An artificial intelligence system for predicting prostate cancer risk based on multiple routine physical examination indicators

Abstract

The present invention provides an artificial intelligence system for predicting the risk of prostate cancer based on a variety of routine physical examination indicators, which includes: a data acquisition module, a data cleaning module, a risk assessment module, and a report generation and display module. The data acquisition module is used to directly retrieve routine physical examination indicators from a hospital information system. The data cleaning module is used to clean and preprocess the data. The risk assessment module is used to perform feature selection using a random forest feature selection algorithm, and to construct a prediction model based on the selected features and using the prediction classification method of XGBoost. The report generation and display module is used to generate different reports according to different thresholds, and to display the evaluation results in graphical and tabular form. The input of the prediction model of the present application is the patient's routine physical examination data, which can save costs, and the prediction model adopts a combination of the most effective feature selection algorithm and classification method, which improves the prediction accuracy of prostate cancer risk.

Description

Technical Field

[0001] The present invention relates to a prostate cancer risk prediction system based on multiple clinical routine physical examination indicators, and belongs to the field of tumor diagnosis and tumor risk prediction in the medical field. Background Art

[0002] Prostate cancer (PCa) is a malignant tumor that mainly affects older men. The incidence of prostate cancer is high, and prostate cancer diagnosis is a major medical challenge facing the world. According to the report "Cancer Statistics, 2018" in the United States, it is estimated that in 2018, there will be 164,690 new cases of prostate cancer and 29,430 deaths in the United States. Among male malignant tumors, its incidence and mortality rate rank first and second respectively. In China, with the aggravation of aging, the incidence of prostate cancer in Chinese men is rapidly increasing. Between 2000 and 2016, its annual growth rate was 13%.

[0003] Although prostate cancer has good therapeutic effects in early treatment and is expected to be cured, early diagnosis is extremely challenging because early prostate cancer often has no obvious symptoms. According to statistics from the National Cancer Center (2016), more than 70% of prostate cancer patients in China are already in the late stage when diagnosed, with poor treatment effects, easy metastasis, and high mortality. At the same time, there is regional imbalance in China's economic development. Many regions have not popularized systematic prostate cancer screening, diagnosis and detection methods, resulting in a large number of cases that cannot be discovered and diagnosed in time, further increasing the proportion of advanced patients and causing serious public health problems. Therefore, the current diagnosis of prostate cancer is a major medical challenge facing my country.

[0004] At present, the diagnosis of prostate cancer mainly relies on the detection of serum PSA, but an increase in PSA does not necessarily mean that the patient has prostate cancer. A considerable number of people with an increased PSA do not actually have prostate cancer. Patients with an increased PSA need to undergo puncture and pathological biopsy to confirm prostate cancer. In current medical practice, a considerable number of men who undergo prostate puncture in my country are not diagnosed with prostate cancer. This phenomenon is mainly due to the poor specificity of prostate-specific antigen (PSA) testing. The positive rate of prostate puncture in China is only 33%, which means that 67% of men who undergo prostate puncture do not find prostate cancer. At the same time, prostate puncture is an interventional operation accompanied by pain and complications. It requires the ultrasound probe to be inserted into the patient's rectum and punctured continuously for 12 or more needles under ultrasound guidance. Continuous punctures bring very uncomfortable feelings to patients, and may also be accompanied by certain risks of postoperative infection and bleeding after puncture. In severe cases, patients may suffer from septic shock and hemorrhagic shock, which may affect their life, health and quality of life.

[0005] PSA specificity leads to excessive puncture biopsy, which not only increases the economic and psychological burden on patients, but also has caused widespread doubts and criticisms in recent years.

[0006] (1) Application of multi-parameter magnetic resonance imaging (mpMRI) in the diagnosis of prostate cancer. The popularity of multi-parameter magnetic resonance imaging (mpMRI) is limited by the high cost of the equipment and the need for highly trained doctors to interpret the images. The main reason is that multi-parameter magnetic resonance equipment is difficult to popularize. It is only equipped in medium and large cities and large hospitals. At the same time, the interpretation of multi-parameter magnetic resonance also requires very professional personnel, and there is a problem that it cannot be popularized and promoted in a short period of time. At the same time, there is also a problem of resource shortage for multi-parameter magnetic resonance equipment in large cities. Therefore, although it has gradually been confirmed to be a highly sensitive and specific prostate cancer detection tool in the diagnosis of prostate cancer, it can simultaneously use multiple MRI parameters (such as T1-weighted imaging, T2-weighted imaging, diffusion-weighted imaging, etc.) to provide detailed information about the location, size and invasiveness of prostate cancer, and has significant advantages in diagnosis, its application in clinical practice is still limited. The current clinical diagnosis and treatment load has already put a great burden on related medical resources. If a large number of prostate cancer diagnosis and screening tasks are added, it will lead to a shortage of resources for the diagnosis and treatment of a large number of patients with other diseases.

[0007] (2) Application of molecular markers in the diagnosis of prostate cancer. Molecular markers such as PCA3 and p2PSA have also been used in the diagnosis of prostate cancer. Although these markers help to improve the detection rate of prostate cancer, their use in China has not been approved due to their high price and unsatisfactory results. At the same time, a large number of domestic studies have confirmed that there are huge racial differences in prostate cancer between China and Europe and the United States. These differences will significantly limit the application of molecular diagnostic markers given gene changes discovered abroad in China. In summary, these factors limit the widespread use of molecular diagnostic markers in the diagnosis of prostate cancer.

[0008] In summary, an elevated PSA does not necessarily mean that one has prostate cancer. Therefore, the strategy of performing prostate puncture biopsy based on elevated PSA has certain diagnostic defects, resulting in a considerable number of unnecessary punctures in clinical practice, which has a significant impact on people's health and quality of life.

[0009] New methods for detecting prostate cancer are constantly emerging, but they all have some shortcomings. At the same time, due to regional imbalances in China's economic development, the high cost of some methods makes it unaffordable for patients and the government. In order to improve the predictive effect of prostate cancer screening without increasing the financial burden on patients and reducing the popularity of screening, researchers are working hard to improve and innovate. Summary of the invention

[0010] In order to overcome the above technical defects, the purpose of the present invention is to provide an artificial intelligence system for predicting the risk of prostate cancer based on multiple conventional physical examination indicators, which includes:

[0011] A data acquisition module, which is used to automatically obtain the patient's routine physical examination indicators most closely related to prostate cancer risk through interface integration with the hospital information system;

[0012] A data cleaning module, the data cleaning module includes a data quality check submodule and a missing value filling submodule, the data quality check submodule is used to automatically mark incomplete or erroneous data, and the missing value filling submodule is used to handle the problem of missing data, using the pattern of similar patients to fill in the qualitative data, and using the average value of similar patients to fill in the quantitative data;

[0013] A risk assessment module, wherein the risk assessment module includes a feature selection submodule and a model building submodule, wherein the feature selection submodule is used to perform feature selection using a random forest feature selection algorithm, and the model building submodule is used to build a prediction model based on the selected features and using an XGBoost prediction classification method;

[0014] The report generation and display module includes a report generation submodule and a display submodule. The report generation submodule is used to generate different reports according to different thresholds. The report contains ROC curves, calibration graphs and decision curve analysis results. The display submodule is used to display the evaluation results in graphical and tabular form.

[0015] Furthermore, the artificial intelligence system for predicting the risk of prostate cancer based on multiple conventional physical examination indicators further includes one or more of a system management and security module, a tracking and follow-up and popular science module, and a data storage module;

[0016] The system management and security module includes a system management submodule, a data input submodule and a security submodule. The system management submodule is used to allow an administrator to manage users and system settings. The data input submodule allows manual input or upload of data. The security submodule is used to perform data encryption, access control, audit logs and security updates on sensitive medical data.

[0017] The tracking and follow-up and popular science module includes a tracking and follow-up submodule and a popular science education submodule, wherein the tracking and follow-up submodule is used to formulate a personalized follow-up plan according to the risk assessment results, and the popular science education submodule is used to provide patients with knowledge and preventive measures about prostate cancer;

[0018] The data storage module is used to store the collected raw data in a relational database, equipped with a backup mechanism to prevent data loss, and to safely store the processed data in a NoSQL database optimized for big data.

[0019] Furthermore, the data storage module further includes a version control submodule, and the version control submodule is used to ensure that each data change can be traced back.

[0020] Furthermore, the conventional physical examination indicators most closely related to the risk of prostate cancer include age, urine specific gravity, and ultrasound tumor area (the product of the left and right diameters and the upper and lower diameters of the tumor (cm 2 )), serum alkaline phosphatase, serum albumin, free PSA (fPSA), prostate specific antigen (PSA) value, fPSA / PSA, blood glucose, lymphocyte percentage, blood potassium, blood sodium, platelet distribution width, hematocrit, γ-glutamyl transpeptidase, blood lymphocyte count, blood neutrophil count and mean corpuscular hemoglobin, a total of 18 indicators.

[0021] Furthermore, the data acquisition module further includes an optical character recognition submodule, which is used to capture data from handwritten or printed paper medical records.

[0022] Furthermore, the data acquisition module further includes a natural language processing submodule, and the natural language processing submodule is used to parse medical records and extract relevant clinical indicators.

[0023] Furthermore, the risk assessment module further includes a model optimization submodule and a risk indication submodule, the model optimization submodule is used to perform hyperparameter tuning to improve prediction accuracy, and the risk indication submodule is used to display the risk level using color-coded indicators.

[0024] Furthermore, the method for performing feature selection using the random forest feature selection algorithm in the feature selection submodule includes the following steps:

[0025] First, use the random forest algorithm to extract multimodal information from the data to obtain the importance of each feature.

[0026] In the random forest algorithm, for each decision tree, the corresponding out-of-bag data is selected to calculate the out-of-bag data error. Its basic form is shown in the following formula:

[0027]

[0028] In formula (1), N is the total number of samples, m is the number of misclassified samples,

[0029] Randomly add noise interference to the features of all samples of the out-of-bag data OBB, and calculate the out-of-bag data error again, denoted as err OBB 2. Assuming there are N trees in the forest, the importance of the feature is:

[0030] ∑(err OBB 1-err OBB 2) / N (2)

[0031] Calculate the importance of each feature according to formula (2);

[0032] Then, the features with high importance are screened out to form the modeling feature set, while reducing the data dimension.

[0033] Furthermore, the method for constructing a prediction model based on the selected features and using the prediction classification method of XGBoost in the model building submodule includes the following steps:

[0034] First, based on the modeling feature set, the XGBoost algorithm is used to build the initial classification model of multimodal fusion, and the model is trained in the classifier. A new tree is added in each iteration. This tree is generated by efficiently searching in the feature space and selecting the optimal split point. After each new tree is added, the weight of each leaf node is adjusted according to the gradient of the objective function and the information of the Hessian matrix, so that each new tree attempts to correct the residual left by the previous tree to minimize the prediction error. The cross-validation strategy is used to evaluate and optimize the performance of the trained model to obtain the optimal prediction model.

[0035] Then, when the trained prediction model classifies a case, it will fall into a corresponding leaf node in each tree according to the characteristics of the case. Each leaf node corresponds to a score. The prediction scores corresponding to each tree are added together to obtain the disease risk value of the case.

[0036] In XGBoost, the objective function is simplified to:

[0037]

[0038] In formula (3), G j represents the sum of the first-order derivatives of all samples, H j represents the sum of the second-order derivatives of all samples, γ represents the adjustment function for this item, T represents the number of leaf nodes in a tree, and f t Represents the predicted output of the model for the input sample after the tth iteration.

[0039] Compared with the prior art, the above technical solution has the following beneficial effects:

[0040] 1. The technical solution of this application can give full play to the advantages of artificial intelligence in risk prediction. Models containing more modalities and parameters and using multiple algorithms should be attempted to be established to improve the accuracy of current risk prediction models. For prostate cancer risk prediction, multimodal data may include clinical data, imaging data, biomarkers, genetic data, etc. Each modality can provide different information, and the combination of this information can help us understand the patient's condition more comprehensively. In addition, multi-parameter models can better capture the complexity of the disease and provide more accurate predictions. For example, in clinical data, multiple parameters such as age, gender, race, body mass index, family history, and lifestyle habits should be considered as much as possible. In addition, the use of multiple algorithms is also an important strategy for optimizing risk prediction models. Different algorithms have their specific advantages and limitations, and combining multiple algorithms can improve the stability and accuracy of the model. Combining multiple machine learning algorithms may achieve better predictive performance. Specifically, this patent introduces a multi-modal, multi-parameter and multi-algorithm strategy into the prostate cancer prediction model to use machine learning to establish a new risk prediction model based on physical examination data to improve the prediction effect, reduce overdiagnosis and treatment, and at the same time reduce the economic burden on patients and popularize screening as much as possible.

[0041] 2. In response to the problem that new methods for diagnosing prostate cancer are expensive and difficult to promote, we proposed a prostate cancer prediction model based on routine physical examination data. The input of this model is the patient's routine physical examination data, which means that its application does not require additional equipment or examinations. These data can be easily obtained in grassroots hospitals and will not bring additional burdens to patients, which also makes our model highly feasible and universal. At the same time, routine physical examination data can provide enough information. The model can predict whether a patient is likely to have prostate cancer by learning the relationship between these features and prostate cancer. Moreover, the model can be easily downloaded and applied online, which greatly reduces the cost of prostate cancer diagnosis and enables the new diagnostic method to be widely promoted.

[0042] 3. In response to the problem of low positive rate of prostate biopsy, this patent proposes a pre-screening model based on machine learning. The training data set of this model contains rich clinical information, such as PSA level, age, weight, height, prostate volume, etc., as well as labels of whether it has been diagnosed with prostate cancer. A machine learning model was trained using this data set. By learning the complex patterns of these features, the model can predict whether patients are at risk of prostate cancer based on their clinical information. In practice, we can use this model to conduct a risk assessment for every patient who undergoes a prostate examination, and then only perform a prostate biopsy on those patients who are predicted by the model to be at high risk. In this way, unnecessary prostate biopsies can be reduced, thereby reducing clinical burden and excessive puncture biopsies.

[0043] 4. Regarding the impact of racial differences on prostate cancer diagnosis, a prostate cancer prediction model based on Asian data is proposed. This model uses a large amount of clinical data of Asians for training, which can better adapt to the characteristics of Asians, thereby improving the prediction accuracy in Asian populations. By cooperating with medical institutions in various countries, we have collected a wealth of clinical data on Asians, which greatly improves the prediction performance of the model in Asian populations. In this way, we can overcome the challenges of racial differences and improve the diagnostic accuracy of prostate cancer.

[0044] 5. In the construction of the prediction model for prostate cancer detection, there are several important technical issues that need to be addressed. First, the processing of big data requires high costs. While maintaining the quality of screening, we need to effectively process a large amount of clinical indicator data to avoid the high cost of processing complex data. Second, current prostate cancer prediction models and traditional regression methods often ignore a large number of nonlinear features in the data set. In addition, in clinical practice, a large amount of missing data may have an adverse effect on the accuracy of the prediction model. Finally, due to the large number of possible combinations of feature screening and prediction methods, we need an effective method to select the best among these possible combinations. To solve these technical problems, we proposed the following solutions: First, we used data dimensionality reduction technology to reduce data complexity while retaining key information, so as to effectively process a large amount of clinical indicator data. To address the problem of ignoring nonlinear features, we used the xgboost algorithm, which is an efficient machine learning algorithm that is particularly suitable for processing a large number of research variables and nonlinear problems, thereby improving the accuracy of prostate cancer prediction. For the problem of missing clinical data, we used data cleaning and processing methods to convert medical symbols into digital symbols, delete cases and features with incomplete information, and men with high PSA levels, and fill in the missing data appropriately to ensure that our model can be trained on the complete data set. Finally, in order to solve the difficulties in feature selection and prediction method selection, we proposed a full-dimensional feature-classification method combination strategy. Through this strategy, we can comprehensively evaluate the prediction effect of each combination, avoid blindly selecting combinations, and find the best-performing feature selection algorithm and classification method combination.

[0045] This model uses advanced machine learning technology to improve the accuracy of predicting prostate cancer risk. In practice, it can correctly identify 84.50% of non-prostate cancer cases while ensuring that 90.45% of prostate cancer cases are correctly identified. Such results mean that the model has high accuracy and reliability. This result exceeds the performance of other related studies. In clinical practice, the application of the model can also bring significant benefits. The DCA curve (decision curve analysis, a tool to evaluate whether the prediction model has actual benefits for patients) shows that our model can substantially improve the clinical prognosis of patients and improve the benefits of patients.

[0046] 6. At present, the value of conventional physical examination indicators used alone in clinical diagnosis is relatively limited and cannot effectively predict the risk of prostate cancer. PSA is a typical example. PSA screening has been widely used for early detection of prostate cancer, but its specificity needs to be improved. This means that although PSA screening can detect most cases of prostate cancer, it will also misdiagnose some non-cancer cases as cancer cases, which will bring additional burdens to patients and the medical system. However, our algorithm gives more value to the original single indicator by reasonably combining these conventional physical examination indicators and combining the computing power of artificial intelligence. Therefore, this model significantly improves the accuracy of the strategy of screening prostate cancer based on PSA. The model of this patent greatly improves the specificity of PSA screening and reduces the misdiagnosis rate at this point, making prostate cancer screening more accurate and more in line with the needs of personalized medicine. It is not only more specific than PSA, the most commonly used tumor marker for prostate cancer, but also avoids bringing additional burdens to patients, further improving patient compliance. This not only improves the operational feasibility of diagnosis and screening, but also substantially improves the clinical prognosis of patients. Therefore, this algorithm has far-reaching practical significance and application value in clinical practice.

[0047] 7. This model can be provided as an online self-service model or as a real-time non-intervention model based on HIS, which is convenient and fast.

[0048] 8. The model takes into account multiple factors including patient age, family history, prostate-specific antigen (PSA) value, digital rectal examination (DRE) results, prostate volume, free PSA, PSA density, PSA rate, etc., all of which have been shown to play an important role in the early diagnosis and prediction of prostate cancer.

[0049] 9. The model uses a variety of classifiers and feature selection methods, has high flexibility and adaptability, and can be applied to different data sets and prediction tasks.

[0050] 10. The ROC curve, calibration plot and decision curve analysis results of the APCAI score of this model in predicting PCa all show that its predictive ability is very good and can help clinicians make more accurate diagnosis and treatment decisions.

[0051] 11. The model was developed by comparing multiple classification methods and feature selection methods. The best combination of feature selection algorithm and classification algorithm was determined to maximize the predictive ability of the model.

[0052] 12. The results of internal and external validation by multiple independent cohorts showed that the model has good robustness and generalization ability.

[0053] 13. The model also considers different thresholds to meet the needs of different patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a module structure diagram of the artificial intelligence system for predicting prostate cancer risk based on multiple conventional physical examination indicators of this application;

[0055] Figure 2 A flowchart for predicting the risk of prostate cancer using the artificial intelligence system of the present application for predicting the risk of prostate cancer based on multiple conventional physical examination indicators;

[0056] Figure 3 A modeling flowchart of an artificial intelligence system for predicting prostate cancer risk based on multiple conventional physical examination indicators in this application;

[0057] Figure 4 The receiver operating characteristic curve (ROC) of the artificial intelligence system for predicting prostate cancer risk based on multiple conventional physical examination indicators of this application;

[0058] Figure 5 A calibration curve for the artificial intelligence system for predicting prostate cancer risk based on multiple conventional physical examination indicators in this application;

[0059] Figure 6 This is the decision curve of the artificial intelligence system for predicting prostate cancer risk based on multiple routine physical examination indicators in this application. DETAILED DESCRIPTION

[0060] 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.

[0061] like Figure 1As shown, this embodiment provides an artificial intelligence system for predicting prostate cancer risk based on multiple routine physical examination indicators, which includes: a data acquisition module, a data cleaning module, a risk assessment module, a report generation and display module, a system management and security module, a tracking and follow-up and popular science module, and a data storage module.

[0062] like Figure 2 As shown, the method for predicting the risk of prostate cancer using the above artificial intelligence system includes the following steps:

[0063] Step S1: Pre-configure and deploy related resources

[0064] Before formally applying the artificial intelligence system to predict prostate cancer risk, the following configuration and deployment should be performed in advance:

[0065] 1. Hardware and software resource configuration, establishment of hospital information system (HIS) interface:

[0066] First, technicians need to configure the necessary hardware and software resources. This includes a high-performance server that can run machine learning algorithms, necessary database management systems (such as MySQL, MongoDB, etc.), OCR and natural language processing toolkits, and system development and management tools. Next, establish an interface with the hospital information system (HIS), which may require cooperation with the HIS vendor to obtain the required interface documents and permissions. Technicians need to be familiar with the data structure and query methods of the HIS to ensure that the required clinical data can be correctly obtained.

[0067] 2. Configure the data acquisition module:

[0068] After establishing the HIS interface, you need to configure the data collection module, set the type of data and frequency to be obtained, and manually set the path. Write a query script to extract data from the HIS regularly.

[0069] 3. Set up the data storage module:

[0070] Select an appropriate database management system, establish a database, define the structure of the data table, and implement data import and export functions. At the same time, a regular backup mechanism needs to be set up to prevent data loss.

[0071] 4.Configure data cleaning module:

[0072] Write R language-based cleaning and filling codes to implement functions such as data quality check and missing value filling to determine the completeness and correctness of the data and handle missing values.

[0073] 5. Configure the risk assessment module:

[0074] The risk assessment module is the core of the system, which requires the selection of appropriate feature selection and machine learning algorithms to train a model that can accurately predict the risk of prostate cancer.

[0075] 6.Configure multi-platform report generation and display module:

[0076] After all data processing and analysis functions are completed, the report generation and display functions are realized, and the format and content of the report are designed and established on multiple platforms (WeChat official accounts, mini-programs, web pages and APPs), to achieve graphical display of data and report export and printing function interfaces.

[0077] 7. Set up a multi-platform tracking and follow-up module:

[0078] Write regular push code, and set different follow-up frequencies and contents for patients with different risk levels based on the results of the risk assessment module. Regularly push follow-up reminders and health education reminders to patients' accounts based on risk assessment results on multiple platforms.

[0079] 8. Configure system management and security modules:

[0080] In order to protect the security of the system and data, it is necessary to write programs to configure the system management and security modules.

[0081] 9. Testing, Validation, Training and Deployment:

[0082] After the system is configured, extensive testing is required to verify the functionality and performance of the system. This includes functional testing, performance testing, and security testing. Feedback should be collected and adjustments should be made as needed. Once the system passes testing, deployment can begin. Training needs to be provided to hospital staff so they understand how to use the system. Technical support should also be provided to handle any issues that may arise.

[0083] Step S2: Automatically obtain the patient's routine physical examination data most closely related to prostate cancer risk through interface integration with the hospital information system

[0084] The data acquisition module of the present invention is intended to realize automatic and efficient collection of clinical data related to prostate cancer. The module is integrated with the interface of the hospital information system (HIS) to automatically obtain the patient's clinical data, including but not limited to 18 indicators such as age, family history, prostate-specific antigen (PSA) value, GGT, prostate volume, free PSA, fPSA / tPSA. For example, the hospital information system (HIS) collects all physical examination clinical data of patients who undergo prostate biopsy in 8 hospitals (Shanghai tertiary hospitals, Guangzhou tertiary hospitals, Suzhou tertiary hospitals, Sichuan tertiary hospitals, Xi'an tertiary hospitals, Korean teaching hospitals, Hong Kong teaching hospitals and Malaysian teaching hospitals), including age, abdominal ultrasound tumor size, PSA, fPSA and other routine physical examination clinical indicators that can be collected. The data acquisition module is integrated with the interface of the hospital information system (HIS) to automatically obtain the patient's routine physical examination data that is most closely related to the risk of prostate cancer.

[0085] The data acquisition module further includes an optical character recognition (OCR) submodule for capturing data from handwritten or printed paper medical records.

[0086] The data collection module further includes a natural language processing submodule, which is used to parse medical records and extract relevant clinical indicators.

[0087] The collected Shanghai hospital data are divided into two parts in a ratio of 3:7, namely, a training set and a validation set (internal and external validation sets). The training set is used to build and train the prediction model.

[0088] Step S3: Data cleaning and preprocessing

[0089] After collecting the data, data cleaning and preprocessing are performed. The data cleaning module aims to enhance the quality and integrity of the data. The data cleaning module includes a data quality check submodule and a missing value filling submodule. The data quality check submodule is used to automatically mark and delete incomplete or erroneous data. The missing value filling submodule is used to deal with the problem of missing data by taking some way to fill in the missing parts in the original data. For qualitative data, the pattern of similar patients is used to fill in, and for quantitative data, the average value of similar patients is used to fill in.

[0090] Step S4: Constructing a prediction model for predicting prostate cancer risk

[0091] like Figure 3As shown in the figure, after data preprocessing is completed, feature selection is used to find out which data is most likely to affect the risk of prostate cancer, and a risk prediction model is constructed based on this. Feature selection refers to selecting the most useful features for problem solving from the extracted feature set, reducing the complexity of the model and avoiding overfitting in order to improve the accuracy and efficiency of the model. The construction of the classification method refers to providing training data to the machine learning classifier and letting it learn the patterns and rules in the data. The classifier will learn based on the input features and adjust the internal parameters and weights to find the best classification decision boundary.

[0092] In this patent, we used 7 feature selection algorithms to select features from the data, including random forest, extreme gradient boosting tree, sparse neural network, mutual information metric, LASSO regression, linear correlation metric, and maximum correlation and minimum redundancy algorithm. At the same time, 7 classification methods were applied, including random forest classifier, extreme gradient boosting tree, adaptive classifier, K-nearest neighbor classifier, multilayer perceptron classifier, support vector machine classifier, and decision tree classifier. Through the combination of the above methods, 49 possible combinations of feature selection methods and classification methods were tested, and the prediction effect (AUC value) was evaluated using an internal validation cohort to determine the best performing combination to provide a high degree of flexibility and adaptability. Finally, the feature selection method of random forest and the prediction classification method of XGBoost were selected as the best modeling combination.

[0093] The risk assessment module includes a feature selection submodule and a model building submodule. The feature selection submodule is used to perform feature selection using a random forest feature selection algorithm, and the model building submodule is used to build a prediction model based on the selected features and using the XGBoost prediction classification method.

[0094] 1. The feature selection submodule uses the random forest feature selection algorithm to perform feature selection, which includes the following steps:

[0095] First, use the random forest algorithm to extract multimodal information from the data to obtain the importance of each feature.

[0096] In the random forest algorithm, for each decision tree, the corresponding out-of-bag data is selected to calculate the out-of-bag data error. Its basic form is shown in the following formula:

[0097]

[0098] In formula (1), N is the total number of samples, m is the number of misclassified samples,

[0099] Randomly add noise interference to the features of all samples of the out-of-bag data OBB, and calculate the out-of-bag data error again, denoted as err OBB 2. Assuming there are N trees in the forest, the importance of the feature is:

[0100] ∑(err OBB 1-err OBB 2) / N (2)

[0101] Calculate the importance of each feature according to formula (2);

[0102] Then, the features with high importance are screened out to form the modeling feature set, while reducing the data dimension.

[0103] 2. Model building submodule The method of building a prediction model based on the selected features and using the prediction classification method of XGBoost includes the following steps:

[0104] First, based on the modeling feature set, the XGBoost algorithm is used to build the initial classification model of multimodal fusion, and the model is trained in the classifier. A new tree is added in each iteration. This tree is generated by efficiently searching in the feature space and selecting the optimal split point. After each new tree is added, the weight of each leaf node is adjusted according to the gradient of the objective function and the information of the Hessian matrix, so that each new tree attempts to correct the residual left by the previous tree to minimize the prediction error. The cross-validation strategy is used to evaluate and optimize the performance of the trained model to obtain the optimal prediction model.

[0105] Then, when the trained prediction model classifies a case, it will fall into a corresponding leaf node in each tree according to the characteristics of the case. Each leaf node corresponds to a score. The prediction scores corresponding to each tree are added together to obtain the disease risk value of the case.

[0106] In XGBoost, the objective function is simplified to:

[0107]

[0108] In formula (3), G j represents the sum of the first-order derivatives of all samples, H j represents the sum of the second-order derivatives of all samples, γ represents the adjustment function for this item, T represents the number of leaf nodes in a tree, and f t Represents the predicted output of the model for the input sample after the tth iteration.

[0109] For the best combined model, we need to perform model optimization, i.e. hyperparameter tuning, to further improve the prediction accuracy. The risk assessment module further includes a model optimization submodule and a risk indication submodule. The model optimization submodule is used to perform hyperparameter tuning to improve the prediction accuracy. For example, the important final hyperparameters of the model are as follows: learning rate = 0.3, N-estimators = 100, Max-depth = 6, Subsample = 1, Colsample-ytree = 1.

[0110] The risk indicator submodule is used to display the risk level using color-coded indicators, such as green for low risk (risk score less than 25%), yellow for medium risk (risk score 25% to 50%), and red for high risk (risk score greater than 50%).

[0111] Model Evaluation:

[0112] For the constructed prediction model, the validation set was then used to test and evaluate the performance of the prediction model. First, we used descriptive statistics, and then calculated multiple indicators such as sensitivity and specificity to conduct a preliminary evaluation of the model. We also used a variety of methods to evaluate the indicators. The receiver operating characteristic curve (ROC) was used to evaluate the discrimination of the model, and the area under the curve (AUC) was used to compare the diagnostic performance (as shown in Table 1, Figure 4 Decision curves are also used to compare clinical utility (e.g. Figure 5 Finally, the calibration curve was used to assess the overestimation and underestimation of each cohort at a specific risk probability cutoff value (as shown in Figure 6 shown).

[0113] Table 1 Receiver characteristic curve, sensitivity, specificity, positive predictive value, negative predictive value of APCAI score, and AUC for predicting HGPCa

[0114]

[0115] By automatically inputting 18 routine physical examination indicators (age, urine specific gravity, tumor area under ultrasound, serum alkaline phosphatase, serum albumin, free PSA, prostate-specific antigen (PSA) value, fPSA / PSA, blood sugar, lymphocyte percentage, blood potassium, blood sodium, platelet distribution width, hematocrit, γ-glutamyl transpeptidase, absolute value of blood lymphocytes, absolute value of blood neutrophils and mean corpuscular hemoglobin) retrieved from the hospital HIS system into the above-constructed prediction model, the risk assessment results of prostate cancer can be automatically and quickly generated.

[0116] Step S5: Generate report and present evaluation results

[0117] The report generation and display module includes a report generation submodule and a display submodule. The report generation submodule is used to generate different reports based on different thresholds to meet the needs of different patients. The report contains ROC curves, calibration graphs, and decision curve analysis results. The display submodule is used to display the evaluation results in graphical and tabular form to help clinicians make more accurate diagnosis and treatment decisions. Users can choose to view or download reports online on multiple platforms (WeChat official account, mini program, web page, and APP).

[0118] Step S6: Develop a personalized follow-up plan

[0119] The follow-up and popular science module includes a follow-up sub-module and a popular science education sub-module.

[0120] The tracking and follow-up submodule is used to formulate a personalized follow-up plan based on the risk assessment results. The tracking and follow-up module of the present invention is to improve the long-term treatment effect and health management of patients. This module formulates a personalized follow-up plan based on the risk assessment results. For high-risk patients (red, risk score greater than 50%), the system will automatically set a monthly follow-up reminder and notify the patient by email or text message when the follow-up time is approaching. For medium-risk patients (yellow, risk score 25% to 50%), the system will set a follow-up reminder every 3 months. For low-risk patients (green, risk score less than 25%), the system recommends an examination every six months, and pushes it to the patient account on time on multiple platforms for reminders.

[0121] The popular science education submodule is used to provide patients with knowledge and preventive measures about prostate cancer, helping them understand and manage their health conditions.

[0122] Step S7: Save data

[0123] The data storage module is used to store the collected raw data in a relational database, such as MySQL or PostgreSQL, and is equipped with a backup mechanism to prevent data loss. The processed data is securely stored in a NoSQL database optimized for big data, such as MongoDB or HBase. The data storage module further includes a version control submodule, which is used to ensure that each data change can be traced back.

[0124] Step S8: System Management

[0125] The system management and security module includes the system management submodule, the data input submodule and the security submodule. The system management submodule is used to allow administrators to manage users and system settings. The data input submodule allows manual input or upload of data.

[0126] Since the system involves sensitive medical data, in order to ensure the confidentiality and integrity of the data, the security submodule is used to perform data encryption, access control, audit logs and security updates on sensitive medical data. This includes:

[0127] a. Data encryption: Use encryption algorithms (such as AES) to encrypt stored and transmitted data to ensure data security.

[0128] b. Access control: Set user roles and permissions, for example, administrators can access all functions, while ordinary users can only access reporting and follow-up functions.

[0129] c. Audit log: records the system's operation log for easy auditing and troubleshooting.

[0130] d. Security Updates: Regularly check and apply system and software security updates.

[0131] 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. An artificial intelligence system for predicting prostate cancer risk based on multiple routine physical examination indicators, characterized in that: include: A data acquisition module, wherein the data acquisition module is used to automatically obtain the patient's routine physical examination indicators most closely related to the risk of prostate cancer through an interface integration with a hospital information system; the routine physical examination indicators most closely related to the risk of prostate cancer include age, urine specific gravity, tumor area under ultrasound, serum alkaline phosphatase, serum albumin, free PSA, prostate cancer specific antigen (PSA) value, fPSA / PSA, blood sugar, lymphocyte percentage, blood potassium, blood sodium, platelet distribution width, hematocrit, γ-glutamyl transpeptidase, blood lymphocyte absolute value, blood neutrophil absolute value and mean corpuscular hemoglobin; A data cleaning module, the data cleaning module includes a data quality check submodule and a missing value filling submodule, the data quality check submodule is used to automatically mark incomplete or erroneous data, and the missing value filling submodule is used to handle the problem of missing data, using the pattern of similar patients to fill in the qualitative data, and using the average value of similar patients to fill in the quantitative data; A risk assessment module, comprising a feature selection submodule, a model building submodule, a model optimization submodule and a risk indication submodule; The feature selection submodule is used to perform feature selection using the random forest feature selection algorithm; The model building submodule is used to build a prediction model based on the selected features and use the prediction classification method of XGBoost; the model optimization submodule is used to perform hyperparameter tuning to improve prediction accuracy; The Risk Indicator submodule is used to display the risk level using color-coded indicators; A report generation and display module, the report generation and display module includes a report generation submodule and a display submodule, the report generation submodule is used to generate different reports according to different thresholds, the report contains ROC curves, calibration graphs and decision curve analysis results, and the display submodule is used to display the evaluation results in graphical and tabular forms; In the risk assessment module, the feature selection submodule uses the random forest feature selection algorithm to perform feature selection, including the following steps: First, use the random forest algorithm to extract multimodal information from the data to obtain the importance of each feature. In the random forest algorithm, for each decision tree, the corresponding out-of-bag data is selected to calculate the out-of-bag data error. Its basic form is shown in the following formula: In formula (1), N is the total number of samples, m is the number of misclassified samples, Randomly add noise interference to the features of all samples of the out-of-bag data OBB, and calculate the out-of-bag data error again, denoted as err OBB 2. Assuming there are N trees in the forest, the importance of the feature is: ∑(err OBB 1-err OBB 2) / N (2) Calculate the importance of each feature according to formula (2); Then, the features with high importance are selected to form the modeling feature set, while reducing the data dimension; In the risk assessment module, the model building submodule constructs a prediction model based on the selected features and uses the prediction classification method of XGBoost, including the following steps: First, based on the modeling feature set, the XGBoost algorithm is used to build the initial classification model of multimodal fusion, and the model is trained in the classifier. A new tree is added in each iteration. This tree is generated by efficiently searching in the feature space and selecting the optimal split point. After each new tree is added, the weight of each leaf node is adjusted according to the gradient of the objective function and the information of the Hessian matrix, so that each new tree attempts to correct the residual left by the previous tree to minimize the prediction error. The cross-validation strategy is used to evaluate and optimize the performance of the trained model to obtain the optimal prediction model. Then, when the trained prediction model classifies a case, it will fall into a corresponding leaf node in each tree according to the characteristics of the case. Each leaf node corresponds to a score. The prediction scores corresponding to each tree are added together to obtain the disease risk value of the case. In XGBoost, the objective function is simplified to: In formula (3), G j represents the sum of the first-order derivatives of all samples, H j represents the sum of the second-order derivatives of all samples, γ represents the adjustment function for this item, T represents the number of leaf nodes in a tree, and f t Represents the predicted output of the model for the input sample after the tth iteration; In the risk assessment module, the model building submodule uses seven feature selection algorithms, including random forest, extreme gradient boosting tree, sparse neural network, mutual information metric, LASSO regression, linear correlation metric, and maximum correlation and minimum redundancy algorithm, to select features from the data. At the same time, seven classification methods, including random forest classifier, extreme gradient boosting tree, adaptive classifier, K-nearest neighbor classifier, multilayer perceptron classifier, support vector machine classifier, and decision tree classifier, are applied. Through the combination of the above methods, 49 possible combinations of feature selection methods and classification methods are tested, and the prediction effect is evaluated using the internal validation cohort to determine the best performance combination. Finally, the feature selection method of random forest and the prediction classification method of XGBoost are selected as the best modeling combination, and the model hyperparameters of the best model combination are: learning rate = 0.3, N-estimators = 100, Max-depth = 6, Subsample = 1, Colsample-ytree = 1; In the risk assessment module, the model building submodule is also used to conduct a comprehensive validation evaluation of the model performance of the constructed prediction model based on the validation set using the following evaluation methods: AUC, sensitivity, specificity, APCAI score positive predictive value, APCAI score negative predictive value, decision curve and calibration curve.

2. The artificial intelligence system for predicting prostate cancer risk based on multiple routine physical examination indicators as claimed in claim 1, characterized in that: Further includes one or more of a system management and security module, a tracking and follow-up and science popularization module, and a data storage module; The system management and security module includes a system management submodule, a data input submodule and a security submodule. The system management submodule is used to allow an administrator to manage users and system settings. The data input submodule allows manual input or upload of data. The security submodule is used to perform data encryption, access control, audit logs and security updates on sensitive medical data. The tracking and follow-up and popular science module includes a tracking and follow-up submodule and a popular science education submodule, wherein the tracking and follow-up submodule is used to formulate a personalized follow-up plan according to the risk assessment results, and the popular science education submodule is used to provide patients with knowledge and preventive measures about prostate cancer; The data storage module is used to store the collected raw data in a relational database, equipped with a backup mechanism to prevent data loss, and to safely store the processed data in a NoSQL database optimized for big data.

3. The artificial intelligence system for predicting prostate cancer risk based on multiple routine physical examination indicators as claimed in claim 2, characterized in that: The data storage module further includes a version control submodule, and the version control submodule is used to ensure that each data change can be traced back.

4. The artificial intelligence system for predicting prostate cancer risk based on multiple routine physical examination indicators as claimed in claim 1, characterized in that: The data acquisition module further includes an optical character recognition submodule for capturing data from handwritten or printed paper medical records.

5. The artificial intelligence system for predicting prostate cancer risk based on multiple routine physical examination indicators as claimed in claim 1, characterized in that: The data acquisition module further includes a natural language processing submodule, which is used to parse medical records and extract relevant clinical indicators.

Images