deploying the toolkit, prioritize using the age acceleration method that does not contain glycated hemoglobin for cross-center applications.

[0075] It should be noted that the toolkit's encapsulation ensures the reproducibility and clinical applicability of the method. Prioritizing the use of the sugar-free version of age acceleration can effectively reduce bias caused by differences in testing methods between different medical institutions and improve consistency in cross-center applications. The toolkit's adaptive calibration function is implemented through an incremental learning strategy. When new local healthy population data becomes available, the parameters of the E(BA|CA,sex) linear regression model can be iteratively updated without retraining the entire model.

[0076] The executable script is configured to output a visual risk assessment report, which presents risk factors such as pancreatic cancer, including age, risk level, contribution of each input variable to the current risk prediction results, and management recommendations based on the risk trajectory simulation results.

[0077] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Therefore, all equivalent changes made to the structure, shape, and principle of this application should be included within the scope of protection of this application.

Claims

1. A method for predicting the risk of diabetes-related pancreatic cancer based on machine learning models and biological age, characterized in that, Includes the following steps: S1: Based on preset health screening criteria, construct a healthy reference population cohort and a type 2 diabetes application validation cohort; The baseline clinical physiological indicators and biomarker data of the healthy reference population cohort were standardized and preprocessed to obtain a standardized dataset; semantic embedding features of structured indicators and unstructured text were extracted from the electronic health records of the type 2 diabetes application validation cohort. S2: On the standardized dataset, using all-cause mortality, major adverse cardiovascular events, and serious renal events as joint supervised endpoints, a panel of core indicators was determined through an automated machine learning process; a machine learning regularized survival analysis model incorporating adversarial regularization was trained to obtain biological age (BA); S3: Calculate the age acceleration of individuals in the type 2 diabetes application validation cohort, use a machine learning competitive risk model to predict the absolute risk probability of pancreatic cancer in the next 3-5 years, classify the risk level accordingly, and calculate the age of pancreatic cancer and other risks. Risk trajectory simulation is achieved based on the aforementioned bio-age-based risk model that competes with machine learning; S4: The validated machine learning competitive risk model, risk trajectory simulation algorithm and related parameters are packaged into a risk prediction toolkit containing an executable script; the toolkit has an adaptive calibration function and outputs a risk assessment report.

2. The method for predicting the risk of diabetes-related pancreatic cancer based on machine learning models and biological age according to claim 1, characterized in that, Step S1 includes: S11: From a large-scale prospective cohort database, individuals who do not have diabetes, any malignant tumors, systemic inflammatory diseases, grade III-IV heart failure, grade B or higher liver failure, or stage 3 or higher chronic kidney disease at baseline are electronically screened to construct a healthy reference population cohort; and baseline clinical physiological indicators and biomarker data of the healthy reference population cohort are obtained. S12: From the aforementioned large-scale prospective cohort database, define a type 2 diabetes population that did not have pancreatic cancer at baseline as a type 2 diabetes application validation cohort; S13: Standardize and preprocess the baseline clinical physiological indicators and biomarker data of the healthy reference population cohort to obtain a standardized dataset; the preprocessing includes imputing missing values ​​using the random forest algorithm, Winsorize tail reduction processing for continuous variables, and transforming non-normally distributed variables to make their distribution normal. S14: Extract preset structured clinical indicators from the electronic health records associated with the type 2 diabetes application validation cohort; and extract unstructured clinical text data, and use a natural language processing model to extract semantic embedding features related to pancreatic cancer risk representation.

3. The method for predicting the risk of diabetes-related pancreatic cancer based on machine learning models and biological age according to claim 1, characterized in that, Step S2 includes: S21: Through a four-layer automated screening process, the core indicator panel for constructing biological age and the simplified sub-panels derived therefrom are determined from the indicators of the standardized dataset; the indicators of the core indicator panel are clustered into three levels according to their physiological significance, including the basic physiological layer reflecting the basic organ function, the metabolic regulation layer reflecting the state of glucose and lipid metabolism, and the cellular senescence layer reflecting the level of cell damage and repair. S22: On the standardized dataset, using all-cause mortality, major adverse cardiovascular events, and severe renal events as joint supervised endpoints, a Cox-Elastic Net model incorporating adversarial regularization terms is constructed and trained as a machine learning regularized survival analysis model. With the objective of minimizing the combination of the negative logarithmic partial likelihood sum of the joint supervised endpoints and the adversarial loss, the data matrix of the core indicator panel or simplified sub-panel is trained to obtain the weight coefficient vector w, and a 1-dimensional aging axis Z is calculated based on this vector. The calculation formula is as follows: Where X represents the data matrix of the core indicator panel or the simplified sub-panel; w represents the weight coefficient vector obtained by Cox-Elastic Net regularization through the fusion of adversarial regularization; S23: Based on the data from the healthy reference population cohort, fit the Gompertz mortality law between the one-dimensional aging axis Z and mortality risk to obtain calibration parameters. and The one-dimensional aging axis Z is calibrated as biological age BA, and the calculation formula is as follows: in, and The calibration parameters are obtained by fitting the Gompertz mortality law between the 1D aging axis Z and the risk of death. S24: Calculate age-related acceleration using the following formula: Where AgeAccel represents age acceleration; BA represents biological age; sex represents sex; CA represents actual age; and E(BA|CA,sex) represents the expected value of biological age BA predicted by a linear regression model based on actual age CA and sex.

4. The method for predicting the risk of diabetes-related pancreatic cancer based on machine learning models and biological age according to claim 3, characterized in that, The four-layer automated screening process in step S21 includes: S211: Conduct usability and stability screening, and remove indicators with a missing rate higher than 30% or an intragroup correlation coefficient lower than 0.6 from all initial indicators; S212: Perform age-related trend screening. For indicators that pass the screening in S211, retain those whose q-value after error detection rate correction is less than 0.05 and whose correlation with actual age is related to actual age. S213: Perform multi-endpoint outcome-oriented stable selection, guided by joint supervised endpoints, using a Cox-Elastic Net machine learning model with shared coefficients and combined with Bootstrap resampling technology to screen out indicators that are robustly related to disease risk and form a candidate indicator set; S214: Perform feature dimensionality reduction and level assignment determination. Sort the indicators according to the absolute value of the coefficients of each indicator in the candidate indicator set, extract the top K indicators, and classify each indicator into the basic physiological layer, metabolic regulation layer or cell senescence layer according to the division rules of the three levels, so as to form the core indicator panel and the simplified sub-panels with different numbers of indicators, where K is a preset positive integer.

5. The method for predicting the risk of diabetes-related pancreatic cancer based on machine learning models and biological age according to claim 3, characterized in that: In step S24, when calculating age acceleration, two versions are calculated in parallel: one is the age acceleration that includes glycated hemoglobin when constructing the core indicator panel; the other is the age acceleration that does not include glycated hemoglobin when constructing the core indicator panel.

6. The method for predicting the risk of diabetes-related pancreatic cancer based on machine learning models and biological age according to claim 1, characterized in that, Step S3 includes: S31: Obtain the corresponding data of the core indicator panel and the extracted semantic embedding features for individuals in the type 2 diabetes application validation cohort; S32: Based on the core indicator panel, weight coefficient vector w, and calibration parameters and Calculate the age acceleration of individuals in the type 2 diabetes application validation cohort; S33: The age acceleration, semantic embedding features, new-onset diabetes status, body mass index, and estimated glomerular filtration rate are used as input variables, input into a machine learning competitive risk model and executed to predict the absolute risk probability of pancreatic cancer in the next 3-5 years. S34: According to the preset risk classification algorithm, individuals are classified into low-risk, medium-risk, or high-risk levels based on the absolute risk probability; S35: Calculate the risk age for conditions such as pancreatic cancer by matching an individual’s current absolute risk probability with the age-risk distribution curve based on the health reference population cohort; S36: Based on the biological age BA and machine learning competition risk model, the risk trajectory is simulated; according to the preset intervention-indicator mapping rule, the simulated intervention value for the individual is converted into the simulated adjustment amount of the corresponding indicator in the core indicator panel, and substituted into steps S32 and S33 to recalculate the age acceleration and absolute risk probability of pancreatic cancer after intervention, and the risk probability trajectory map of multiple future time points is generated by iterating this process.

7. The method for predicting the risk of diabetes-related pancreatic cancer based on machine learning models and biological age according to claim 6, characterized in that: In step S33, the machine learning-based competitive risk model is the Fine-Gray machine learning competitive risk model; the age acceleration is used as the core predictor variable, and is included in the model together with the new-onset diabetes status, body mass index, and estimated glomerular filtration rate as covariates; if the age acceleration used is a version calculated based on the core indicator panel that does not include glycated hemoglobin, then the glycated hemoglobin level is included in the model as a supplementary covariate.

8. The method for predicting the risk of diabetes-related pancreatic cancer based on machine learning models and biological age according to claim 1, characterized in that, Step S4 includes: S41: Divide the validation set from the type 2 diabetes application validation cohort, execute time extrapolation and geographic extrapolation validation procedures, and evaluate the generalization performance of the machine learning competitive risk model and risk trajectory simulation function. S42: The validated machine learning competitive risk model and its parameters, semantic embedding feature extraction model parameters, risk trajectory simulation algorithm, and adaptive calibration function implementation algorithm based on incremental learning strategy are packaged into a risk prediction toolkit. The toolkit includes the definition of the core indicator panel, weight coefficient vector w, and biological age calibration parameters. and The system includes coefficients of a machine learning competitive risk model and an executable script that can be loaded and executed by a computing device; the executable script is configured to output a visualized risk assessment report, which presents risk age (e.g., pancreatic cancer), risk level, contribution of each input variable to the current risk prediction result, and management recommendations based on the risk trajectory simulation results.

9. The method for predicting the risk of diabetes-related pancreatic cancer based on machine learning models and biological age according to claim 8, characterized in that: When deploying the risk prediction toolkit, the age acceleration that does not include glycated hemoglobin is used preferentially for calculation; the risk prediction toolkit supports receiving updated data from individuals after a six-month follow-up examination, and re-estimates the probability of pancreatic cancer risk based on changes in core indicators; The toolkit's adaptive calibration function, when deployed in a new medical institution, uses local healthy population sample data provided by the medical institution to calibrate the linear regression model that calculates E(BA|CA,sex).

Images