TWI827043B - A method to establish a clinical decision support system for second primary cancers in colorectal cancer using predictive models and visualization methods - Google Patents

Abstract

A method to establish a clinical decision support system for second primary cancer in colorectal cancer using predictive models and visualization. First, the cancer characteristics are combined into a second primary cancer risk assessment feature combination, and the subject's corresponding feature combination is obtained. Clinical data and establish a second primary cancer risk assessment database; input the second primary cancer risk assessment database into the machine learning machine, build a second primary cancer risk assessment model through the machine learning algorithm, and interpret the model through the feature interpreter ; Form a visual clinical decision support system. Obtain the clinical data of new subjects, input it into the clinical decision support system, predict the risk of second primary cancer and the impact of each cancer characteristic on the risk, and present it in a visual way. This way, patients can be given the most appropriate clinical guidance.

Description

A method to establish a clinical decision support system for second primary cancers in colorectal cancer using predictive models and visualization methods

The present invention provides a method for establishing a clinical decision support system, in particular, a method for establishing a clinical decision support system for colorectal cancer second primary cancer in a predictive model and visualization manner. Through this, the risk of second primary cancer can be assessed specifically for colorectal cancer, appropriate clinical decisions can be made, and appropriate personalized advice can be given to patients.

Second Primary Cancers (SPCs), also known as multiple cancers, refer to cancers that have more than two independent lesions in a patient. As cancer screening and treatments become more effective, cancer cure and survival rates have increased. However, the number of second primary cancers diagnosed has also increased. In addition, second primary cancer is the main reason for reducing cancer survival rate. It not only reduces the cure rate of cancer, but also reduces the survival time. Therefore, early detection of second primary cancers is crucial to patient health.

However, nowadays, most of them are carried out through routine and passive follow-up examinations. There is no specific diagnosis and test for second primary cancer, and it is impossible to effectively assess the risk of second primary cancer. Patients may also miss opportunities for early detection. Among them, there is currently no feasible and objective technology in clinical practice for risk assessment and clinical decision-making recommendations for second primary cancers in colorectal cancer.

Based on the above, there is currently a lack of feasible technology that can effectively and conveniently provide clinical Healthcare professionals assess the risk of second primary cancers in colorectal cancer and develop personalized clinical decision-making recommendations.

In view of this, the present invention provides a method for establishing a clinical decision support system for colorectal cancer second primary cancer in a predictive model and visualization manner, including: (a) combining multiple cancer characteristics into a second primary cancer risk assessment Feature combination; (b) Obtain clinical data of multiple subjects corresponding to the second primary cancer risk assessment feature combination to establish a second primary cancer risk assessment database; (c) Combine the second primary cancer risk The assessment database is input into a machine learning machine, and a second primary cancer risk assessment model is established through a machine learning algorithm; (d) the second primary cancer risk assessment model is interpreted through a feature interpreter, and each cancer feature pair is calculated. The impact value of the risk; (e) Present it in the form of images to form a visual clinical decision support system; (f) Obtain the clinical data of new subjects corresponding to the second primary cancer risk assessment feature combination, Input the clinical decision support system, and use the machine learning algorithm to compare and analyze the risk of second primary cancer, as well as the impact of the above-mentioned new subjects' cancer characteristics on the risk, and present it in a visual way. The clinical decision support system; and (g) based on the clinical decision support system presented in step (f), provide risk reduction recommendations based on the impact value of each cancer characteristic, and track changes in risk.

Among them, the influence value of each cancer characteristic is the Sharp power value. When the above-mentioned Sharp power value is positive, the risk of colorectal cancer developing second primary cancer is increased; when the above-mentioned Sharp power value is negative, the risk of colorectal cancer developing second primary cancer is reduced.

The above image is in the form of a bar chart.

The present invention uses the second primary cancer risk assessment feature combination and the machine learning to Learn the machine to establish the second primary cancer risk assessment model, and finally establish the clinical decision support system in a visual way so that medical staff can make a more comprehensive assessment. Because the second primary cancer risk assessment feature combination includes different aspects of the cancer characteristics, medical staff can consider the patient's multi-faceted conditions at the same time, thereby improving the accuracy and effectiveness of the second primary cancer risk assessment. In addition, the clinical decision support system is presented in a visual manner, allowing clinicians to make appropriate clinical decisions conveniently, simply and quickly. In addition, the clinical decision support system can timely display the impact of numerical changes in each cancer characteristic on the risk of second primary cancer, assist clinicians in assessing the cancer characteristics that increase or reduce risk, and provide patients with appropriate personalized recommendations.

10: Second primary cancer risk assessment feature combination

101: Cancer Characteristics

12: Second primary cancer risk assessment database

14:Machine learning machine

16: Second primary cancer risk assessment model

18:Feature Interpreter

20:Clinical decision support system

Figure 1 is a block flow chart of the present invention;

Figure 2 is an interface simulation diagram of the clinical decision support system according to the first embodiment of the present invention;

Figure 3 shows the clinical decision support system in the first embodiment of the present invention graphically presenting the impact values of the cancer characteristics on risk;

Figure 4 is an interface simulation diagram of the clinical decision support system according to the second embodiment of the present invention;

FIG. 5 is a graphical representation of the impact of cancer characteristics on risk in the clinical decision support system according to the second embodiment of the present invention.

Referring to Figure 1, the present invention discloses a method for establishing a clinical decision support system for second primary cancer in colorectal cancer using a predictive model and visualization method, including: (a) combining multiple cancer characteristics 101 into a second primary cancer risk Evaluation feature combination 10; (b) Obtain multiple subjects The clinical data corresponding to the second primary cancer risk assessment feature combination 10 is used to establish a second primary cancer risk assessment database 12; (c) inputting the second primary cancer risk assessment database 12 into a machine learning The machine 14 establishes a second primary cancer risk assessment model 16 through a machine learning algorithm; (d) interprets the second primary cancer risk assessment model 16 through a feature interpreter 18 and calculates the effect of each cancer feature 101 on the risk. Impact value; (e) Presented in the form of images to form a visual clinical decision support system 20; (f) Obtain clinical data of new subjects corresponding to the second primary cancer risk assessment feature combination 10, Input the clinical decision support system 20 and use the machine learning algorithm to compare and analyze the risk of second primary cancer and the impact of the above-mentioned new subjects' cancer characteristics 101 on the risk in a visual manner. Presented in the clinical decision support system 20; and (g) according to the clinical decision support system 20 presented in step (f), provide risk reduction recommendations based on the impact value of each cancer feature 101, and track changes in risk.

Next, referring to Figures 2 to 5, together with Figure 1, the following is based on the clinical data of the second primary cancer risk assessment feature combination 10 of patients with colorectal cancer in a medical center in Taiwan, using the method disclosed in the present invention to establish a large intestine The clinical decision support system 20 for the occurrence of second primary cancer in rectal cancer is an example of an embodiment of the present invention. The method is as follows:

1. Conditions (inclusion and exclusion conditions) and number of subjects:

The subjects are patients with colorectal cancer. This example uses medical record review and does not require additional recruitment of subjects.

2. Data review period and execution period of this embodiment:

The data is backdated from January 1, 2004 to December 31, 2018.

3. Design and methods:

(1) First, the cancer registration data of 32,990 colorectal cancer patients were collected. the The second primary cancer risk assessment feature set 10 includes 42 such cancer features 101, which are as follows: age at diagnosis, gender, primary site, tissue morphology, sex code, grade and differentiation, tumor size, regional lymph node invasion Number, clinical tumor stage T, clinical tumor stage N, clinical tumor stage M, clinical stage, pathological tumor stage T, pathological tumor stage N, pathological tumor stage M, pathological stage, consolidation stage, surgery, primary site Surgical margins, radiation therapy target range summary, radiation therapy equipment, radiation therapy, surgery and radiation therapy sequence, highest radiation dose radiation dose, highest radiation dose radiation therapy number, lower radiation dose radiation dose, lower radiation dose radiation therapy number , systemic treatment, BMI, smoking behavior, betel nut chewing behavior, drinking behavior, SSF1, SSF2, SSF3, SSF4, SSF5, SSF6, SSF7, SSF8, SSF9 and cancer order.

(2) Establish the second primary cancer risk assessment database 12 based on the clinical data of the above 32,990 colorectal cancer patients corresponding to the second primary cancer risk assessment feature combination 10 . Then, through the above machine learning algorithm, a basic classifier is established and optimized, and finally the basic classifiers are filtered, leaving basic classifiers with different viewpoints, and then a meta-classifier is established to integrate the classification results of the basic classifiers. The second primary cancer risk assessment model 16 is thereby established.

(3) After the second primary cancer risk assessment model 16 is established, use the SHapley Additive exPlanations (SHAP) algorithm as the feature interpreter 18 to analyze the best second primary cancer risk assessment model 16. The cancer risk assessment model 16 calculates the impact value of each cancer feature 101 on the risk in the model, such as the Shapley value, and presents the risk impact value result of the subject in a visual manner.

(4) Use the second primary cancer risk assessment model 16 and the feature interpreter 18 to establish the interactive clinical decision support system 20 . Clinicians can use Input the cancer characteristics 101 of the new patient's second primary cancer, and predict the risk of the second primary cancer and the impact of each cancer characteristic 101 on the risk through the established second primary cancer risk assessment model 16 , and present the original values of the cancer characteristics 101 of the above-mentioned new patients and their impact values on the risk of second primary cancer in a visual manner in the clinical decision support system 20 .

Results:

Through the feature interpreter 18, personalized prediction interpretation is provided. For example, a patient's cancer in situ treatment method is surgery, which increases the risk of second cancer. Age, BMI, non-smoking and other current conditions reduce the risk of second cancer. Physicians can observe changes in risk by adjusting system values, such as adjusting BMI. If lowering BMI can reduce the risk of secondary cancer, they can give patients appropriate weight loss recommendations.

Figures 2 to 5 are examples, which respectively present simulation diagrams of the clinical decision support system 20 for different subjects. The first subject is presented in Figures 2 and 3, and the second subject is presented in Figures 4 and 5. . Referring to Figure 2, the medical staff inputs the clinical data of the first subject. As shown in Figure 2, the gender of the first subject is male, age is 65 years old, BMI is 16, the surgical margin of the primary site is No, the primary site is the right colon, the number of regional lymph node invasion is Unexamined, and the grade is The tumor is well differentiated, the tumor size is 1-49 mm, the tumor is stage II, the patient has received radiation therapy, and the patient smokes and drinks alcohol. After entering the above information and clicking the prediction button, the second primary cancer risk assessment in Figure 3 will be displayed. The left side shows that the risk of colorectal cancer in the first subject to develop a second primary cancer is three times higher than that of the study population. The Sharpe values of these cancer characteristics 101 are presented in a bar graph on the right, allowing doctors to quickly see the influence of each cancer characteristic 101 on the risk of second primary cancer. A positive value of the Shapley value represents an increased risk of colorectal cancer developing a second primary cancer, and a negative value represents a decreased risk of colorectal cancer developing a second primary cancer. Among the cancer characteristics 101 of the first subject, the primary site is the right colon, with a Sharp power value of 0.138 being the highest, indicating that the primary site is the right colon. The first subject's risk of second primary cancer is increased, with the greatest impact. Other cancer characteristics 101 that increase the risk of second primary cancer in the first subject are tumor size, smoking behavior, drinking behavior, number of regional lymph node invasion, betel nut chewing behavior, age, gender, radiotherapy and primary cancer. site surgical margins, etc. Thereby, the physician can evaluate which cancer characteristics 101 to provide recommendations to the first subject to help the first subject reduce the risk of a second primary cancer. In addition, doctors can also observe changes in the risk of second primary cancers for each cancer feature 101 by adjusting the values of the cancer features 101 .

Referring to Figure 4, the medical staff inputs the clinical data of the second subject. As shown in the figure, the gender of the second subject is male, age is 52 years old, BMI is 28, the surgical margin of the primary site is No, the primary site is the rectum, the number of regional lymph node invasion is Unexamined, and the grade and differentiation are The tumor is well differentiated, the tumor size is 50-99 mm, the tumor is stage II, and the patient has received radiation therapy, smokes and drinks alcohol. After inputting the above information and clicking the prediction button, the second primary cancer risk assessment in Figure 5 will be displayed. The left side shows that the second subject's risk of developing colorectal cancer as a second primary cancer was 0.15 times that of the study population. The bar graph on the right shows that among the cancer characteristics 101 of the second subject, the Sharpe value of -0.176 for the second subject's cancer stage is the lowest, which means that the second subject's cancer stage is lower than the second stage. It is the risk of second primary cancer and has the greatest impact. Other cancer characteristics that reduce the risk of a second primary cancer in the second subject are age, tumor size, primary site and number of regional lymph node invasion, etc.

Based on the above, medical staff can consider the patient's multi-faceted conditions at the same time, thereby improving the accuracy and effectiveness of second primary cancer risk assessment.

The above machine learning algorithm uses

Logistic regression, multivariate adaptive regression splines, Decision tree classifiers, Rule-based classifier, Nearest Neighbor Classifiers, Naïve Bayes classifier, Bayesian Networks, Artificial Neural Network, Deep learning, support vector machine, Random forest forest), extreme gradient boosting (eXtreme Gradient Boosting), categorical boosting, light gradient boosting machine (Light Gradient Boosting Machine), ensemble learning (Ensemble learning Methods), bagging method and boosting method-based classifiers ( Bagging and Boosting-based Classifiers), Adaptive Boosting-based classifiers (Adaptive Boosting-based classifiers), Fuzzy Set-based classifiers (Fuzzy Set-based classifiers), Genetic Algorithms -based classifiers), Genetic Programming-based classifiers, Meta heuristic-based classifiers, linear and non-linear discriminant analysis (linear and non-linear discriminant analysis) or any combination of the above algorithms.

The feature interpreter includes Local Interpretable Model-Agnostic Explanations (LIME), Deep Learning Important FeaTures (DeepLIFT), and Layer-Wise Relevance Propagation. ,LRP), Classic Shapley Value Estimation, SHapley Additive exPlanations (SHAP), Shapley value based model explanations or Any combination of the above algorithms.

The second primary cancer risk assessment feature combination 10 is gender, birth year, initial diagnosis date, initial pathological diagnosis date, cancer diagnosis method, primary site, laterality, tissue type, sex code, grade and differentiation, clinical Tumor size, pathological tumor size, number of regional lymph nodes examined, number of regional lymph node invasion, distance between surgical margin and tumor cells, surgical margin at primary site, tumor staging version, clinical tumor stage T, clinical tumor stage N, clinical tumor stage M, Clinical tumor stage, pathological tumor stage T, pathological tumor stage N, pathological tumor stage M, pathological tumor stage, apply for a hospital to perform surgical treatment on the primary site of the tumor, apply for a hospital to perform surgical treatment on the primary site, apply for a hospital to perform peripheral lymph node resection, apply for a hospital Scope of regional lymph node surgery, date of first surgery, radiation therapy at the reporting hospital for the primary site, radiation therapy method for the primary site, radiation dose for external beam therapy at the primary site, number of external beam radiation treatments, start of external beam radiation therapy at the reporting hospital Date, end date of external radiation therapy in the applying hospital, brachytherapy in the applying hospital, dose of brachytherapy, chemotherapy in the applying hospital, simultaneous chemotherapy and radiation therapy, chemotherapy method, number of chemotherapy courses, start date of chemotherapy in the applying hospital, application Hospital hormone treatment, declaration of start date of hospital hormone treatment, date of last contact or death, survival status, cancer status, date of first recurrence, type of first recurrence, cause of death, carcinoembryonic Antigen (CEA) test value, tumor shrinkage grade, Pathological circumferential resection margin, nerve invasion, KRAS (Kirsten rat sarcoma virus) test value, whether there is intestinal obstruction before or during surgery, whether there is intestinal perforation before or during surgery, or any combination of the above.

10: Second primary cancer risk assessment feature combination

101: Cancer Characteristics

12: Second primary cancer risk assessment database

14:Machine learning machine

16: Second primary cancer risk assessment model

18:Feature Interpreter

20:Clinical decision support system

Claims (9)

A method to establish a clinical decision support system for second primary cancers in colorectal cancer using predictive models and visualization, including: (a) Combining multiple cancer features into a second primary cancer risk assessment feature combination; (b) Obtain clinical data of multiple subjects corresponding to the second primary cancer risk assessment characteristic combination to establish a second primary cancer risk assessment database; (c) Input the second primary cancer risk assessment database into a machine learning machine and establish a second primary cancer risk assessment model through a machine learning algorithm; (d) Interpret the second primary cancer risk assessment model through a feature interpreter and calculate the impact value of each cancer feature on the risk; (e) Presented in the form of images to form a visual clinical decision support system; (f) Obtain the clinical data of new subjects corresponding to the second primary cancer risk assessment feature combination, input it into the clinical decision support system, and predict the risk of second primary cancer through comparison and analysis with the machine learning algorithm, And the impact value of each cancer characteristic on the risk of the above-mentioned new subjects is presented in a visual way in the clinical decision support system; and (g) Based on the clinical decision support system presented in step (f), give risk reduction recommendations based on the impact value of each cancer characteristic, and track changes in risk. A method for establishing a clinical decision support system for second primary cancer in colorectal cancer using a predictive model and visualization method as described in claim 1, wherein the above machine learning algorithm uses Logis regression, multivariate adaptive cloud Type regression, decision tree, rule classification, nearest neighbor classification, naive Bayesian classifier, Bayesian network, neural network, deep learning, support vector machine, random forest, extreme gradient boosting, class boosting, optical gradient Boosting machine, ensemble learning, bagging method and boosting method-based classifiers, adaptive boosting-based classifiers, fuzzy set-based classifiers, genetic algorithm-based classifiers, basic Classifiers based on programming methods, classifiers based on universal heuristics, linear and nonlinear discriminant analysis, or any combination of the above. A method for establishing a clinical decision support system for second primary cancer in colorectal cancer using a predictive model and visualization method as described in claim 1, wherein the feature interpreter includes a locally understandable model-independent interpretation method, a deep learning important Feature method, layer-by-layer correlation propagation, classic Sharpe force value estimation method, Sharpe force additivity interpretation method, Sharpe force value-based interpretation method, or any combination of the above. A method for establishing a clinical decision support system for colorectal cancer second primary cancer in a predictive model and visualization manner as described in claim 1, wherein the second primary cancer risk assessment feature combination is gender, year of birth, initial Diagnosis date, initial pathological diagnosis date, cancer diagnosis method, primary site, laterality, tissue type, sex code, grade and differentiation, clinical tumor size, pathological tumor size, number of regional lymph nodes examined, number of regional lymph node invasion, surgical margins Distance from tumor cells, surgical margin of primary site, tumor staging version, clinical tumor stage, pathological tumor stage, apply for hospital to perform surgical treatment on the primary site of tumor, apply for hospital's surgical method for primary site, apply for hospital to remove peripheral lymph nodes, apply for The scope of regional lymph node surgery in the hospital, the date of the first operation, the applying hospital’s radiotherapy for the primary site, the radiotherapy method for the primary site, the dose of external radiotherapy for the primary site, the number of external radiotherapy treatments, the applied hospital’s external radiotherapy Start date, end date of external radiation therapy in the applied hospital, brachytherapy in the applied hospital, dose of brachytherapy, chemotherapy in the applied hospital, simultaneous chemotherapy and radiation therapy, chemotherapy method, number of chemotherapy courses, start date of chemotherapy in the applied hospital, Declaration of hospital hormone treatment, declaration of start date of hospital hormone treatment, date of last contact or death, survival status, cancer status, date of first recurrence, type of first recurrence, cause of death, carcinoembryonic antigen test value, tumor shrinkage grade, pathological circumcision margin, nerve invasion, KRAS test value, intestinal obstruction found before or during surgery, intestinal perforation found before or during surgery, or any combination of the above. A method for establishing a clinical decision support system for colorectal cancer second primary cancer in a predictive model and visualization manner as described in claim 4, wherein the above-mentioned clinical tumor stage is clinical tumor stage T, clinical tumor stage N or clinical tumor Stage M. A method for establishing a clinical decision support system for second primary cancer in colorectal cancer using a predictive model and visualization method as described in claim 4, wherein the above-mentioned pathological tumor is pathological tumor stage T, pathological tumor stage N or pathological tumor stage M. A method for establishing a clinical decision support system for second primary cancer in colorectal cancer using a predictive model and visualization method as described in claim 1, wherein the influence value of each cancer characteristic is a Sharpe force value. A method for establishing a clinical decision support system for the occurrence of second primary cancers in colorectal cancer using a predictive model and visualization method as described in claim 7, wherein when the above-mentioned Sharpe force value is positive, the occurrence of second primary cancers in colorectal cancer The risk increases; when the above-mentioned Sharp power value is negative, the risk of colorectal cancer developing second primary cancer decreases. A method for establishing a clinical decision support system for second primary cancer in colorectal cancer using a predictive model and visualization method as described in claim 1, wherein the image is in the form of a bar graph.

Images