Machine learning-based ampulla cancer prognosis model and construction method thereof

Abstract

The invention belongs to the technical field of biological information, and provides an ampulla cancer patient prognosis model based on machine learning and a construction method thereof. According to the method, clinical information of an ampulla cancer patient is obtained through a cancer statistical database, clinical characteristics independently related to survival of the patient are analyzed and positioned through single-factor and multi-factor COX, a prognosis model of the ampulla cancer is established through multiple machine learning algorithms according to the clinical characteristics, and the method can be used for survival prediction of the patient. The method has the advantages of being rapid in prediction, high in accuracy and easy and convenient to use, gene determination does not need to be specially conducted on tissue or blood sampling of the patient, accurate survival rate judgment can be achieved only through clinical information, a clinician can make a personalized treatment scheme of the patient according to the prognosis state, and therefore the survival rate of the patient is increased.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics, and in particular relates to a prognostic model for ampullary carcinoma based on machine learning and a construction method thereof. Background Art

[0002] Ampullary carcinomas (AC) are a rare cancer, accounting for about 0.2% of all gastrointestinal tumors, with an annual cancer incidence of about 5.7 cases per million people, of which the incidence in men is higher, at 3.8 cases per million people, while the incidence in women is 2.7 cases per million people. Due to its rare nature, there is still a lack of large-scale randomized clinical studies on ampullary carcinoma. In addition, ampullary carcinoma originates from the junction of the pancreatic duct, bile duct and duodenum, and its histological type is complex, with different subtypes such as intestinal type, pancreaticobiliary type and mixed type. Due to its rare and complex histological characteristics, the treatment and prognosis of patients with ampullary carcinoma are still unclear. At present, the treatment of ampullary carcinoma is usually pancreaticoduodenectomy, and the role of adjuvant therapy in this cancer is still controversial, and there is still no specific targeted drug for ampullary carcinoma. In terms of prognosis, patients with different histological subtypes have different prognostic tendencies, and it is difficult to recruit enough patients for prognostic studies. Currently, the field of survival prediction for ampullary carcinoma is still blank, so it is very important to establish a new prognostic tool for patients with ampullary carcinoma, which can be used to determine the patient's survival period and guide clinical treatment.

[0003] Machine Learning (ML) is a new type of artificial intelligence technology that can extract useful data from large-scale information through data mining, and improve performance through continuous learning and self-improvement. In recent years, machine learning has been widely used in the field of medical health, and is commonly used in the diagnosis, prognosis and treatment of diseases. Due to the lack of clinical samples caused by the rarity of ampullary carcinoma, the prediction of the current ampullary carcinoma prognosis model is still not accurate enough. Through machine learning methods, the complex interrelationships between patient information and survival time can be extracted from clinical data, and computers can be used to simulate human learning behavior, thereby continuously improving the accuracy of predictions to improve model performance. Therefore, the present invention uses a machine learning method to construct an ampullary carcinoma prognosis model, trains the model with the clinical information of ampullary carcinoma patients, achieves high-accuracy prediction of patient survival time, and evaluates characteristic indicators related to prognosis. Summary of the invention

[0004] In view of the existing technical problems, the present invention provides a prognostic model for ampullary carcinoma patients based on machine learning and a construction method thereof, which specifically includes the following contents:

[0005] In a first aspect, the present invention provides a prognostic model for ampullary carcinoma based on machine learning, and the method for constructing the prognostic model for ampullary carcinoma comprises:

[0006] (1) Download clinical information of ampullary carcinoma patients related to prognosis from the SEER database; clinical information of ampullary carcinoma patients includes: basic information of patients, disease information and treatment records;

[0007] (2) Performing prognostic correlation analysis on the clinical information obtained in step (1) by univariate and multivariate COX regression to obtain clinical features significantly correlated with prognosis;

[0008] (3) Performing feature processing on the clinical features related to prognosis obtained in step (2), standardizing the numerical variables, and converting the reward type variables into dummy variables that can be recognized by the algorithm;

[0009] (4) Dividing the data processed in step (3) into a test set and a training set, constructing 1-year, 3-year, and 5-year survival prediction models for the training set using a machine learning algorithm, and optimizing the model by traversing possible hyperparameter combinations using a grid search during the construction process;

[0010] (5) The established models were evaluated using AUC, accuracy, balanced accuracy, recall, specificity, precision, and F1 score indicators to select the best model in survival prediction.

[0011] Preferably, the basic patient information includes age, gender, median family income, race and marital status; the disease information includes tumor grade, tumor stage, histological type, tumor metastasis; and the treatment records include surgery, regional lymph node surgery, radiotherapy, chemotherapy and systemic treatment.

[0012] Preferably, the tumor staging includes: comprehensive tumor staging and TNM staging.

[0013] Preferably, the tumor metastasis includes bone metastasis, brain metastasis, liver metastasis, lung metastasis, distant lymph node metastasis, and other metastases.

[0014] Preferably, the treatment record also includes surgery and systemic treatment, radiotherapy sequence, systemic treatment and surgery sequence.

[0015] Preferably, the machine learning algorithm in step (3) includes: random forest algorithm, support vector machine algorithm, logistic regression algorithm, neural network algorithm, K nearest neighbor algorithm and extreme gradient boosting algorithm.

[0016] Preferably, the SEER database in step (1) is a Surveillance Epidemiology and EndResults database.

[0017] Preferably, the prognostic correlation analysis in step (2) uses the coxph function of the Survival package in the R language to perform COX regression; features with P<0.05 in the univariate COX regression are included in the multivariate COX regression; and features with P<0.05 in the multivariate COX regression are considered to be significantly correlated with prognosis.

[0018] Preferably, in the feature processing in step (3), missing data values ​​are filled with the median, the StandardScaler function in the Sklearn package of the Python language is used to convert the numerical variables into a standard normal distribution, and the categorical variables are converted into dummy variables in the form of one-hot encoding through the OneHot Encoder function.

[0019] Preferably, in step (4), the ratio of the training set to the test set is 8:2.

[0020] Preferably, in step (4), the GridSearchCV function in the Sklearn package is used for grid search, and the KNeighborsClassifier, LogisticRegression, MLPClassifier, RandomForestClassifier, SVC and XG BClassifier functions are used for model construction.

[0021] In a second aspect, the present invention provides an electronic device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, the prognosis model for ampullary carcinoma patients described in the first aspect is implemented.

[0022] In a third aspect, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the prognostic model for ampullary carcinoma patients described in the first aspect.

[0023] The beneficial effects of the present invention are as follows: the present invention constructs a prognostic model for ampullary carcinoma based on the clinical characteristics of patients based on machine learning methods for the first time; the model incorporates the largest-scale ampullary carcinoma patient data to date for training, and survival prediction can be achieved by only providing the patient's clinical information, without the need for special blood or tissue sampling and genetic testing, thereby reducing costs; in addition, the model can be used to analyze characteristic factors related to prognosis, thereby determining whether the patient is suitable for chemotherapy and guiding doctors in making clinical decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A machine learning-based prognostic model for ampullary carcinoma patients and a technical roadmap for its construction method;

[0025] Figure 2 Model evaluation heatmaps for six machine learning algorithms;

[0026] Figure 3 ROC curves for six machine learning algorithms;

[0027] Figure 4 Confusion matrix diagram of the optimal model;

[0028] Figure 5 Feature importance analysis of the optimal model;

[0029] Figure 6 KM survival analysis of patients receiving chemotherapy stratified by the three-year survival prediction model. DETAILED DESCRIPTION

[0030] The following examples further describe and illustrate the present invention, but they are not intended to limit the scope of protection claimed by the present invention. The experimental methods used in the following examples are conventional methods unless otherwise specified.

[0031] Example 1: Constructing a machine learning prognostic model for ampullary carcinoma patients and screening patients suitable for chemotherapy

[0032] The prognostic model of ampullary carcinoma patients based on machine learning and its construction method technology roadmap are as follows Figure 1 As shown, specifically including: 1. Data source and preprocessing

[0033] (1) Data acquisition:

[0034] Patient information on ampullary carcinoma was obtained from the SEER database (https: / / seer.cancer.gov / ) and accessed using SEER*Stat software. The downloaded patient characteristics included survival time, survival status, age, sex, median household income, time from diagnosis to treatment, race, tumor grade, histology, comprehensive tumor stage, TNM stage, regional lymph node surgery, other site surgery, surgery and radiotherapy sequence, surgery, radiotherapy, chemotherapy, systemic therapy and surgery sequence, bone metastasis, brain metastasis, liver metastasis, lung metastasis, distant lymph node metastasis, other metastasis, marital status, and urban and rural residence. A total of 8204 patient data between 2000 and 2021 that met the following criteria were obtained: ① The tumor site was recorded as ampullary in ICD-O-3; ② Patients with multiple primary cancers were excluded; ③ Patients diagnosed only by autopsy were excluded; ④ Patients with only death certificates were excluded; ⑤ Patient data contained clear survival time.

[0035] (2) Data preprocessing: Preprocess the acquired patient data, merge the patient's tumor staging status, normalize the data, and supplement or delete some missing data.

[0036] 2. Univariate and multivariate COX analysis

[0037] (1) Univariate COX analysis: Univariate COX analysis was performed on the data of patients who had completed pretreatment to perform feature selection and determine the correlation between different clinical characteristics and patient prognosis. The results of univariate COX analysis showed that among the clinical characteristics of all patients included in the data, brain metastasis, bone metastasis, lung metastasis, liver metastasis, other metastasis and advanced age were closely related to poor prognosis, while surgery, regional lymph node surgery and systemic therapy were favorable factors for patient prognosis, and no correlation was shown between gender and prognosis.

[0038] (2) Multivariate COX analysis: Further multivariate COX analysis was performed on the patient data to screen for independent prognostic factors that were still associated with prognosis after excluding the mutual influence between variables. The results of multivariate COX analysis showed that age, gender, median family income, time from diagnosis to treatment, race, tumor grade, histology, comprehensive tumor stage, T stage, N stage, regional lymph node surgery, surgery and radiotherapy sequence, surgery, chemotherapy, systemic treatment and surgery sequence, bone metastasis, other metastasis and marital status were still closely related to prognosis after excluding the mutual influence.

[0039] 3. Machine learning algorithms predict patient outcomes

[0040] (1) Data preprocessing: The independent prognostic factors selected by multivariate COX analysis were used as features to build the model to improve the model effect and reduce data redundancy. The median was used to fill the missing values ​​of the numerical variables, and the StandardScaler function was used to transform the numerical variables into a form that conforms to the standard normal distribution. The categorical variables were transformed into dummy variables through one-hot encoding using the OneHotEncoder function.

[0041] (2) Model construction: The test set and training set were divided into a ratio of 8:2. Hyperparameter traversal was performed through grid search to improve model performance. Six machine learning methods, including random forest, support vector machine, logistic regression, neural network, K nearest neighbor, and extreme gradient boosting, were used to construct 1-year, 3-year, and 5-year survival prediction models for patients with ampullary cancer.

[0042] (3) Model evaluation: After the model is established, the model prediction performance is evaluated by AUC, accuracy, balanced accuracy, recall, specificity, precision and F1 score, and the heat map and ROC curve of each parameter of the model evaluation are drawn, such as Figure 2 , Figure 3As shown in Figure 2. The best machine learning models for 1-year, 3-year, and 5-year predictions were selected, and it was found that the extreme gradient boosting algorithm performed best for 1 year, while the logistic regression algorithm had the highest accuracy for 3-year and 5-year survival predictions. Further confusion matrix analysis was performed on the best performing model to evaluate its performance, as shown in Figure 2. Figure 4 shown.

[0043] (4) Feature importance analysis: Feature importance analysis was performed on the best prediction models for 1 year, 3 years, and 5 years. The results are as follows: Figure 5 As shown. The importance of model features was evaluated by the SHAP method and ranked according to their impact on patient survival. The results showed that surgery ranked highest in feature importance and was the preferred standard treatment for ampullary carcinoma. In addition, other metastases, age, grade, N stage, T stage, and comprehensive tumor stage of the patient were also important prognostic features. While chemotherapy played an important role in the 1-year prognostic model, its importance decreased in predicting 3-year and 5-year survival rates.

[0044] 4. Screening patients suitable for chemotherapy through the model

[0045] (1) Patient screening: Patient data is imported into the established 3-year survival prediction model for survival judgment. Patients judged by the model to be alive are classified into the low-risk group, while patients judged to be dead are classified into the high-risk group. Since patients with ampullary carcinoma with a poor prognosis are usually more suitable for chemotherapy, while patients with a good prognosis may be adversely affected by chemotherapy, chemotherapy is recommended for patients in the high-risk group classified by the model.

[0046] (2) KM survival analysis: KM survival analysis was performed on patients in the high-risk and low-risk groups determined by the model, such as Figure 6 As shown, patients in the high-risk group had a higher survival rate after chemotherapy, while the survival rate of patients in the low-risk group tended to decrease after chemotherapy.

[0047] 5. Results Analysis

[0048] In summary, based on the data of patients with ampullary carcinoma in the SEER database, the present invention conducted a correlation analysis on the clinical characteristics of patients with ampullary carcinoma through univariate and multivariate COX analysis, developed a prognostic model using six machine learning algorithms, and iterated the model through grid search to improve the performance, thereby achieving accurate survival judgment.

[0049] The 1-year, 3-year and 5-year survival prediction models provided by the present invention have AUC values ​​greater than 0.8 in the test set validation, and have excellent predictive performance, which can help doctors determine the prognosis of patients and guide clinical decision-making. The present invention can stratify patients by combining a prognostic model with different factors. The survival probability of patients in the high-risk group (predicted survival period less than 3 years) is significantly improved after chemotherapy, while the long-term survival rate of the low-risk group (predicted survival period greater than 3 years) after chemotherapy is reduced, providing a new tool for distinguishing the applicability of chemotherapy to different populations of patients with ampullary cancer.

[0050] The above description is a detailed description of the preferred feasible embodiments of the present invention, but the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modified changes completed under the technical spirit suggested by the present invention should fall within the patent scope covered by the present invention.

Claims

1. A prognostic model for ampullary carcinoma based on machine learning, characterized in that: The method for constructing the prognostic model for ampullary carcinoma comprises: (1) Download clinical information of ampullary carcinoma patients related to prognosis from the SEER database; clinical information of ampullary carcinoma patients includes: basic information of patients, disease information and treatment records; (2) Performing prognostic correlation analysis on the clinical information obtained in step (1) by univariate and multivariate COX regression to obtain clinical features significantly correlated with prognosis; (3) Performing feature processing on the clinical features related to prognosis obtained in step (2), standardizing the numerical variables, and converting the reward type variables into dummy variables that can be recognized by the algorithm; (4) Dividing the data processed in step (3) into a test set and a training set, constructing 1-year, 3-year, and 5-year survival prediction models for the training set using a machine learning algorithm, and optimizing the model by traversing possible hyperparameter combinations using a grid search during the construction process; (5) The established models were evaluated using AUC, accuracy, balanced accuracy, recall, specificity, precision, and F1 score indicators to select the best model in survival prediction.

2. The prognostic model for ampullary carcinoma according to claim 1, characterized in that: The basic patient information includes age, gender, median household income, race and marital status; the disease information includes tumor grade, tumor stage, histological type, and tumor metastasis; and the treatment records include surgery, regional lymph node surgery, radiotherapy, chemotherapy, and systemic therapy.

3. The prognostic model for ampullary carcinoma according to claim 2, wherein: The machine learning algorithms in step (3) include: random forest algorithm, support vector machine algorithm, logistic regression algorithm, neural network algorithm, K nearest neighbor algorithm and extreme gradient boosting algorithm.

4. The prognostic model for ampullary carcinoma according to claim 3, characterized in that: The SEER database in step (1) is the Surveillance Epidemiology and End Results database.

5. The prognostic model for ampullary carcinoma according to claim 4, characterized in that: The prognostic correlation analysis in step (2) uses the coxph function of the Survival package in the R language to perform COX regression; features with P < 0.05 in the univariate COX regression are included in the multivariate COX regression; and features with P < 0.05 in the multivariate COX regression are considered to be significantly correlated with prognosis.

6. The prognostic model for ampullary carcinoma according to claim 5, characterized in that: In the step (3), the feature processing fills the missing data values ​​with the median, and the StandardScaler function in the Sklearn package of the Python language is used to convert the numerical variables into a standard normal distribution, and the categorical variables are converted into dummy variables in the form of one-hot encoding through the OneHotEncoder function.

7. The prognostic model for ampullary carcinoma according to claim 6, characterized in that: In step (4), the ratio of the training set to the test set is 8:

2.

8. The prognostic model for ampullary carcinoma according to claim 7, wherein: In the step (4), the GridSearchCV function in the Sklearn package is used for grid search, and the KNeighborsClassifier, LogisticRegression, MLP Classifier, RandomForestClassifier, SVC and XGBClassifier functions are used for model construction.

9. An electronic device, characterized in that: The invention comprises a processor, a memory and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, the prognosis model for ampullary carcinoma patients according to any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the prognostic model for ampullary carcinoma patients according to any one of claims 1 to 8 is implemented.

Images