Construction method and application of postoperative II-III stage colorectal cancer long-term pulmonary metastasis prediction model

Abstract

The invention relates to a construction method and application of a postoperative II-III stage colorectal cancer long-term lung metastasis prediction model, and belongs to the technical field of machine learning. The construction method of the postoperative early colorectal cancer long-term pulmonary metastasis prediction model comprises the following steps: screening through machine learning and a logistic step-by-step backward regression method to obtain four indexes formed by combined prediction of postoperative early colorectal cancer long-term pulmonary metastasis, and presenting the prediction model through a combined diagnosis equation and a column diagram; the four indexes for predicting the long-term lung metastasis formation of the postoperative early colorectal cancer comprise baseline ctDNA, serum carcino-embryonic antigen, cancer nodule and plasma PIK3CA gene. The model constructed by the invention is the first model for predicting the long-term lung metastasis probability of the patient with the early colorectal cancer, and the ROC of the model is superior to that of each element in the model.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a method for constructing and applying a prediction model for long-term pulmonary metastasis of stage II-III colorectal cancer after surgery. Background Art

[0002] With the progress of diagnostic techniques and the improvement of the understanding of screening, the occurrence of tumors can be monitored earlier before distant metastasis of the tumor appears. Surgical operation is still the main treatment method for colorectal cancer patients. For colorectal cancer patients, adjuvant treatment is based on clinical and pathological risk stratification to guide treatment decisions. However, 15%-30% of colorectal cancer patients cannot benefit from adjuvant chemotherapy. Even after completing standard adjuvant treatment, recurrence or metastasis will still occur, and they are exposed to the toxic effects of chemotherapy. The high mortality rate of colorectal cancer patients is mainly due to distant metastasis. Pulmonary metastasis is the second most common metastatic site, which greatly affects long-term survival. Due to the relatively slow growth of lung lesions, the diagnosis of colorectal cancer lung metastasis faces challenges, and many suspicious lung nodules cannot be clearly determined as benign or malignant by conventional imaging, making early detection difficult. Currently, there is no reliable biomarker in the adjuvant treatment of colorectal cancer that can accurately predict the occurrence of lung metastasis.

[0003] Postoperative follow-up usually includes imaging, such as computed tomography and the serum oncology marker carcinoembryonic antigen (CEA). However, studies have shown that the sensitivity of carcinoembryonic antigen (CEA) in detecting recurrence is only 68% to 82%, and imaging examinations can only detect visible lesions, with low sensitivity to lung metastasis. Clinically, imaging and serum oncology are commonly used to predict lung metastasis nodules. CEA is an important tumor marker for the diagnosis, efficacy evaluation, and recurrence monitoring of colorectal cancer. A large number of studies have shown that the preoperative serum CEA level has significant prognostic value in both early and stage IV colorectal cancer, but CEA has a certain false positive and false negative rate, and its accuracy in recurrence monitoring is not high. A large part of suspicious lung nodules cannot be determined by imaging. Even using PET-CT for monitoring, for nodules with a diameter less than 10 mm, false negative results may also occur. Therefore, seeking a method for accurately predicting long-term pulmonary metastasis of early colorectal cancer after surgery has become an urgent technical problem in this field. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method for constructing and applying a prediction model for long-term pulmonary metastasis of stage II-III colorectal cancer after surgery.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0006] In a first aspect, the present invention provides a method for constructing a prediction model for distant lung metastasis of stage II-III colorectal cancer after surgery. The method for constructing the prediction model for distant lung metastasis of stage II-III colorectal cancer after surgery includes: screening four indicators for jointly predicting the formation of distant lung metastasis of stage II-III colorectal cancer after surgery through machine learning and logistic stepwise regression methods, and presenting the prediction model through a joint diagnostic equation and a nomogram; the four indicators for predicting the formation of distant lung metastasis of stage II-III colorectal cancer after surgery include: baseline ctDNA, serum carcinoembryonic antigen, cancer nodules, and plasma PIK3CA gene.

[0007] Preferably, the method includes the following steps:

[0008] Step 1: Obtain the data of patients meeting the screening criteria;

[0009] Step 2: Screen variables through Logistic forward stepwise regression analysis, and screen out variables with p < 0.10;

[0010] Step 3: Further screen variables for the variables screened in Step 2 through LASSO regression analysis to obtain the optimal factor combination for constructing the prediction model;

[0011] Step 4: Use multiple stepwise backward regression analysis to further screen influencing factors, screen out variables that have a significant impact on the outcome, establish a prediction model for distant lung metastasis of early-stage colorectal cancer after surgery through the screened variables, and present the model in the form of a nomogram; the variables include baseline ctDNA, serum carcinoembryonic antigen, cancer nodules, and plasma PIK3CA gene.

[0012] Preferably, the method further includes the following steps: evaluating the discrimination of the model through an ROC curve; evaluating the application value of the model through a clinical decision curve; evaluating the prediction ability and accuracy of the model through a calibration curve.

[0013] Preferably, the method further includes the following steps: internally validating the model through repeated sampling by the Bootstrap method.

[0014] In a second aspect, the present invention provides a prediction model constructed by the method for constructing a prediction model for distant lung metastasis of stage II-III colorectal cancer after surgery. The model is presented in the form of a nomogram, and the variables of the model are baseline ctDNA, CEA, cancer nodules, and plasma PIK3CA gene.

[0015] Preferably, in the nomogram, the scores corresponding to the variables and their scores, and the prediction probabilities corresponding to the total scores are as follows:

[0016] The variable baseline ctDNA includes the options of No and Yes; the score corresponding to the No option is: 0 points, and the score corresponding to the Yes option is: 60 points;

[0017] The variable CEA includes the options of content ≤ 5 ug / L and content > 5 ug / L; the scores corresponding to the option of content ≤ 5 ug / L are: 0 points, and the scores corresponding to the option of content > 5 ug / L are: 66 points;

[0018] The variable cancer nodules includes the options of No and Yes; the score corresponding to the No option is: 0 points, and the score corresponding to the Yes option is: 100 points;

[0019] The variable plasma PIK3CA includes the options of No and Yes; the score corresponding to the No option is: 0 points, and the score corresponding to the Yes option is: 95 points;

[0020] The total score corresponding to the predicted probability is determined as follows:

[0021] When the total score is 53 points, the predicted probability is 10%; when the total score is 112 points, the predicted probability is 20%; when the total score is 151 points, the predicted probability is 30%; when the total score is 183 points, the predicted probability is 40%; when the total score is 212 points, the predicted probability is 50%; when the total score is 241 points, the predicted probability is 60%; when the total score is 273 points, the predicted probability is 70%; when the total score is 312 points, the predicted probability is 80%.

[0022] Thirdly, the present invention provides a program storage medium for receiving user input, and the stored computer program enables an electronic device to execute the method for constructing a long-term lung metastasis prediction model for stage II-III colorectal cancer after surgery, which includes the following steps:

[0023] Step 1: Obtain the data of patients meeting the screening criteria;

[0024] Step 2: Screen variables through Logistic forward stepwise regression analysis, and screen out variables with p < 0.10;

[0025] Step 3: Further screen variables for the variables screened in Step 2 through LASSO regression analysis to obtain the best combination of elements for constructing the prediction model;

[0026] Step 4: Further screen influencing factors by using multiple stepwise regression analysis, screen out variables that have a significant impact on the outcome, establish a long-term lung metastasis prediction model for stage II-III colorectal cancer after surgery through the screened variables, and present the model in the form of a nomogram;

[0027] Step 5: Evaluate the discrimination of the model through the ROC curve; evaluate the application value of the model through the clinical decision curve; evaluate the prediction ability and accuracy of the model through the calibration curve;

[0028] Step 6: Repeated sampling by the Bootstrap method is used to perform internal validation on the model.

[0029] Fourthly, the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor is caused to execute the steps of the method for constructing a prediction model for long-term pulmonary metastasis of stage II-III colorectal cancer after surgery.

[0030] Fifthly, the present invention provides an information data processing terminal, which is used to implement the method for constructing a prediction model for long-term pulmonary metastasis of stage II-III colorectal cancer after surgery.

[0031] Sixthly, the present invention provides the application of the prediction model for long-term pulmonary metastasis of stage II-III colorectal cancer after surgery in predicting the probability of long-term pulmonary metastasis of early-stage colorectal cancer after surgery.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] The model constructed by the present invention is the first model for predicting the probability of long-term pulmonary metastasis in patients with stage II-III colorectal cancer. It contains ctDNA and clinicopathological risk factor variables. The accuracy of the model prediction is relatively high, and the elements of the prediction model can be obtained conveniently and quickly, several months earlier than imaging diagnosis. In the comparison of multiple models, the ROC of our model is better than each element in the model. Our model is presented in the form of a nomogram, which can more intuitively and conveniently prompt the high-risk population of long-term pulmonary metastasis clinically, and perform more rigorous clinical follow-up monitoring on this part of patients for timely intervention, which has great clinical significance. Description of the Drawings

[0034] Figure 1 Schematic diagram of elements obtained by LASSO regression analysis screening and CVLASSO verification for obtaining the best combination for constructing the model (Figure A is the LASSO screening diagram; Figure B is the CVLASSO diagram);

[0035] Figure 2 Nomogram of the prediction model for long-term pulmonary metastasis of early-stage colorectal cancer constructed by the present invention;

[0036] Figure 3 Schematic diagram of the results of external validation by ROC curve, clinical decision curve, and calibration curve (Figure A is the ROC curve of the training set; Figure B is the ROC curve of the validation set; Figure C is the DCA curve of the training set; Figure D is the DCA curve of the validation set; Figure D is the calibration curve of the training set; Figure E is the calibration curve of the validation set);

[0037] Figure 4Schematic diagram of the results of internal validation of the model by resampling using the Bootstrap method (Figure A shows the ROC curve for 500 Bootstrap resamplings; Figure B shows the ROC curve for each element in the model; Figure C shows the clinical impact CIC curve). Detailed implementation manners

[0038] To better illustrate the objectives, technical solutions and advantages of the present invention, the present invention will be further described below in conjunction with specific embodiments.

[0039] Embodiment 1

[0040] 1. Research subjects:

[0041] Patients diagnosed with colorectal cancer in the Sixth Affiliated Hospital of Sun Yat-sen University from January 2010 to December 2022, who had undergone radical surgery and had circulating tumor DNA (ctDNA) detection 2-8 weeks after surgery were collected.

[0042] 1.1 Inclusion criteria:

[0043] (1) Age ≥ 18 years old

[0044] (2) Histopathological diagnosis of colorectal adenocarcinoma;

[0045] (3) Primary colorectal tumor;

[0046] (4) According to the colon cancer TNM staging system of the Union for International Cancer Control (UICC) / American Joint Committee on Cancer (AJCC) (8th edition, 2017), the stage is II or III;

[0047] (5) Have undergone radical surgical treatment;

[0048] (6) ctDNA detection was performed 2-8 weeks after surgery and before adjuvant treatment.

[0049] 1.2 Exclusion criteria:

[0050] (1) Multiple primary malignant tumors;

[0051] (2) Incomplete surgical resection (non-R0 resection);

[0052] (3) Neoadjuvant treatment before surgery;

[0053] (4) Imaging or intraoperative confirmation of distant organ metastasis;

[0054] (5) Incomplete clinical information.

[0055] 2. Information collection

[0056] (1) Clinical information: including the age, gender, serum carcinoembryonic antigen (CEA), serum carbohydrate antigen 199 (CA199), serum carbohydrate antigen 125 (CA125), primary tumor location, plasma ctDNA results of the patient during surgery. The plasma ctDNA detection method is as follows:

[0057] ① Extract plasma-free DNA from peripheral blood samples;

[0058] ② Perform DNA fragmentation to construct a sequencing library;

[0059] ③ Amplify and purify the library fragments;

[0060] ④ Use the Illumina platform to perform chip-based sequencing-by-synthesis on the machine. All ctDNA results have passed quality control samples and process quality control. The effective sequencing depth of this panel is 10,000x, and the sensitivity is 0.25;

[0061] ⑤ Definition of ctDNA:

[0062] Definition of ctDNA positive: When at least one somatic mutation is detected in the ctDNA sample and class I or class II mutations are detected in the ctDNA, it is defined as ctDNA positive;

[0063] Definition of ctDNA negative: When class III mutations are detected in the ctDNA or no somatic variation is detected in the ctDNA sample, it is defined as ctDNA negative.

[0064] According to the Clinical Practice Expert Consensus on ctDNA High-Throughput Sequencing (2022 Edition) for the interpretation of the clinical significance of tumor somatic mutations, it is recommended to classify according to companion diagnosis, clinical guidelines, database or literature evidence. Tumor somatic gene mutations can be divided into those with clear clinical significance (grade I), potential clinical significance (grade II), unclear clinical significance (grade III), and benign or likely benign mutations (grade IV).

[0065] (2) Pathological information: Pathological T stage, pathological N stage, cancer nodules, vascular invasion, nerve invasion, pathological TNM after radical surgery;

[0066] (3) Gene mutation information: Next-generation sequencing (NGS) was used to detect mutant genes in pathological tissues, and the same NGS 88 genes (products of Guangzhou KingMed Diagnostics Co., Ltd.) as those for plasma ctDNA were used for gene mutation detection.

[0067] (4) Lung metastasis information: Two experienced senior physicians from the oncology department, pathology department, and radiology department each judged the patients for lung metastasis according to the Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1, and the first occurrence of lung metastasis, regardless of whether other organ metastases were present, was defined as lung metastasis.

[0068] 3. Statistical methods for variable screening

[0069] All statistical analyses were performed using R-4.3.1 software for data processing, statistical analysis, and visualization of results. The R software packages used were: survival, pROC, calibrate, MASS, rms, foreign, nricens, glmnet, Matrix, ggplot2, lattic, survex.

[0070] All factors were included in the univariate logistic regression analysis, as shown in Table 1:

[0071] Table 1: Univariate Logistic regression related to lung metastasis

[0072]

[0073]

[0074] Note: The primary tumor site was divided into the left and right halves. The left half included the splenic flexure of the colon, descending colon, sigmoid colon, and rectum, and the right half included the cecum, ascending colon, hepatic flexure, and transverse colon. Tissue MSI was MSI (microsatellite instability); OR was the odds ratio; CI was the confidence interval.

[0075] According to Table 1 above, variables with p < 0.10 (including baseline ctDNA, baseline serum CEA, cancer nodules, pathological TNM stage, tissue TP53 mutation, tissue BRAF mutation, plasma TP53 mutation, plasma PIK3CA mutation, plasma KRAS mutation, plasma FBXW7 mutation, and plasma BRAF mutation, a total of 11) were included in the Least Absolute Shrinkage and Selection Operator (LASSO) regression ( Figure 1A); Perform 10-fold cross-validation (CVLASSO), and select the best combination of model elements based on the minimum lambda. A total of 5 factors were included, namely baseline ctDNA, baseline CEA level, cancer nodules, plasma KRAS mutation, and plasma PIK3CA mutation ( Figure 1 B).

[0076] Multivariate stepwise regression analysis was used to further screen the influencing factors. The Akaike information criterion (AIC) was used to screen out the variables that had a significant impact on the outcome, and a model was constructed. The final model included 4 variables: baseline ctDNA, CEA, cancer nodules, and plasma PIK3CA mutation.

[0077] 4. Construction of a nomogram prediction model

[0078] To facilitate clinical visualization, we used the rms package in R software to construct a nomogram prediction model. By adding the scores of these 4 variables and drawing a vertical line according to the points score, the probability of lung metastasis of the patient can be obtained ( Figure 2 ).

[0079] The corresponding scores for each variable are as follows:

[0080] The scores corresponding to baseline ctDNA (No, Yes) are 0 and 60 points respectively;

[0081] The scores corresponding to CEA (content ≤ 5 μg / L, content > 5 μg / L) are 0 and 66 points respectively;

[0082] The scores corresponding to cancer nodules (No, Yes) are 0 and 100 points respectively;

[0083] The scores corresponding to plasma PIK3CA (No, Yes) are 0 and 95 points respectively.

[0084] The total score corresponds to the predicted probability:

[0085] The predicted probability for 53 points is 10%; the predicted probability for 112 points is 20%; the predicted probability for 151 points is 30%; the predicted probability for 183 points is 40%; the predicted probability for 212 points is 50%; the predicted probability for 241 points is 60%; the predicted probability for 273 points is 70%; the predicted probability for 312 points is 80%.

[0086] 5. Model validation

[0087] We used the clinical data of different ethnic groups in Spanish hospitals (including the age, gender, serum carcinoembryonic antigen (CEA), serum carbohydrate antigen 199 (CA199), serum carbohydrate antigen 125 (CA125), primary tumor location, histopathological conditions, tissue gene mutation information, and plasma ctDNA results of patients during surgery) for external validation. The results are as Figure 3 shown in Figures A and B. The area under the curve (AUC) of the ROC curve of the model is 0.775, and the Hosmer-Lemeshow test, p = 0.967, indicating that the model has a good fit. The AUC of the validation set is 0.724, indicating that our model has good discrimination. The clinical decision curve (decision curve analysis, DCA) and calibration curve were used to evaluate the clinical net benefit and predictive ability of the model. The results are as Figure 3 shown in Figures C and D. The DCA curves of both cohorts showed that for patients within a threshold probability of 80%, the model provided a greater net benefit than either treating all or treating none. The results are as Figure 3 shown in Figures E and F. The calibration curves of both cohorts were also near the diagonal.

[0088] We also performed 500 internal validations using Bootstrap. The results are as Figure 4 shown in Figure A. The concordance index (C-index) is 0.726, indicating that our model has good discrimination. We also compared the ROC of each element within the model. The results are as Figure 4 shown in Figure B. The ROC area of our overall model is better than that of each element in the model. We also plotted the clinical impact curve (CIC). The results are as Figure 4 shown in Figure C, showing the clinical efficiency of the prediction model.

[0089] In summary, the present invention constructs a model that can predict the long-term probability of lung metastasis in patients with early colorectal cancer. Within a relatively wide and practical threshold probability range (probability from 10% to 80%, total score from 53 points to 312 points), the population determined to have lung metastasis by the prediction model of the present invention highly matches the population with actual lung metastasis, indicating that the lung metastasis prediction model has significant predictive value.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the protection scope of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A method for constructing a prediction model for long-term lung metastasis of postoperative stage II-III colorectal cancer, characterized in that: The method for constructing a prediction model for long-term lung metastasis of postoperative stage II-III colorectal cancer comprises: screening four indicators for jointly predicting the formation of long-term lung metastasis of postoperative stage II-III colorectal cancer by machine learning and logistic stepwise regression method, and presenting the prediction model by a combined diagnostic equation and a nomogram; The four indicators for predicting the formation of long-term lung metastasis of stage II-III colorectal cancer after surgery include: baseline ctDNA, serum carcinoembryonic antigen, cancer nodules and plasma PIK3CA gene.

2. The construction method according to claim 1, characterized in that: The method comprises the following steps: Step 1: Obtain data of patients who meet the screening criteria; the patients are colorectal cancer patients who have undergone radical surgery and have undergone circulating tumor DNA testing 2-8 weeks after surgery; Step 2: Screen variables through logistic forward stepwise regression analysis and select variables with p < 0.10; Step 3: The variables screened in step 2 are further screened through LASSO regression analysis to obtain the best combination of factors for constructing the prediction model; Step 4: Multivariate stepwise backward regression analysis was used to further screen the influencing factors and select the variables that have a significant impact on the outcome. A prediction model for long-term lung metastasis of early postoperative colorectal cancer was established based on the selected variables, and the model was presented in the form of a nomogram; the variables included baseline ctDNA, serum carcinoembryonic antigen, cancer nodules, and plasma PIK3CA gene.

3. The construction method according to claim 2, characterized in that: The method also includes the following steps: evaluating the discrimination of the model through the ROC curve; evaluating the application value of the model through the clinical decision curve; and evaluating the predictive ability and accuracy of the model through the calibration curve.

4. The construction method according to claim 3, characterized in that: The method also includes the following steps: performing internal validation on the model by repeated sampling using the Bootstrap method.

5. A prediction model constructed by the method for constructing a prediction model for long-term lung metastasis of postoperative stage II-III colorectal cancer according to any one of claims 1 to 4, characterized in that: The model was presented in the form of a nomogram, and the variables of the model were baseline ctDNA, CEA, cancer nodules, and plasma PIK3CA gene.

6. The prediction model according to claim 5, characterized in that The scores of the variables and their scores in the nomogram, as well as the predicted probabilities corresponding to the total scores, are as follows: The variable baseline ctDNA includes No and Yes options; the score corresponding to the No option is 0 points, and the score corresponding to the Yes option is 60 points; The variable CEA includes the options of content ≤5ug / L and content >5ug / L; the corresponding scores of the option of content ≤5ug / L are 0 points, and the corresponding scores of the option of content >5ug / L are 66 points; The variable cancer nodules includes No and Yes options; the score corresponding to the No option is 0 points, and the score corresponding to the Yes option is 100 points; The variable plasma PIK3CA includes No and Yes options; the score corresponding to the No option is 0 points, and the score corresponding to the Yes option is 95 points; The predicted probability corresponding to the total score is determined as follows: The predicted probability of a total score of 53 points is 10%; the predicted probability of a total score of 112 points is 20%; the predicted probability of a total score of 151 points is 30%; the predicted probability of a total score of 183 points is 40%; the predicted probability of a total score of 212 points is 50%; the predicted probability of a total score of 241 points is 60%; the predicted probability of a total score of 273 points is 70%; and the predicted probability of a total score of 312 points is 80%.

7. A program storage medium for receiving user input, characterized in that: The stored computer program enables the electronic device to execute the method for constructing a prediction model for long-term lung metastasis of postoperative stage II-III colorectal cancer as described in any one of claims 1 to 4, comprising the following steps: Step 1: Obtain data on patients who meet the screening criteria; Step 2: Screen variables through logistic forward stepwise regression analysis and select variables with p < 0.10; Step 3: The variables screened in step 2 are further screened through LASSO regression analysis to obtain the best combination of factors for constructing the prediction model; Step 4: Multiple stepwise regression analysis was used to further screen the influencing factors and select the variables that have a significant impact on the outcome. A prediction model for long-term lung metastasis of stage II-III colorectal cancer after surgery was established based on the selected variables, and the model was presented in the form of a nomogram. Step 5: Evaluate the discrimination of the model through the ROC curve; evaluate the application value of the model through the clinical decision curve; evaluate the predictive ability and accuracy of the model through the calibration curve; Step 6: Perform internal validation of the model through repeated sampling using the Bootstrap method.

8. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method for constructing a long-term lung metastasis prediction model for postoperative stage II-III colorectal cancer as described in any one of claims 1-4.

9. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the method for constructing a prediction model for long-term lung metastasis of postoperative stage II-III colorectal cancer as described in any one of claims 1-4.

10. Use of the model for predicting long-term lung metastasis of postoperative stage II-III colorectal cancer as described in claim 5 or 6 in predicting the probability of long-term lung metastasis of postoperative stage II-III colorectal cancer.

Images