Pancreatic cancer risk prediction system based on artificial intelligence

Abstract

The invention provides a pancreatic cancer risk prediction system based on artificial intelligence. The pancreatic cancer risk prediction system comprises a data acquisition module, a risk assessment module and an output module. The data acquisition module is used for at least acquiring the body mass index, fasting blood glucose, blood glucose control condition, symptom and age of the current to-be-evaluated person. And the risk assessment module is used for receiving the data and calculating a pancreatic cancer risk probability value of the current to-be-assessed person based on a pre-constructed pancreatic cancer risk prediction artificial intelligence model. And the output module is used for outputting the pancreatic cancer risk probability value of the current to-be-assessed person and the variable contribution graph based on the SHAP. According to the technical scheme, a large prospective queue is utilized, the random forest algorithm with the optimal performance is selected to construct the accurate pancreatic cancer risk prediction model with high interpretability, accurate identification of pancreatic cancer high-risk groups is realized, and the efficiency and accuracy of early screening are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a pancreatic cancer risk prediction system based on artificial intelligence. Background Art

[0002] Pancreatic cancer is the most malignant tumor in the digestive system. Its five-year survival rate has long stagnated at 13%. The global ratio of new cases to deaths is close to 1:0.89 each year, highlighting the urgency of early diagnosis (CA Cancer J Clin. 2024;74(1):12-49). Epidemiological data show that the five-year survival rate of stage IA patients after surgical resection can reach 83.7%, while the survival rate of stage IV patients drops sharply to less than 3.2% (J Natl Cancer Inst. 2020;112(11):1162-1169). This huge prognostic difference has shaped the early detection window into a core battlefield for improving clinical outcomes. However, pancreatic cancer screening faces a double dilemma: on the one hand, its annual incidence rate is only 13.5 / 100,000 and there is a lack of specific biomarkers, which leads to a high risk of false positives in traditional population screening strategies (JAMA. 2019;322(5):438-444); on the other hand, existing clinical screening guidelines mainly target genetic high-risk populations, which account for less than 10% of the total number of cases, while 90% of sporadic cases often miss the opportunity for early diagnosis due to the lack of effective risk assessment tools (Lancet Gastroenterol Hepatol. 2020;5(7):698-710; Gastroenterology. 2023;164(5):752-765).

[0003] Breakthroughs in artificial intelligence technology have provided new ideas for solving this dilemma. Risk prediction models based on medical data have shown significant advantages: the END-PAC model developed by the Mayo Clinic achieved an AUC value of 0.87 in a retrospective cohort by integrating blood sugar fluctuations, weight change trajectories and age parameters (Gastroenterology. 2018;155(3):730-739); Harvard University's CancerRiskNet system analyzed 9 million electronic health records (EHRs) and constructed a time series deep learning model with a 3-year risk prediction efficiency of 0.879 (Nat Med. 2023;29(5):1113-1122.); the University of Pennsylvania's THIN model focused on the newly diagnosed diabetes group, and the prediction system constructed by combining BMI dynamic changes, smoking exposure and metabolic indicators had an AUC of 0.82 (Gastroenterology. 2017;152(4):840-850). Although these breakthroughs have confirmed the feasibility of data-driven models, their limitations are also significant: the model construction is based on retrospective data of patients visiting medical institutions, which is subject to selection bias and spectrum bias (Gastroenterology. 2019;156(7):2024-2040.); the predictive factors rely on laboratory indicators collected in specialized medical settings (such as alkaline phosphatase and hemoglobin A1c), which are difficult to adapt to community primary health care scenarios (Gastroenterology. 2017;152(4):840-850); more importantly, there is a lack of verification of the model's screening efficacy for sporadic cases in real community settings.

[0004] Based on this, this study innovatively designed a community-oriented pancreatic cancer risk prediction model: through a nested case-control study of a prospective multicenter screening cohort, a pancreatic cancer risk prediction model based on machine learning was developed, and the model was externally validated using data from a prospective community natural population cohort. The model incorporates community accessible parameters and builds an online risk assessment platform (https: / / pc.chhospital.com.cn) to achieve two-way risk assessment functions on the community doctor side and the resident side. Compared with previous models, this study has triple breakthrough value: for the first time, the screening efficacy of the model for sporadic cases was verified in the community natural population; a lightweight assessment tool suitable for primary medical scenarios was developed; and the temporal bias of retrospective studies was avoided through a prospective design. The research results will provide evidence-based basis for community early screening of sporadic pancreatic cancer, promote the precise allocation of medical resources to high-risk groups, and lay a theoretical foundation for establishing a hierarchical diagnosis and treatment model of "community initial screening-hospital intensive examination". Summary of the invention

[0005] In order to overcome the above technical defects, the object of the present invention is to provide a pancreatic cancer risk prediction system based on artificial intelligence, which comprises:

[0006] A data collection module, used to obtain at least the body mass index, fasting blood sugar, blood sugar control status, symptoms and age of the person to be evaluated;

[0007] A risk assessment module is used to receive the above data and calculate the pancreatic cancer risk probability value of the current person to be assessed based on a pre-built pancreatic cancer risk prediction artificial intelligence model;

[0008] The output module is used to output the pancreatic cancer risk probability value of the current person to be evaluated and the variable contribution diagram based on SHAP.

[0009] Furthermore, the data acquisition module includes:

[0010] Data interface, used to receive body mass index, fasting blood sugar, blood sugar control status, symptoms and age input by users through the Web terminal or local terminal;

[0011] A data storage unit, used to store de-privacy residents’ health data, including basic personal information, medical history records, and on-site measurement data;

[0012] The outlier detection unit is used to identify and remove outliers that meet specific criteria using the box plot method.

[0013] Furthermore, the risk assessment module further includes:

[0014] The risk classification unit is used to compare the calculated risk probability value of the current person to be assessed with a preset threshold value and make a risk classification determination;

[0015] The sensitivity and specificity optimization unit is used to adjust the threshold of the model through cross-validation and Bootstrap method to balance the sensitivity and specificity of the model.

[0016] Furthermore, the preset threshold includes a first threshold and a second threshold that is smaller than the first threshold; if the risk probability value is ≥ the preset threshold, it is judged as a high risk; if the risk probability value is < the preset threshold, it is judged as a low risk.

[0017] Furthermore, the preset threshold is 0.51.

[0018] Furthermore, the method for constructing the pancreatic cancer risk prediction artificial intelligence model includes:

[0019] Step S1: A multicenter prospective nested case-control application design was used to collect electronic health information from the target population cohort data that met the inclusion criteria, and all data were divided into a training set and an external validation set according to a specific ratio;

[0020] Step S2: Use LASSO regression to screen key variables related to pancreatic cancer;

[0021] Step S3: Adaptive synthetic oversampling technology is used to preprocess the training set, and five machine learning algorithms, including random forest, XGBoost, naive Bayes, KNN and Logistic regression, are used for modeling, and the random forest algorithm with the best model performance is selected as the final model;

[0022] Step S4: Optimize model parameters through cross-validation and evaluate the AUC, accuracy, sensitivity and specificity of the model using an external validation set.

[0023] Furthermore, the output module is also used to: dynamically adjust the risk prediction probability according to the data input by the user, and update the risk classification results in real time; output inspection suggestions and health management guidance for different risk levels.

[0024] Furthermore, the output module includes:

[0025] A graphical user interface for real-time display of user-entered variable information, calculated pancreatic cancer risk probability values, and risk classification results;

[0026] The risk report generation unit is used to output an individualized risk prediction report, which includes:

[0027] Risk classification results, variable contribution diagram based on SHAP, and inspection recommendations and health management guidance for different risk levels.

[0028] Furthermore, further MRI or CT imaging examinations are recommended for the high-risk group, while only routine health examinations are required for the low-risk group.

[0029] Furthermore, the system is integrated with the community physical examination platform through a Web interface or a local terminal interface to automatically obtain residents' physical examination data, conduct real-time pancreatic cancer risk assessment, and output a screening list for the target population to assist in formulating a graded screening plan.

[0030] Compared with the prior art, the above technical solution has the following beneficial effects:

[0031] The technical solution of this application utilizes a large prospective cohort, integrates large-scale real-world data, and uses a variety of machine learning algorithms to develop models. Ultimately, the random forest algorithm with the best performance is selected to build an accurate and highly interpretable pancreatic cancer risk prediction model, thereby achieving accurate identification of high-risk groups for pancreatic cancer, improving the efficiency and accuracy of early screening, and improving patient prognosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It shows the flow of applicant participants in the queue;

[0033] Figure 2 For Lasso regression and feature engineering graphs;

[0034] Figure 3 A is a global analysis using the SHAP method, B is the SHAP contribution of asymptomatic pancreatic cancer patients, and C is the SHAP contribution of symptomatic low-risk people;

[0035] Figure 4 This is the interactive interface of the pancreatic cancer risk prediction system deployed on the web page. The side is the data input module, which is input through the drop-down menu, check, fill, etc. The right side is the output module, which outputs the results obtained by the internal calculation of the model. From top to bottom are the prediction result output box, input data display box and prediction model SHAP explanation box. DETAILED DESCRIPTION

[0036] The advantages of the present invention are further described below in conjunction with the accompanying drawings and specific embodiments. Those skilled in the art should understand that the following specific description is illustrative rather than restrictive, and should not be used to limit the scope of protection of the present invention.

[0037] This embodiment provides an artificial intelligence-based pancreatic cancer risk prediction system, which consists of a data acquisition module, a risk assessment module and an output module. The data acquisition module is responsible for collecting various types of health data related to pancreatic cancer, which will be passed to the risk assessment module. The risk assessment module processes the input data based on a pre-trained artificial intelligence prediction model, calculates the individual's pancreatic cancer risk probability, and classifies the risk. Finally, the output module displays the results of the risk assessment to the user in an intuitive and easy-to-understand manner, including risk classification information and visual charts of the contribution of key variables to the prediction.

[0038] The data acquisition module is used to obtain the body mass index, fasting blood sugar, blood sugar control status, symptoms and age of the person to be evaluated. The data acquisition module includes a data interface, a data storage unit and an abnormal value detection unit.

[0039] Preferably, the system is integrated with the community physical examination platform through a Web interface or a local terminal interface to automatically obtain residents' physical examination data, such as body mass index, fasting blood sugar, blood sugar control, symptoms and age, so as to conduct real-time pancreatic cancer risk assessment. Data interface: 1) On the Web side, a special user interaction interface is developed, and users can enter information such as body mass index, fasting blood sugar, blood sugar control, symptoms and age by filling out a form on the web page. The interface design is concise and clear, and detailed instructions and prompts are provided for each input item to ensure that users can enter data accurately. For example, for symptom options, a drop-down menu or multiple-choice box is provided to list common symptoms related to pancreatic cancer, such as abdominal pain, loss of appetite, jaundice, etc., and users can choose according to their actual situation. 2) On the local terminal, client software adapted to different operating systems (such as Windows, Mac, Linux) is developed. The client software establishes a secure connection with the system's data acquisition module. After the user enters the data on the local terminal, the data will be encrypted and transmitted to the system server. The local client also provides a friendly interface similar to the Web side to facilitate user operation.

[0040] The data storage unit is used to store de-privatized residents' health data, including basic personal information, medical history records, and on-site measurement data. The data storage unit establishes a safe and reliable database system to store de-privatized residents' health data. During the data storage process, personal basic information (such as name, ID number, etc.) is desensitized, and encryption algorithms such as hash functions are used to convert sensitive information into irreversible encrypted strings, retaining only the necessary identification information for data association and analysis. Medical history records are stored in the form of structured data, for example, information such as disease name, illness time, and treatment process are recorded in different fields to facilitate subsequent query and analysis. On-site measurement data (such as weight, height, fasting blood sugar measurement values, etc.) are stored in the order of measurement time, and relevant information of the measurement equipment is recorded to trace and evaluate the data quality. The database is backed up regularly to prevent data loss. The backup data is stored in a disaster recovery center in a different location to ensure that data can be quickly restored in the event of a natural disaster or system failure in the local area. At the same time, a strict data access rights management mechanism is established, and only authorized personnel (such as medical professionals, data analysts, etc.) can access specific residents' health data, and access operations are recorded in logs for auditing and tracking.

[0041] The outlier detection unit is used to identify and remove outliers that meet specific standards using the box plot method. Outlier detection uses the box plot method, that is, values ​​that exceed 1.5 times the interquartile range of the upper and lower quartiles are considered outliers, and data filling is performed after the outliers are deleted. In order to retain the original data information as much as possible, the median is used to fill missing data for measurement data. For categorical variables, the majority class is used to fill missing values. Measurement data are expressed as mean (standard deviation), and the independent sample t test is used for comparison between the two groups; measurement data that do not conform to the normal distribution are expressed as median (lower quartile, upper quartile), and the Wilcoxon rank sum test is used for comparison between groups. Count data are expressed as frequency and percentage, and the chi-square test is used for comparison between groups. When new data enters the system, the outlier detection unit first obtains the variable type corresponding to the data (such as measurement data or categorical variable). For measurement data (such as body mass index, fasting blood sugar, age, etc.), the upper and lower quartiles are calculated, and the range of 1.5 times the interquartile range is determined (refer to the standard calculation process of the box plot method in statistics). If a data point is outside this range, it is identified as an outlier. For categorical variables, the frequency of occurrence of each category is counted, and categories with extremely low frequencies (below a certain threshold, such as 5%) are considered outliers. Once an outlier is detected, the system will automatically process it. For outliers in measurement data, the median is used for filling. For example, if an individual's fasting blood glucose value is detected as an abnormally high value, the system will replace this outlier with the median of all data for this variable. For outliers in categorical variables, the majority class is used for filling. For example, if an individual's symptom option enters an extremely rare symptom (determined to be an outlier), the system will replace this outlier based on the majority class of the variable where the symptom is located in other normal samples. The processed valid data will be stored in the data storage unit for use by the subsequent risk assessment module.

[0042] The risk assessment module is used to receive the above data and calculate the pancreatic cancer risk probability value of the current person to be assessed based on a pre-built pancreatic cancer risk prediction artificial intelligence model.

[0043] This embodiment provides a method for constructing an artificial intelligence model for predicting pancreatic cancer risk, which includes steps S1 to S4:

[0044] Step S1: A multicenter prospective nested case-control application design was used to collect electronic health information from the target population cohort data that met the inclusion criteria, and all data were divided into a training set and an external validation set according to a specific ratio.

[0045] like Figure 1As shown in Figure 1, the flow of applicant participants in the cohort is shown. This embodiment is based on the primary health electronic records of annual physical examinations of 51,490 residents in 12 community hospitals, and prospectively collects pancreatic cancer risk-related information and informed consent of all participants. To ensure the integrity and standardization of multi-center data collection, cohort data collection follows a preset information list, and primary health information is electronically recorded through a unified data collection website. De-privacy clinical records are regularly transferred to the cohort database. Medical history and personal information are collected by on-site consultations by community doctors who have undergone standardized training, while fasting blood sugar, weight, height, etc. are obtained through on-site instrument measurements. Electronic records include basic demographic information, standardized medical history collection, and on-site measurement data. During the information collection process, the system automatically checks each piece of data to identify abnormalities, ensure the standardization of electronic records, and conduct quality assessments. Data that fail the quality assessment are not included in the independent database.

[0046] The target applicant population is set as residents aged 50-75 who signed informed consent for pancreatic cancer screening and participated in annual physical examinations in 12 community hospitals between 2020 and 2023. Exclusion criteria include: residents with incomplete identity information; residents diagnosed with pancreatic cancer or other malignancies in the past 5 years; residents who are unable to undergo MR / CT examinations due to various physical factors; residents with abnormal data records; and residents who refuse to accept follow-up.

[0047] A total of 51,490 patients from 12 centers were included in the prospective cohort application. Data from 9 centers were extracted as training sets, and data from 3 centers were used as external validation sets. The training set contained a total of 40,099 residents, of which 44 were diagnosed with pancreatic cancer. According to the ratio of case group: matched control group = 1:20, 880 subjects were extracted from the training set according to gender and enrollment time to form the control group. The cohort was divided into a training set (n=40,099, including 44 cases and 880 nested controls) and a test set (n=11,391, including 16 cases and 11,375 controls). See Table 1 below for details.

[0048] Table 1 Data queue statistics

[0049]

[0050] Step S2: Use LASSO regression to screen key variables related to pancreatic cancer.

[0051] After a comprehensive and systematic literature review, combing of postoperative care guidelines, multiple retrospective applications, and consensus meetings, this example identified 122 predictor variables (including measured physical signs, lifestyle and dietary habits, disease and drug history, etc.) from the literature. According to the clinical significance of the predictor variables and the target application scenarios of the model, 12 features were selected after scoring by frontline medical workers: family history (FDR), history of chronic pancreatitis (Chronic.pancreatitis), history of diabetes (Diabetes.mellitus), blood sugar control (Glycemic.control), symptoms (Symptom), height (Height), weight (Weight), waist circumference (Waist.circumference), fasting blood sugar (FBG), gender (Gender), age (Age), body mass index (BMI) for variable screening, questionnaire design and prospective information collection.

[0052] Before model training, Figure 2 As shown in the figure, using Lasso regression and feature engineering, 5 variable factors were finally selected from 12 features for model training: symptoms (Symptom), fasting blood glucose (FBG), age (Age), body mass index (BMI), and blood glucose control (Glycemic-control). Body mass index (BMI) and fasting blood glucose (FBG) were obtained through on-site measurements, age can be obtained from registered identity information, and symptoms and blood glucose control were recorded by clinicians in an electronic database through inquiries. Among them, no discomfort was recorded as (Symptom=0), and any symptom such as jaundice, abdominal discomfort, loss of appetite, and weight loss was recorded as (Symptom=1). Blood glucose control was registered according to the long-term self-tested blood glucose maintenance range reported by the subjects.

[0053] Step S3: Adopt adaptive synthetic oversampling technology to preprocess the training set, adopt at least two machine learning algorithms to build a model, and select the algorithm with the best model performance as the final model.

[0054] For example, this embodiment uses five machine learning algorithms, namely Random Forest, XGBoost, KNN, Naive Bayes, and Logistic Regression, for modeling. The performance comparison data of these five machine learning algorithms are shown in Table 2 below:

[0055] Table 2 Performance comparison of five machine learning algorithms

[0056] From the data comparison in Table 2, we can see that the optimal algorithm is the random forest model.

[0057] During the training process, in order to solve the problem of imbalanced outcome distribution and strong similarity between positive and negative samples, which may cause the model to have bias in the classification prediction of the majority category, the Adaptive Synthetic Sampling Approach for Imbalanced Learning (ADASYN) was combined to perform oversampling preprocessing on the minority category (pancreatic cancer patients) in the training set to increase the number of samples so that the model can better learn the characteristic patterns of pancreatic cancer patients. To ensure robustness and avoid overfitting, 50 repeated 5-fold cross validation and random grid search were used in the training set processed by the ADASYN sampling method for hyperparameter optimization.

[0058] To ensure the robustness and generalization ability of the prediction model, we used a 10-fold cross-validation method for internal validation. This method involves dividing the entire dataset into ten equal segments. In each iteration of the cross-validation process, nine segments are used as training sets to develop the model, while the remaining segments are used as validation sets to evaluate its performance. This process is repeated ten times, with each segment serving as a validation set exactly once, ensuring that every data point is used for both training and validation. The mean and standard deviation of these metrics for these ten folds are calculated to provide an estimate of the model performance and its reliability. The positive predictive value, accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC-ROC) were used to evaluate the predictive performance of the model.

[0059] Optimize model parameters through cross-validation (such as 10-fold cross-validation). Divide the data set into ten parts, select nine parts as training sets and one part as validation sets each time, and repeat training and validation ten times. Select the optimal model parameter combination based on the performance indicators (such as AUC, sensitivity, specificity, etc.) on the validation set.

[0060] Among them, the optimal parameters of the random forest model are: n_estimators (maximum number of iterations of weak learners) = 100, min_samples_split (minimum number of samples required for internal node re-division) = 2, min_samples_leaf (minimum number of samples for leaf nodes) = 1, max_depth (maximum depth of decision tree) = None.

[0061] Step S4: Optimize model parameters through cross-validation, and use an external validation set to evaluate the model's AUC, sensitivity, specificity, and positive predictive value.

[0062] After the training of the random forest model was completed, in order to verify the generalization ability of the model and comprehensively evaluate the performance of the model, data from 11,561 residents in three centers were extracted as independent external validation, and the trained model was predicted in an independent external validation cohort. The receiver operating characteristic curve was constructed, and the positive predictive value, sensitivity, specificity, and accuracy were calculated as the evaluation indicators of the model. The predictive performance of the prediction model in an independent data set was evaluated. Among them, the area under the receiver operating characteristic curve and the positive predictive value were used as the main indicators, and the 95% confidence interval of the AUC value was calculated by the Bootstrap method repeated 1000 times, and the ROC curve was drawn. The final model was evaluated for discrimination using confusion matrix and ROC curve. In the external validation set, the AUC of the random forest model was 0.74. The confusion matrix was constructed based on the prediction probability and threshold, the positive predictive value was calculated, and the sensitivity, specificity and accuracy were selected to evaluate the discrimination performance of the model. The statistical analysis used R4.2.2, all tests were two-sided, and P<0.05 was considered statistically significant. In the external validation set, the model showed a sensitivity of 50.00%, a specificity of 93.14%, an accuracy (combined accuracy of positive and negative outcomes) of 93.08%, and a positive predictive value of 1.04%. Among 11,107 individuals, the model predicted 768 high-risk individuals (accounting for 6.91% of the total population) and identified 8 PC patients in the high-risk population. This proved that the model has a good risk enrichment effect in an independent cohort, increasing the prevalence of the screened population to 1.04%, significantly enriching the risk population.

[0063] The risk assessment module further includes a risk classification unit and a sensitivity and specificity optimization unit. The risk classification unit is used to compare the calculated risk probability value of the current person to be assessed with the preset threshold and make a risk classification judgment. If the risk probability value is ≥ the preset threshold, it is judged as high risk; if the risk probability value is < the preset threshold, it is judged as low risk. The sensitivity and specificity optimization unit is used to adjust the threshold of the model through cross-validation and Bootstrap methods to balance the sensitivity and specificity of the model. After receiving the individual health data from the data acquisition module, the risk assessment module inputs it into the trained random forest prediction model to calculate the individual's pancreatic cancer risk probability. Then, the risk probability value is compared with the preset threshold. The preset threshold is determined during the model development process by analyzing the risk distribution of a large amount of sample data and combining clinical actual needs and cost-effectiveness considerations. For example, after analyzing the training set and validation set data, it is found that when the threshold is set to a certain value, the model can have a higher specificity and reduce false positive results while ensuring a certain sensitivity. In order to balance the sensitivity and specificity of the model, cross-validation and Bootstrap methods are used. During the cross-validation process, the threshold of the model is continuously adjusted to observe the changes in the sensitivity and specificity indicators of the model on the validation set under different thresholds. At the same time, the Bootstrap method is used to sample the data set with replacement to generate multiple new training sets and test sets, and the above threshold adjustment and performance evaluation process are repeated on these different data sets. Through repeated experiments, an optimal threshold is found so that the model achieves a good balance between sensitivity and specificity. Preferably, the preset threshold is 0.51. For example, in some cases, it may be necessary to appropriately increase the sensitivity to ensure that as many pancreatic cancer patients as possible are detected, but this may result in a slight decrease in specificity and increase some false positive results; in other cases, more emphasis may be placed on specificity to reduce unnecessary further examinations and waste of medical resources. According to the actual application scenario and needs, the most appropriate threshold setting is determined through this optimization method.

[0064] The output module is used to output the pancreatic cancer risk probability value of the current person to be evaluated and the variable contribution diagram based on SHAP. In order to solve the problem that the black box model is difficult to explain, this application uses the average SHAP value based on the SHAP method to explain the importance of the overall variable characteristics of the model. And use force diagrams for individual interpretation to analyze the contribution of variables in individual predictions. Model interpretability analysis (based on the SHAP method) is divided into two aspects:

[0065] (1) Global analysis: Global analysis is performed by applying the SHAP method to the trained random forest model ( Figure 3A). Calculate the average SHAP value of each variable, and sort the variables according to the size of the average SHAP value to determine the importance ranking of the variables. For example, after analysis, it was found that the average SHAP value of age was the largest, indicating that age contributed the most to the prediction of pancreatic cancer risk in the entire model, followed by variables such as body mass index, fasting blood sugar, blood sugar control and symptoms. This global analysis helps to understand the decision-making basis of the model as a whole, and provides an important reference for further optimizing the model and in-depth research on pancreatic cancer risk factors. This application uses the importance ranking of variables at the global level to understand the role of each variable in the screening model in predicting the risk of pancreatic cancer. The importance ranking of the variables is plotted in descending order on Figure 3 Among them, the order of importance is age, body mass index, fasting blood sugar, blood sugar control and discomfort symptoms.

[0066] (2) Individual analysis: SHAP is used to draw force diagrams for interpretation of the individual results. Figure 3 B is the SHAP contribution of asymptomatic pancreatic cancer patients, indicating that the model identifies them as high-risk people when their blood sugar control is poor and their body mass index is low. The importance of these individual variables is blood sugar control, low body mass index, and high blood sugar. Figure 3 C is the SHAP contribution of symptomatic low-risk people. The model assigns a higher risk to those with discomfort symptoms, but a higher body mass index level and normal fasting blood sugar and blood sugar control indicate a higher probability of low risk, making the model score low risk. For the prediction results of each individual, the SHAP method is used to draw a force diagram. In the force diagram, the horizontal axis represents the value of each variable, and the vertical axis represents the contribution of the variable to risk prediction (SHAP value). For example, for a specific individual, if his blood sugar control is poor, the point of the corresponding blood sugar control variable in the force diagram will be located at a lower position on the horizontal axis (indicating poor blood sugar control) and the vertical axis value will be larger (indicating a greater positive contribution to risk prediction, that is, an increased risk of cancer). In this way, the contribution of each variable in individual risk prediction is intuitively displayed, making the model results more transparent and easy to understand for individual users, and enhancing users' trust in risk assessment results.

[0067] The output module is also used to dynamically adjust the risk prediction probability according to the data input by the user and update the risk classification results in real time. When the user enters data through the data interface, the system monitors the data changes in real time. Whenever the user modifies or supplements a data, the data acquisition module immediately passes the new data to the risk assessment module. The risk assessment module recalculates the individual's pancreatic cancer risk probability based on the new data, and returns the updated risk probability value and classification results to the output module in real time. The graphical user interface of the output module dynamically updates and displays the risk probability and classification results (real-time data updates and dynamic interface display effects can be achieved through front-end technology), allowing users to see the impact of data changes on risk assessment in real time. For example, after the user enters data such as age or blood sugar control, the risk probability value and classification results on the interface will immediately change accordingly, helping users to intuitively understand the impact of different factors on their own pancreatic cancer risk, so that users can adjust their lifestyles or take other intervention measures in a timely manner.

[0068] The output module is also used to output inspection recommendations and health management guidance for different risk levels. For high-risk individuals, the report provides detailed inspection recommendations, such as recommending MRI or CT imaging examinations (refer to clinical diagnosis and treatment guidelines and recommendations for pancreatic cancer inspection methods in related studies), and provides information on nearby medical institutions that can conduct relevant inspections and appointment guidance (relevant information can be obtained by connecting to the medical institution information system), and explains the necessity of the inspection and the time interval recommendation. At the same time, health management guidance is given, including dietary adjustment recommendations (such as controlling sugar and fat intake, increasing dietary fiber intake, etc., refer to nutrition research and dietary recommendations for related diseases such as diabetes and obesity, because these factors are related to pancreatic cancer risk), weight management recommendations (such as formulating a reasonable weight loss or weight gain plan based on body mass index) and regular follow-up recommendations (such as recommending a review every 3 or 6 months, etc., combined with clinical follow-up practice and pancreatic cancer risk monitoring needs), to help high-risk individuals take positive measures to reduce the risk of pancreatic cancer. For the low-risk group, only routine health checks are required, such as a comprehensive physical examination once a year. The system can remind users to have routine physical examinations at appropriate times and pay attention to their health. In terms of dietary adjustment, specific diet plans are provided based on the individual's body mass index, blood sugar, and other conditions. For example, if the individual's body mass index is high and fasting blood sugar is high, it is recommended to reduce the intake of high-calorie, high-sugar foods, increase the intake of vegetables, fruits and whole grains, and control the calorie intake of each meal (refer to nutrition research and dietary recommendations for related diseases such as diabetes and obesity). In terms of weight management, a reasonable weight management plan is formulated based on the individual's current weight and target body mass index. For example, it is recommended to perform moderate aerobic exercise (such as brisk walking, jogging, etc.) and strength training every week, and the intensity and time of exercise are adjusted according to the individual's physical condition, and set stage-by-stage weight goals (refer to sports medicine research and health management practices). For regular follow-up recommendations, individuals should be clearly informed of the time, follow-up items, and follow-up location of the next follow-up, and follow-up reminder services should be provided (such as text message reminders or reminder notifications displayed on the system interface) to ensure that high-risk individuals can follow up on time and monitor changes in health status in a timely manner, so as to detect pancreatic cancer early or take effective intervention measures to reduce the risk of disease.

[0069] The output module includes: a graphical user interface and a risk report generation unit. The graphical user interface is used to display the variable information input by the user, the calculated pancreatic cancer risk probability value and the risk classification result in real time. The graphical interface of the present application is user-friendly. After the user enters the data in the data acquisition module, the interface displays the variable information input by the user in real time, including specific values ​​or options such as body mass index, fasting blood sugar, blood sugar control, symptoms and age. At the same time, the calculated pancreatic cancer risk probability is dynamically displayed on the interface, presented in the form of a percentage or an intuitive progress bar, so that users can quickly understand their risk level. When the risk assessment module completes the risk classification, the interface prominently displays the risk classification results (high risk or low risk), for example, distinguished by different color fonts or icons, high risk is displayed as a red warning icon, and low risk is displayed as a green safety icon, so that users can get key information at a glance. The risk report generation unit is used to output an individualized risk prediction report, which includes risk classification results, variable contribution diagrams based on SHAP, and inspection recommendations and health management guidance for different risk levels. The report first clearly gives the classification results of high risk or low risk, so that users can clearly know their risk status. Then, draw a visual chart of the contribution of each variable to risk prediction (such as Figure 3 ), using a similar visual approach to display variable contributions, such as using a bar chart or radar chart to intuitively display the relative importance of variables such as body mass index, fasting blood sugar, blood sugar control, symptoms, and age in individual risk prediction.

[0070] This system is integrated with the community physical examination platform through a Web interface or a local terminal interface. It is not only used to automatically obtain residents' physical examination data and conduct real-time pancreatic cancer risk assessment, but also to output a screening list for the target population and assist in formulating a graded screening plan. (1) Automatically obtain residents' physical examination data and conduct real-time pancreatic cancer risk assessment: Develop a data docking interface with the community physical examination platform (the interface design must comply with data security and medical information system interface specifications) to ensure the security and stability of the interface. Through this interface, the system can obtain relevant data generated by residents during the community physical examination in real time, including height, weight (used to calculate body mass index), fasting blood glucose measurement values, personal basic information (age, etc.) in the physical examination items, and symptom information and medical history information filled in the physical examination questionnaire (used to determine blood sugar control and other related factors). After the data is acquired, it is directly transmitted to the data acquisition module of this system. According to the above data acquisition and processing process, the risk assessment module performs real-time pancreatic cancer risk assessment. The assessment results are fed back to the community physical examination platform in real time. The physical examination doctor can directly view the residents' pancreatic cancer risk assessment results on the platform and provide residents with corresponding health advice or recommendations for further examinations based on the results. (2) Output a screening list for high-risk groups in the community to assist in developing a graded screening plan: The system regularly (such as monthly or quarterly) conducts statistical analysis on the pancreatic cancer risk assessment results of all residents in the community to screen out high-risk groups. Generate a screening list for high-risk groups in the community, which contains basic information of high-risk individuals (such as name, age, contact information, etc.), risk probability values, and a brief risk factor analysis. Provide the screening list to community health service centers or relevant medical institutions to assist them in developing a graded screening plan. For example, for key subjects in the high-risk population (those with extremely high risk probability values ​​or multiple high-risk factors), arrange for further detailed examinations and diagnoses as a priority; for other high-risk groups, arrange screening work in an orderly manner according to a certain time sequence and resource allocation principles, improve the efficiency and pertinence of community pancreatic cancer screening, and rationally allocate medical resources.

[0071] In order to make the trained prediction model easy to use, we built a web-based model calculator using the shiny package (version 1.2.1) of the programming language Python. In the graphical user interface (UI) section, Symptom is a multiple-choice option, where "None" means no symptom, which is mutually exclusive with the other options. Enter Age, FB GAs well as Weight and Height as model input. After receiving the input data, the model will convert the Symptom variable, select "Without Symptom" to convert Symptom=0, and automatically calculate the body mass index (BMI) based on Weight and Height for model prediction. The model outputs two results. According to the model prediction probability, those above the risk threshold are judged as high risk, and "Prediction result: High risk" is displayed. At the same time, a variable contribution graph based on SHAP is output as an interpretable visualization of the black box machine learning algorithm. Figure 4 The SHAP values ​​of variables that are judged as high risk (High risk) or low risk (Low risk) for symptomatic and asymptomatic people are shown as examples, which can guide the model contribution of each variable in clinical practice.

[0072] The system in this embodiment is applicable to the following three application scenarios:

[0073] 1. Used as an auxiliary diagnostic tool for pancreatic cancer in medical institutions to help doctors identify high-risk individuals

[0074] In medical institutions, doctors can guide patients to use this system. Patients input relevant health data through the Web terminal or local terminal device provided by the hospital (such as the hospital self-service terminal). After the system conducts risk assessment, doctors can refer to the risk classification results, risk prediction reports and visual variable contribution charts given by the system during the diagnosis and treatment process (doctors can view this information in the electronic medical record system or a dedicated system interface). This information can help doctors understand the patient's pancreatic cancer risk status more comprehensively and assist doctors in making diagnostic decisions. For example, if the system prompts that the patient is at high risk, the doctor can combine the patient's clinical manifestations and other examination results to arrange further examinations in a more targeted manner (such as pancreatic enhanced CT, tumor marker detection, etc., refer to the provisions of the pancreatic cancer diagnosis process in the clinical diagnosis and treatment guidelines), and improve the early diagnosis rate of pancreatic cancer. At the same time, doctors can develop personalized treatment and health management plans for patients based on the health advice provided by the system, such as adjusting treatment drugs based on the patient's blood sugar control, or guiding patients to exercise and control their diet reasonably according to weight management recommendations.

[0075] 2. Used for large-scale population screening in community physical examinations

[0076] In the annual physical examination or other health screening activities organized by the community, this system is integrated with the community physical examination platform to play a role. During the physical examination, the system automatically obtains physical examination data and conducts real-time pancreatic cancer risk assessment. For low-risk people, community doctors can give general health guidance in the physical examination report, encourage them to continue to maintain good living habits, such as a reasonable diet, moderate exercise, quitting smoking and limiting alcohol, and inform them of the importance of regular physical examinations. For high-risk people, community doctors will promptly notify them to conduct further examinations and assist them in making appointments for relevant examination items, such as recommending MRI or CT examinations at higher-level hospitals, and report the information of high-risk people to higher-level medical institutions or CDCs for follow-up management and intervention. Through this large-scale population screening method, potential pancreatic cancer patients can be detected early, the health level of community residents can be improved, and the incidence and mortality of pancreatic cancer can be reduced.

[0077] 3. Use it as a testing and optimization platform for pancreatic cancer risk prediction models in scientific research to support multi-center data analysis and model promotion

[0078] In the field of scientific research, multiple research centers can use this system as a unified pancreatic cancer risk prediction model testing and optimization platform. Each research center inputs the health data of community residents collected by itself into the system in the format and standards required by the system, and the system uses these multi-center data to evaluate the performance of the model. By comparing the performance of the model under the data of different centers, the generalization ability and stability of the model are analyzed, and factors that may affect the performance of the model (such as regional differences, population characteristics, etc.) are found. For example, the dietary habits of residents in the area where one research center is located may be mainly high in sugar and high in fat, while the diet of residents in the area where another research center is located is relatively light. By comparing the prediction results of the model under data from different regions, the impact of dietary habits on the pancreatic cancer risk prediction model can be studied. Researchers can further optimize the model based on these analysis results, such as adjusting variable weights, improving algorithm parameters, etc. At the same time, through the verification and optimization of multi-center data, the reliability and universality of the model are improved, providing a scientific basis for the promotion and application of the model on a larger scale. In addition, researchers can use the system's output results (such as risk classification results, variable contribution analysis, etc.) to conduct in-depth scientific research, explore the complex relationship between the pathogenesis of pancreatic cancer and risk factors, and provide new ideas and methods for the prevention and treatment of pancreatic cancer. For example, by analyzing the characteristics of a large number of high-risk individuals, it is possible to study whether there are new potential risk factors or biomarkers, which can provide directions for further basic and clinical research.

[0079] It should be noted that the embodiments of the present invention have better practicability and do not impose any form of limitation on the present invention. Any technician familiar with the field may use the technical content disclosed above to change or modify it into an equivalent effective embodiment. However, any modification or equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. The pancreatic cancer risk prediction system based on artificial intelligence is characterized by: include: A data collection module, used to obtain at least the body mass index, fasting blood sugar, blood sugar control status, symptoms and age of the person to be evaluated; A risk assessment module is used to receive the above data and calculate the pancreatic cancer risk probability value of the current person to be assessed based on a pre-built pancreatic cancer risk prediction artificial intelligence model; An output module, used to output the pancreatic cancer risk probability value of the current person to be evaluated and a variable contribution diagram based on SHAP; The method for constructing the pancreatic cancer risk prediction artificial intelligence model includes: Step S1: A multicenter prospective nested case-control application design was used to collect electronic health information from the target population cohort data that met the inclusion criteria, and all data were divided into a training set and an external validation set according to a specific ratio; Step S2: Use LASSO regression to screen key variables related to pancreatic cancer; Step S3: Adaptive synthetic oversampling technology is used to preprocess the training set, and five machine learning algorithms, including random forest, XGBoost, KNN, naive Bayes, and Logistic regression, are used for modeling, and the random forest model with the best model performance is selected as the final model; Step S4: Optimize model parameters by cross-validation and use an external validation set to evaluate the model's AUC, accuracy, sensitivity, and specificity; The optimal parameters of the random forest model are: maximum number of iterations of weak learners = 100, minimum number of samples required for internal node repartitioning = 2, minimum number of samples for leaf nodes = 1, and maximum depth of decision tree = None.

2. The pancreatic cancer risk prediction system based on artificial intelligence according to claim 1, characterized in that: The data acquisition module comprises: Data interface, used to receive body mass index, fasting blood sugar, blood sugar control status, symptoms and age input by users through the Web terminal or local terminal; A data storage unit, used to store de-privacy residents’ health data, including basic personal information, medical history records, and on-site measurement data; The outlier detection unit is used to identify and remove outliers that meet specific criteria using the box plot method.

3. The pancreatic cancer risk prediction system based on artificial intelligence according to claim 1, characterized in that: The risk assessment module further comprises: The risk classification unit is used to compare the calculated risk probability value of the current person to be assessed with a preset threshold value and make a risk classification determination; The sensitivity and specificity optimization unit is used to adjust the threshold of the model through cross-validation and Bootstrap method to balance the sensitivity and specificity of the model.

4. The pancreatic cancer risk prediction system based on artificial intelligence according to claim 3, characterized in that: If the risk probability value is ≥ the preset threshold, it is judged as high risk; if the risk probability value is < the preset threshold, it is judged as low risk.

5. The artificial intelligence-based pancreatic cancer risk prediction system according to claim 4, characterized in that: The preset threshold is 0.

51.

6. The pancreatic cancer risk prediction system based on artificial intelligence according to claim 1, characterized in that: The output module is also used to: dynamically adjust the risk prediction probability according to the data input by the user, and update the risk classification results in real time; output inspection suggestions and health management guidance for different risk levels.

7. The artificial intelligence-based pancreatic cancer risk prediction system according to claim 6, characterized in that: The output module comprises: A graphical user interface for real-time display of user-entered variable information, calculated pancreatic cancer risk probability values, and risk classification results; The risk report generation unit is used to output an individualized risk prediction report, which includes: Risk classification results, variable contribution diagram based on SHAP, and inspection recommendations and health management guidance for different risk levels.

8. The artificial intelligence-based pancreatic cancer risk prediction system according to claim 7, characterized in that: Further MRI or CT imaging examinations are recommended for the high-risk group, while only routine health examinations are required for the low-risk group.

9. The pancreatic cancer risk prediction system based on artificial intelligence according to any one of claims 1 to 8, characterized in that: The system is integrated with the community physical examination platform through a Web interface or a local terminal interface to automatically obtain residents' physical examination data, conduct real-time pancreatic cancer risk assessment, and output a screening list for the target population to assist in formulating a graded screening plan.

Images