Method for predicting the risk of tumor re-progression after transcatheter arterial chemoembolization

Abstract

The present invention belongs to the field of biological technology, and specifically relates to a method for predicting the risk of tumor re-progression after hepatic artery embolization chemotherapy. The method comprises determining core characteristic variables in collected data, and using a 0-8 week data set, an 8-12 week data set, a 12-16 week data set, and a >16 week data set as training sets, performing data enhancement processing, and respectively training neural networks to obtain a first cycle prediction network, a second cycle prediction network, a third cycle prediction network, and a fourth cycle prediction network; inputting the data to be tested into four prediction networks respectively to obtain corresponding probabilities, selecting the maximum value among the four probability values ​​as a prediction result, and the corresponding time period as a prediction period of tumor re-progression; thereby realizing individualized risk prediction of tumor re-progression, so as to facilitate the formulation of an individualized plan.

Description

Technical Field

[0001] The present invention belongs to the field of biological technology, and in particular relates to a method for predicting the risk of tumor re-progression after hepatic artery chemoembolization. Background Art

[0002] Primary liver cancer (hereinafter referred to as liver cancer) has an insidious onset, rapid progression, and poor prognosis. Although the promotion of liver cancer screening and the progress of early treatment mainly based on surgery in recent years have enabled more patients to have the opportunity to be cured of liver cancer, many patients are still in the middle and late stages when they seek medical treatment and have underlying liver diseases, thus losing the opportunity for surgical resection. Currently, transarterial chemoembolization (TACE), as a local treatment method, has achieved significant results in patients with middle and late stage unresectable liver cancer, and patients have achieved complete or partial remission of tumor treatment. Therefore, it has been recommended by many authoritative guidelines as the priority treatment for such patients. However, given that these patients are often accompanied by underlying liver diseases such as cirrhosis, the re-progression and recurrence of liver cancer after local treatment has become the main cause of tumor treatment failure.

[0003] In addition, the researchers mainly focused on liver cancer patients undergoing surgery to explore the prediction of recurrence and survival after surgical resection of liver cancer. For example, the SLICER score, CLIP score, and ERASL's early postoperative recurrence model for liver tumors can all predict the risk of tumor recurrence after resection in a specific group of liver cancer patients, and have good diagnostic efficacy. However, for patients with advanced liver cancer who cannot be surgically removed, there are significant differences in the underlying causes of liver disease, pathological characteristics of tumor staging, nutritional and immune status of the body, and tumor treatment methods compared with patients with surgically resectable liver cancer. In addition, although the tumor has been inactivated after local treatment, it still exists in the body and can affect tumor progression and recurrence.

[0004] Therefore, exploring accurate prediction methods for tumor progression and recurrence after hepatic arterial chemoembolization will help to detect the progression and recurrence of liver cancer patients as early as possible and provide timely treatment. At the same time, it has important clinical significance for formulating individualized diagnosis and treatment plans for patients, improving the overall treatment effect of patients, and prolonging the survival of patients. Summary of the invention

[0005] The present invention provides a method for predicting the risk of tumor re-progression after hepatic artery chemoembolization.

[0006] The technical solution of the present invention is as follows:

[0007] The present invention provides a method for predicting the risk of tumor re-progression after hepatic artery chemoembolization, comprising the following steps:

[0008] S1: Collect the core characteristic variable data of patients undergoing transcatheter arterial chemoembolization at four time stages before surgery: 0-8 weeks, 8-12 weeks, 12-16 weeks, and >16 weeks, and construct the 0-8 week data set, 8-12 week data set, 12-16 week data set, and >16 week data set respectively;

[0009] S2: The 0-8 week data set, the 8-12 week data set, the 12-16 week data set, and the >16 week data set are used as training sets. After data enhancement processing, the neural networks are trained respectively to obtain the first cycle prediction network, the second cycle prediction network, the third cycle prediction network, and the fourth cycle prediction network;

[0010] S3: input the data to be tested into the first cycle prediction network, the second cycle prediction network, the third cycle prediction network, and the fourth cycle prediction network, respectively, and obtain the probability of tumor re-progression in the first cycle, the probability of tumor re-progression in the second cycle, the probability of tumor re-progression in the third cycle, and the probability of tumor re-progression in the fourth cycle, respectively, and select the maximum value of the four probability values ​​as the prediction result, and the corresponding time period as the prediction period of tumor re-progression;

[0011] The core characteristic variables in S1 are age, alkaline phosphatase, alanine aminotransferase, glutamyl transpeptidase, aspartate aminotransferase, total bilirubin, direct bilirubin, hemoglobin, prothrombin normalized ratio, platelet count, white blood cell count, alpha-fetoprotein, red blood cell count, tumor number and tumor diameter.

[0012] In S2 of the present invention, the 0-8 week data set, the 8-12 week data set, the 12-16 week data set, and the >16 week data set are used as training sets to train the neural network respectively. During the training process, each core feature variable has different weights in the training process of different time periods. The specific weights are shown in the table below.

[0013]

[0014] After the prediction cycle of tumor re-progression in S3 described in the present invention, it also includes decomposing the prediction results according to the SHAP algorithm to obtain the contribution value of each core characteristic variable to tumor re-progression in the prediction cycle, and according to the contribution value of tumor re-progression, the probability of tumor re-progression in the first cycle, the probability of tumor re-progression in the second cycle, the probability of tumor re-progression in the third cycle, and the probability of tumor re-progression in the fourth cycle are obtained.

[0015] After constructing the 0-8 week data set, 8-12 week data set, 12-16 week data set, and >16 week data set in S1 of the present invention, the data in the data set are cleaned, specifically by deleting collinear features, features of patients with missing baseline and follow-up data, and filling the missing data with the median.

[0016] The neural network in S2 described in the present invention is random forest, extreme gradient boosting, lightweight gradient boosting machine, and classification enhancement.

[0017] In S1 of the present invention, core feature variable data is collected, specifically,

[0018] Data were collected, including preoperative baseline characteristic variable data, characteristic variable data at complete / near-complete remission of the tumor, change in characteristic variable data, and endpoint events of tumor progression and progression-free survival;

[0019] Recursive feature elimination cross-validation was used to screen the core feature variables of the collected data.

[0020] The progression-free survival period described in the present invention starts from the time when complete / near-complete remission is achieved after hepatic artery chemoembolization and ends when liver cancer progression occurs.

[0021] The present invention also provides a system for predicting the risk of tumor re-progression after hepatic artery chemoembolization, comprising:

[0022] Dataset construction module: used to collect core characteristic variable data of patients undergoing transcatheter arterial chemoembolization at four time stages before surgery: 0-8 weeks, 8-12 weeks, 12-16 weeks, and >16 weeks, and to construct the 0-8 week dataset, 8-12 week dataset, 12-16 week dataset, and >16 week dataset respectively;

[0023] The core characteristic variables are age, alkaline phosphatase, alanine aminotransferase, glutamyl transpeptidase, aspartate aminotransferase, total bilirubin, direct bilirubin, hemoglobin, prothrombin normalized ratio, platelet count, white blood cell count, alpha-fetoprotein, red blood cell count, tumor number and tumor diameter;

[0024] Neural network training module: used to use the 0-8 week data set, 8-12 week data set, 12-16 week data set, and >16 week data set as training sets, perform data enhancement processing, and then train the neural network respectively to obtain the first cycle prediction network, the second cycle prediction network, the third cycle prediction network, and the fourth cycle prediction network;

[0025] Prediction module: used to input the data to be tested into the first cycle prediction network, the second cycle prediction network, the third cycle prediction network, and the fourth cycle prediction network, respectively, to obtain the probability of tumor re-progression in the first cycle, the probability of tumor re-progression in the second cycle, the probability of tumor re-progression in the third cycle, and the probability of tumor re-progression in the fourth cycle, respectively, and select the maximum value of the four probability values ​​as the prediction result, and the corresponding time period as the prediction period of tumor re-progression.

[0026] The present invention also provides a device for predicting the risk of tumor re-progression after hepatic arterial chemoembolization, comprising a processor and a memory, wherein the processor implements the method for predicting the risk of tumor re-progression after hepatic arterial chemoembolization when executing a computer program stored in the memory.

[0027] The present invention also provides a storage medium for predicting the risk of tumor re-progression after hepatic arterial chemoembolization, which is used to store a computer program, wherein the computer program, when executed by a processor, implements the method for predicting the risk of tumor re-progression after hepatic arterial chemoembolization.

[0028] Beneficial Effects

[0029] The method for predicting the risk of tumor re-progression provided by the present invention can predict the risk of tumor re-progression after hepatic artery chemoembolization.

[0030] The tumor re-progression risk prediction method provided by the present invention can not only distinguish the population with tumor progression at different time stages, but also obtain the probability of tumor progression at different time stages after hepatic artery chemoembolization, thereby realizing individualized risk prediction of tumor re-progression, so as to formulate individualized plans;

[0031] The tumor re-progression risk prediction method provided by the present invention can also provide a weight and contribution of each core characteristic variable to the prediction result during tumor progression at different time stages, so as to provide a logical explanation for the generation of the prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a screening diagram for the core characteristic variables predicting tumor progression using RFECV technology.

[0033] Figure 2 ROC curves of the prediction efficiency of tumor re-progression of different models at 0-8 weeks, 8-12 weeks, 12-16 weeks, and >16 weeks, including CatBoost model (A), LightGBM model (B), XGBoost model (C), and RF model (D).

[0034] Figure 3 PR curve (A) and confusion matrix diagram (B) of different models in the SHPI test set.

[0035] Figure 4 KM curves for the CatBoost model to distinguish the patients with tumor re-progression in weeks 0-8 (A), 0-12 (B), 0-16 (C), and 0-8, 8-12, 12-16, and >16 (D).

[0036] Figure 5The ROC curves of the CatBoost model for predicting tumor progression in the internal validation set cohort, the external validation set cohort I, and the external validation set cohort II at 0-8 weeks (A, B, C), 8-12 weeks (D, E, F), 12-16 weeks (G, H, I), and >16 weeks (J, K, L) respectively.

[0037] Figure 6 SHAP graphs of the CatBoost model at weeks 0-8 (A), 8-12 (B), 12-16 (C), and >16 (D), and LIME graphs of patients whose tumor re-progression occurred at weeks 0-8 (E) and >16 (F).

[0038] Figure 7 This is a schematic diagram of the web version of the CatBoost model's visualization application interface. DETAILED DESCRIPTION

[0039] The following examples are intended to illustrate the present invention rather than to further limit the present invention.

[0040] The present invention provides a method for predicting the risk of tumor re-progression after transcatheter arterial chemoembolization (TACE), comprising the following steps:

[0041] S1: Collect the core characteristic variable data of patients undergoing transcatheter arterial chemoembolization at four time stages before surgery: 0-8 weeks, 8-12 weeks, 12-16 weeks, and >16 weeks, and construct the 0-8 week data set, 8-12 week data set, 12-16 week data set, and >16 week data set respectively;

[0042] S2: The 0-8 week data set, the 8-12 week data set, the 12-16 week data set, and the >16 week data set are used as training sets. After data enhancement processing, the neural networks are trained respectively to obtain the first cycle prediction network, the second cycle prediction network, the third cycle prediction network, and the fourth cycle prediction network;

[0043] S3: Input the data to be tested into the first cycle prediction network, the second cycle prediction network, the third cycle prediction network, and the fourth cycle prediction network, respectively, and obtain the probability of tumor re-progression in the first cycle, the probability of tumor re-progression in the second cycle, the probability of tumor re-progression in the third cycle, and the probability of tumor re-progression in the fourth cycle, respectively. Select the maximum value of the four probability values ​​as the prediction result, and the corresponding time period as the prediction period of tumor re-progression.

[0044] The tumor re-progression risk prediction method provided by the present invention can not only distinguish the population whose tumors progress at different time stages, but also obtain the probability of tumor progression at different time stages after hepatic artery embolization chemotherapy, thereby realizing individualized risk prediction of tumor re-progression, so as to formulate individualized plans.

[0045] S1: The core characteristic variable data of patients undergoing transcatheter arterial chemoembolization were collected at four time stages before surgery: 0-8 weeks, 8-12 weeks, 12-16 weeks, and >16 weeks. The 0-8 week dataset, 8-12 week dataset, 12-16 week dataset, and >16 week dataset were constructed respectively.

[0046] In order to ensure the accuracy of the prediction method, it is necessary to determine the comprehensiveness of the research object and data collection. Preferably, core characteristic variable data is collected in S1, specifically,

[0047] Data were collected, including preoperative baseline characteristic variable data (including general demographic characteristics, blood routine, liver function, coagulation function, and tumor pathological characteristics), characteristic variable data at the time of complete / near-complete remission of the tumor, the change in characteristic variable data, and endpoint events of tumor progression and progression-free survival;

[0048] Recursive feature elimination cross-validation was used to screen the core feature variables of the collected data.

[0049] The details are as follows:

[0050] 1. Enrollment of research subjects and establishment of prospective research cohort

[0051] 1.1. Inclusion criteria of the research subjects

[0052] 1) Aged between 15 and 75 years old;

[0053] 2) The patient's liver function is classified as Child-Pugh grade A or B;

[0054] 3) ECOG score ≤2;

[0055] 4) According to the diagnostic criteria of the "Guidelines for the Diagnosis and Treatment of Primary Liver Cancer (2019 Edition)", patients with primary liver cancer are diagnosed based on histopathology or typical CT / MR imaging manifestations and are patients who have just been diagnosed and treated;

[0056] 5) According to the "Guidelines for the Diagnosis and Treatment of Primary Liver Cancer (2019 Edition)", the patient is staged as Ib to IIIb and is recommended to receive TACE as the first-line treatment according to the guidelines;

[0057] 1.2 Exclusion criteria for study subjects

[0058] 1) Patients with contraindications to TACE treatment: such as ECOG score 3-4 points, Child-Pugh grade C or D; complete embolism of the portal vein; combined coagulation disorder (PTA < 40%); diffuse liver cancer or distant metastasis; severe allergy to iodine contrast agents.

[0059] 2) Rupture of primary liver cancer;

[0060] 3) receiving surgical resection or liver transplantation after TACE;

[0061] 4) The tumor continued to progress during TACE treatment, and no complete / near-complete remission occurred;

[0062] 5) At the end of the study, no endpoint event of liver cancer progression occurred;

[0063] 6) Combined with other systemic malignant tumors;

[0064] 7) Those with incomplete baseline and follow-up data.

[0065] 1.3. Establishment of a prospective study cohort

[0066] 1.3.1. Patient enrollment:

[0067] like Figure 1As shown in the results, a multicenter prospective study on the prediction of tumor re-progression in patients with unresectable HCC (CNLC stage Ib-IIIb) after TACE treatment was first initiated on December 1, 2011 at two medical centers: the Department of Infectious Diseases (SPHI) of Shandong Provincial Hospital Affiliated to Shandong First Medical University and the Department of Gastroenterology (SPHG) of Shandong Provincial Hospital Affiliated to Shandong First Medical University. In addition, the above prospective study was initiated on May 1, 2012 and May 1, 2014 at two medical centers: the Department of Infectious Diseases (LCPH) of Liaocheng People's Hospital and the Department of Intervention (SPPH) of Shandong Provincial Public Health Clinical Center. The four medical centers finally completed the screening and enrollment of all patients on March 31, 2023, and the follow-up of all enrolled patients was terminated on August 30, 2023. The research methods, inclusion and exclusion criteria and procedures of the study population involved in this study all complied with relevant ethical standards and were approved by the Medical Ethics Committee of the Provincial Hospital Affiliated to Shandong First Medical University, Shandong Provincial Public Health Clinical Center, and Liaocheng People's Hospital.

[0068] After patients were enrolled according to the inclusion criteria, their baseline clinical data were collected, including: (1) demographic information: age, gender, and presence of ascites; (2) laboratory examination indicators within 2 weeks before the first TACE treatment: including peripheral blood cell classification count, liver function indicators, coagulation function, liver cancer tumor markers, hepatitis B virus quantitative detection, etc.; (3) liver imaging (enhanced CT / MRI) examination within 1 month before surgery, recording tumor diameter, tumor number, presence of lymph node metastasis, and presence of portal vein tumor thrombus, to obtain the baseline pathological characteristics of the tumor.

[0069] 1.3.2. Follow-up of patients:

[0070] After receiving TACE series treatment, the enrolled patients will be followed up regularly. The follow-up time and content are as follows:

[0071] (1) Efficacy evaluation after TACE series treatment: Enhanced CT / MR imaging examinations are used to evaluate the efficacy according to the mRECIST standard. The mRECIST standard is as follows:

[0072] Complete Remission (CR): All target lesions (enhanced imaging in the arterial phase) disappear;

[0073] Partial response (PR): The direct sum of target lesions (enhanced imaging in the arterial phase) is reduced by ≥30%;

[0074] Near Complete Remission (nCR): Based on the characteristics of the patients in this study and combined with clinical experience, it will be impossible to block the remaining tumor blood supply with improved technology, which is defined as near complete remission.

[0075] Progressive Disease (PD): The direct sum of target lesions (enhanced imaging in the arterial phase) increases by ≥20% or new lesions appear;

[0076] Stable Disease (SD): The direct sum of target lesions (enhanced imaging in the arterial phase) has not shrunk to PR and has not increased to PD.

[0077] (2) During TACE treatment, the treatment effect of the patient was evaluated according to the mRECIST standard. After achieving complete or near-complete remission, follow-up examinations were performed every 4 to 8 weeks until tumor progression occurred. If the patient still had no tumor progression after 16 weeks, follow-up examinations were performed every 12 weeks until tumor progression occurred.

[0078] 1.3.3. Study endpoint events and progression-free survival (PFS) calculation:

[0079] The endpoint event is defined as: tumor progression (PD) in patients with unresectable liver cancer after TACE treatment. The time to achieve CR / nCR during treatment is used as the starting point for calculation, and the time when patients develop PD during follow-up is used as the endpoint for calculation. The period is defined as progression-free survival, and patients who have not developed PD by the end of the study are treated as censored values.

[0080] Finally, a total of 973 patients with liver cancer who received serial TACE treatment were screened in the four medical centers. According to the patient inclusion criteria and follow-up requirements, four prospective study cohorts (n=483) were established, including: training cohort (SHPI, n=210), internal validation cohort (SHPG, n=102), external validation cohort I (SPPH, n=104) and external validation cohort II (LCPH, n=67).

[0081] After the data is collected, it is necessary to further process and screen the data. Preferably, after constructing the 0-8 week data set, the 8-12 week data set, the 12-16 week data set, and the >16 week data set in S1, the data in the data set is cleaned, specifically deleting data with strong collinearity, data with more missing baselines and follow-up, and filling some missing data with the median. Finally, 13 patients with missing baselines and 126 patients with missing follow-up data were deleted, which is close to 14% (139 / 973) of the samples, and 27 baseline characteristics were obtained, as shown in Tables 1, 2, and 3. The average follow-up time for all liver cancer patients who received TACE series treatment was 35.85 (interquartile range IQR 22.71, 71.00) weeks, and the average time of tumor progression was 19.71 (IQR11.86, 43.43) weeks.

[0082] Table 1 Preoperative baseline characteristic variables of the dataset

[0083]

[0084]

[0085] Note: *Tumor diameter: the sum of the maximum diameters of each tumor.

[0086] Table 2 Characteristic variable data of complete / near complete remission of tumor in data set

[0087]

[0088] Table 3 Changes in characteristic variable data before and after treatment of the dataset

[0089]

[0090]

[0091] Note: △ represents the change between the index at complete / near complete remission and the baseline index.

[0092] In order to ensure that the core features in the constructed data set are more accurate, preferably, the core feature variables in S1 are age, alkaline phosphatase, alanine aminotransferase, glutamyl transpeptidase, aspartate aminotransferase, total bilirubin, direct bilirubin, hemoglobin, prothrombin normalized ratio, platelet count, white blood cell count, alpha-fetoprotein, red blood cell count, tumor number and tumor diameter. Specifically,

[0093] A total of 27 baseline features were extracted and screened for liver cancer patients who received TACE series treatment. The principle of obtaining the best prediction performance of liver cancer progression with the minimum number of features was used, and the recursive feature elimination cross-validation (RFECV) technology was used to screen the core features, reduce the unimportant features in the data set, and find the optimal number of features, such as Figure 1 As shown, a total of 15 core feature variables are obtained and used as the feature basis for building the prediction network.

[0094] S2: The 0-8 week data set, 8-12 week data set, 12-16 week data set, and >16 week data set are used as training sets. After data enhancement processing, the neural networks are trained respectively to obtain the first cycle prediction network, the second cycle prediction network, the third cycle prediction network, and the fourth cycle prediction network.

[0095] In addition, the weights of the core feature variables are set. In S2, the 0-8 week data set, the 8-12 week data set, the 12-16 week data set, and the >16 week data set are used as training sets to train the neural network respectively. During the training process, each core feature variable has different weights in the training process of different time periods. The specific weights are shown in Table 4.

[0096] Table 4 Weights of core characteristic variables in different time periods

[0097]

[0098] Furthermore, in order to ensure that the prediction network can better identify data, it is necessary to reasonably divide the data set for training and testing. After the training set is subjected to data enhancement processing in S2, the training set is further divided into a new training set and a test set in a ratio of 7:3. Specifically:

[0099] The training set cohort (SPHI, n = 210) was expanded to 272 cases through data augmentation to achieve classification balance, and then divided into two independent subsets in a ratio of 7:3: a new training set (n = 190) and a test set (n = 82).

[0100] Specifically, after the training set is reasonably divided, the new training set is divided into 10 subsets, of which 9 subsets are used to train the prediction network and 1 subset is used to verify the prediction network. In each round of training of the prediction network, 9 subsets are used for training and 1 subset is used for verification.

[0101] After each round of training, parameter tuning was performed, and a total of 10 rounds of training were performed.

[0102] In addition, the neural networks in S2 are Random Forest, Extreme Gradient Boosting, Light Gradient Boosting Machine, and Categorical Boosting.

[0103] S3: Input the data to be tested into the first cycle prediction network, the second cycle prediction network, the third cycle prediction network, and the fourth cycle prediction network, respectively, and obtain the probability of tumor re-progression in the first cycle, the probability of tumor re-progression in the second cycle, the probability of tumor re-progression in the third cycle, and the probability of tumor re-progression in the fourth cycle, respectively. Select the maximum value of the four probability values ​​as the prediction result, and the corresponding time period as the prediction period of tumor re-progression.

[0104] In addition, after the prediction cycle of tumor re-progression in S3, the prediction results are decomposed according to the SHAP algorithm to logically explain the generation of the prediction results, obtain the contribution value of each core characteristic variable to tumor re-progression in the prediction cycle, and obtain the probability of tumor re-progression in the first cycle, the probability of tumor re-progression in the second cycle, the probability of tumor re-progression in the third cycle, and the probability of tumor re-progression in the fourth cycle according to the contribution value of tumor re-progression. In this way, individualized risk prediction of tumor re-progression is achieved to facilitate the formulation of individualized plans.

[0105] Experimental Results

[0106] The method for predicting the risk of tumor re-progression after transcatheter arterial chemoembolization (TACE) provided in this application was used, and four models, namely, random forest (RF), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM) and classification boosting (CategoricalBoosting, CatBoost) were selected for experiments. For the sake of simplicity, the relevant terms are abbreviated as shown in Table 5.

[0107] Table 5 Terms and abbreviations

[0108]

[0109] 1. Model evaluation and selection

[0110] ROC curves were used to evaluate the prediction efficiency of tumor re-progression of four models: CatBoost, LightGBM, XGBoost, and RF in four time periods: 0-8 weeks, 8-12 weeks, 12-16 weeks, and >16 weeks. The results are as follows: Figure 2 and as shown in Table 6. The area under the curve (AUC) values ​​of the CatBoost model were: 0.89 (95% CI: 0.80-0.99), 0.74 (95% CI: 0.61-0.88), 0.90 (95% CI: 0.79-1.01), 0.94 (95% CI: 0.87-1.00) ( Figure 2 ,A); the AUC values ​​of the LightGBM model were: 0.84 (95% CI: 0.74-0.95), 0.67 (95% CI: 0.52-0.82), 0.81 (95% CI: 0.67-0.95), 0.87 (95% CI: 0.77-0.96) ( Figure 2 , B); the AUC values ​​of the XGBoost model were: 0.84 (95% CI: 0.73-0.95), 0.68 (95% CI: 0.53-0.83), 0.77 (95% CI: 0.61-0.92), 0.90 (95% CI: 0.82-0.98) ( Figure 2 , C); the AUC values ​​of the RF model were: 0.78 (95% CI: 0.66-0.91), 0.61 (95% CI: 0.46-0.76), 0.78 (95% CI: 0.63-0.93), 0.93 (95% CI: 0.86-1.00) ( Figure 2 ,D). Therefore, compared with LightGBM, XGBoost and RF models, the CatBoost model using the prediction method of the present application has the highest AUC value in the four time stages of 0-8 weeks, 8-12 weeks, 12-16 weeks, and >16 weeks.

[0111] Table 6 AUC values ​​of different models in the four time periods of 0-8 weeks, 8-12 weeks, 12-16 weeks, and >16 weeks

[0112]

[0113] The model was further evaluated using the test set cohort (SPHI, n=82). As shown in Table 7: the accuracy, precision, recall and F1 score of the CatBoost model are 0.900, 0.898, 0.890, and 0.890, respectively; the above indicators of the LightGBM model are 0.835, 0.837, 0.827, and 0.829, respectively. In addition, the above indicators of the XGBoost and RF models are 0.789, 0.789, 0.782, 0.781 and 0.633, 0.654, 0.631, and 0.623, respectively. It can be seen that the CatBoost model and LightGBM model using the prediction method of this application perform well.

[0114] In addition, the average precision (AP) is used to evaluate the precision and recall of the model. The results show that the AP value of the CatBoost model can reach 0.97, and the AP value of the LightGBM model is 0.91. The AP values ​​of the XGBoost and RF models are 0.86 and 0.63 respectively. Figure 3 ,A,Table 7). Therefore, considering multiple evaluation indicators, the CatBoost model is the best model. Figure 3 As shown in Figure ,B, in order to further intuitively demonstrate the classification effect, the confusion matrix diagram of the CatBoost model in the test set (SPHI, n=82) is also provided, which more intuitively shows the prediction effect of the CatBoost model in the four aspects of true positive, true negative, false positive, and false negative.

[0115] Table 7 Accuracy, precision, recall, F1 score and AP value of different models

[0116]

[0117]

[0118] 2. Stratified analysis

[0119] The KM curve was used to perform a stratified analysis of the risk of tumor re-progression in liver cancer patients receiving TACE series treatment. The results showed that the CatBoost model, using the prediction method of this application, was able to distinguish between patients with liver cancer whose tumors progressed within weeks 0-8 ( Figure 4 ,A), can distinguish the group with tumor progression during weeks 0-12 ( Figure 4 , B), and can also distinguish the group whose tumor progressed during weeks 0-16 ( Figure 4 ,C); In addition, it can also distinguish the population with tumor progression in four time periods: 0-8 weeks, 8-12 weeks, 12-16 weeks, and >16 weeks ( Figure 4 ,D).

[0120] 3. Generalization

[0121] The generalization ability of the CatBoost model using the prediction method of the present application was multiple-validated in three cohorts: the internal validation cohort (SHPG, n=102), the external validation cohort I (SPPH, n=104), and the external validation cohort II (LCPH, n=67). Figure 5 As shown in Table 8, the AUC values ​​of the CatBoost model for tumor re-progression in the three cohorts at weeks 0-8 were 0.89 and 0.91, respectively.

[0122] (95CI:0.79-0.98), 0.88(95CI:0.78-0.98), 0.94(95CI:0.85-1.03)( Figure 5 , A, B, C); the AUC values ​​of tumor progression at weeks 8-12 were 0.76 (95CI: 0.65-0.88), 0.74 (95CI: 0.55-0.93), and 0.76 (95CI: 0.54-0.97), respectively. Figure 5 , D, E, F); the AUC values ​​of tumor progression at weeks 12-16 were 0.74 (95CI: 0.56-0.91), 0.72 (95CI: 0.56-0.88), and 0.73 (95CI: 0.51-0.95), respectively. Figure 5 , G, H, I); the AUC values ​​of tumor progression at >16 weeks were 0.89 (95CI: 0.82-0.96), 0.82 (95CI: 0.74-0.90), and 0.90 (95CI: 0.83-0.97), respectively ( Figure 5 , J, K, L).

[0123] Therefore, the above results confirm that the CatBoost model using the prediction method of the present application has strong generalization ability in both internal and external validation set cohorts.

[0124] Table 8 AUC values ​​of tumor re-progression in different time stages of the CatBoost model in different cohorts

[0125]

[0126] 4. Prediction and application

[0127] Figure 6The weight order of the influence of 15 core features of the CatBoost model using the prediction method of the present application on the prediction results is shown. The results show that when predicting tumor progression within weeks 0-8, the top 9 core features are ranked as follows: tumor diameter, age, PLT (platelets), ALP (alkaline phosphatase), AFP (alpha-fetoprotein), HGB (hemoglobin), ALT (alanine aminotransferase), INR (prothrombin normalized ratio), GGT (glutamyl transpeptidase) ( Figure 6 ,A);

[0128] In predicting tumor re-progression within the 8th to 12th week, the top 9 core features were ranked as follows: tumor diameter, AFP (alpha-fetoprotein), TBIL (total bilirubin), ALT (alanine aminotransferase) and ALP (alkaline phosphatase), DBIL (direct bilirubin), age, INR (prothrombin normalized ratio), AST (aspartate aminotransferase) ( Figure 6 ,B);

[0129] In predicting tumor progression within the 12th to 16th week, the top 9 core features were ranked as follows: tumor diameter, age, ALT (alanine aminotransferase), GGT (glutamyl transpeptidase), AFP (alpha-fetoprotein), TBIL (total bilirubin), DBIL (direct bilirubin), PLT (platelets) ( Figure 6 ,C);

[0130] In predicting tumor progression within >16 weeks, the top 9 core features are ranked as follows: tumor diameter, ALP (alkaline phosphatase), GGT (glutamyl transpeptidase), age, PLT (platelets), TBIL (total bilirubin), DBIL (direct bilirubin), AST (aspartate aminotransferase), WBC (white blood cell count) ( Figure 6 ,D). In addition, the LIME diagram is used to explain the decision logic mechanism of the CatBoost prediction model for predicting tumor progression after patients receive TACE series treatment, among which 0-8 weeks ( Figure 6 ,E) and >16 weeks ( Figure 6 ,F) as representative.

[0131] like Figure 7 As shown, the web version of the visual application interface of the CatBoost model using the prediction method of the present application, the user inputs 15 core feature information of a specific patient, clicks "CatBoost model calculation", and calculates the probability of tumor re-progression in four time stages: 0-8 weeks, 8-12 weeks, 12-16 weeks, and >16 weeks. The time stage with the highest probability is selected, which is the prediction result of the CatBoost model for the time stage of tumor re-progression after the patient undergoes TACE surgery.

Claims

1. A method for predicting the risk of tumor re-progression after hepatic arterial chemoembolization, characterized in that: The following steps are involved: S1: Collect the preoperative baseline characteristic variable data of patients undergoing hepatic arterial chemoembolization, the characteristic variable data at the time of complete / near-complete tumor remission after surgery, the change in characteristic variable data, and the endpoint events of tumor progression and progression-free survival, and construct the 0-8 week data set, 8-12 week data set, 12-16 week data set, and >16 week data set respectively; The preoperative baseline characteristic variable data, including age, alkaline phosphatase, alanine aminotransferase, glutamyl transpeptidase, aspartate aminotransferase, total bilirubin, direct bilirubin, hemoglobin, prothrombin normalized ratio, platelet count, white blood cell count, alpha-fetoprotein, red blood cell count, tumor number and tumor diameter, were recorded as core characteristic variables; S2: The 0-8 week data set, 8-12 week data set, 12-16 week data set, and >16 week data set are used as training sets. After data enhancement processing, the classification enhancement neural network is trained respectively. During the training process, each core feature variable has different weights in the training process of different time periods. The specific weights are shown in the table below. ， Obtaining a first cycle prediction network, a second cycle prediction network, a third cycle prediction network, and a fourth cycle prediction network; S3: Input the data to be tested into the first cycle prediction network, the second cycle prediction network, the third cycle prediction network, and the fourth cycle prediction network respectively, decompose the results of each prediction cycle according to the SHAP algorithm, and obtain the contribution value of each core characteristic variable in the prediction cycle to tumor re-progression, and obtain the probability of tumor re-progression in the first cycle, the probability of tumor re-progression in the second cycle, the probability of tumor re-progression in the third cycle, and the probability of tumor re-progression in the fourth cycle according to the contribution value, respectively, and select the maximum value of the four probability values ​​as the prediction result, and the corresponding time period as the prediction period of tumor re-progression.

2. The method for predicting the risk of tumor re-progression after hepatic arterial chemoembolization according to claim 1, characterized in that: After constructing the 0-8 week data set, 8-12 week data set, 12-16 week data set, and >16 week data set in S1, the data in the data set are cleaned, specifically by deleting collinear features, features of patients with missing baseline and follow-up data, and filling missing data with the median.

3. The method for predicting the risk of tumor re-progression after hepatic arterial chemoembolization according to claim 1, characterized in that: The progression-free survival period starts from the time when complete / near-complete remission is achieved after hepatic arterial chemoembolization and ends when liver cancer progression occurs.

4. A system for predicting the risk of tumor re-progression after hepatic arterial chemoembolization, characterized in that: include: Dataset construction module: used to collect preoperative baseline characteristic variable data of patients undergoing hepatic arterial chemoembolization, characteristic variable data at the time of complete / near-complete remission of the tumor after surgery, the change in characteristic variable data, and endpoint events of tumor progression and progression-free survival, and to construct the 0-8 week data set, 8-12 week data set, 12-16 week data set, and >16 week data set respectively; The preoperative baseline characteristic variable data, including age, alkaline phosphatase, alanine aminotransferase, glutamyl transpeptidase, aspartate aminotransferase, total bilirubin, direct bilirubin, hemoglobin, prothrombin normalized ratio, platelet count, white blood cell count, alpha-fetoprotein, red blood cell count, tumor number and tumor diameter, were recorded as core characteristic variables; Neural network training module: It is used to use the 0-8 week data set, 8-12 week data set, 12-16 week data set, and >16 week data set as training sets. After data enhancement processing, the classification enhancement neural network is trained respectively. During the training process, each core feature variable has different weights in the training process of different time periods. The specific weights are shown in the table below. ， Obtaining a first cycle prediction network, a second cycle prediction network, a third cycle prediction network, and a fourth cycle prediction network; Prediction module: used to input the test data into the first cycle prediction network, the second cycle prediction network, the third cycle prediction network, and the fourth cycle prediction network respectively, decompose the results of each prediction cycle according to the SHAP algorithm, and obtain the contribution value of each core characteristic variable to tumor re-progression in the prediction cycle, and obtain the probability of tumor re-progression in the first cycle, the probability of tumor re-progression in the second cycle, the probability of tumor re-progression in the third cycle, and the probability of tumor re-progression in the fourth cycle according to the contribution value of tumor re-progression respectively, and select the maximum value of the four probability values ​​as the prediction result, and the corresponding time period as the prediction period of tumor re-progression.

5. A device for predicting the risk of tumor progression after hepatic artery chemoembolization, characterized in that: The method comprises a processor and a memory, wherein the processor implements the method for predicting the risk of tumor re-progression after hepatic artery chemoembolization as described in any one of claims 1 to 3 when executing the computer program stored in the memory.

6. A storage medium for predicting the risk of tumor progression after hepatic artery chemoembolization, characterized in that: Used to store a computer program, wherein when the computer program is executed by a processor, the method for predicting the risk of tumor re-progression after hepatic arterial chemoembolization according to any one of claims 1 to 3 is implemented.

Images