A method for predicting osteosarcoma survival based on machine learning model

Abstract

The present invention relates to a method for predicting survival of osteosarcoma based on a machine learning model, and the method includes: obtaining data baseline information based on the SEER database; using a variety of methods to screen out preliminary screening variables as optional factor combinations from the baseline information; constructing a variety of machine learning combination prediction models based on the optional factor combinations, and determining the optimal prediction model according to the average C-index value; performing variable screening and importance evaluation on the optional factor combination according to the optimal prediction model to obtain the optimal variable combination, introducing risk score combined with the clinical indicators of the optimal variable combination to perform survival analysis including KM survival curve, univariate Cox, multivariate Cox and establishment of a nomogram to predict the survival rate for several years. The present invention solves the problem that the existing clinical characteristics are difficult to evaluate the true situation of osteosarcoma patients, and can provide osteosarcoma patients with a suitable high-precision prediction model that can incorporate a variety of prognosis-related change factors.

Description

Technical Field

[0001] The present invention relates to the technical field of prognosis prediction, and in particular to a method for predicting osteosarcoma survival based on a machine learning model. Background Art

[0002] Osteosarcoma is the most common and highly malignant primary malignant bone tumor, accounting for about 35% of malignant tumors (Jawad et al., 2011). It mainly originates from primitive transformed cells originating from mesenchymal sources (Mirabello et al., 2009), and is prone to occur in the metaphysis of long bones, preferably around the knee joint. It is characterized by the production of highly malignant spindle-shaped matrix cells. Surgical resection plus preoperative and postoperative chemotherapy is currently the standard treatment for this disease (Messerschmitt, et al., 2011). The presence or absence of metastasis has become an important factor affecting the prognosis of osteosarcoma patients.

[0003] Currently, the most commonly used osteosarcoma staging systems include the American Joint Committee on Cancer (AJCC) staging (Chou et al., 2008) and the Enneking staging (Greene et al., 2017). Both staging systems assess the clinical risk of osteosarcoma based on initial clinical features such as tumor size, pathological grade, and metastasis. However, they can only provide a rough estimate of a patient's disease risk. Due to differences in the pathogenesis, growth rate, invasiveness, and response to treatment of osteosarcoma, each patient's condition is unique (Marina et al., 2004; Messerschmitt et al., 2009; Meyers et al., 2005; Bacci et al., 2003; Edwards et al., 2008). Therefore, relying solely on these early clinical features for evaluation often does not fully reflect the true situation of osteosarcoma patients. At the same time, any current prediction model or evaluation tool cannot measure multiple types of data, and it is difficult to screen suitable models for analysis and prediction for specific data objects. Moreover, existing prediction models or evaluation tools cannot take into account possible changes in the disease during the prognosis process, such as the emergence of new metastases, treatment responses, etc. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides an osteosarcoma survival prediction method based on a machine learning model, which solves the technical problems that the existing clinical characteristics are difficult to evaluate the actual situation of osteosarcoma patients, and it is difficult to provide a suitable high-precision prediction model for osteosarcoma patients that can incorporate multiple prognosis-related changing factors.

[0006] (II) Technical solution

[0007] In order to achieve the above object, the main technical solutions adopted by the present invention include:

[0008] In a first aspect, an embodiment of the present invention provides a method for predicting osteosarcoma survival based on a machine learning model, comprising:

[0009] Data screening, data filling and continuous variable discretization were performed on the data of osteosarcoma patients obtained from the SEER database to obtain baseline data containing multiple variables;

[0010] At least one set of preliminary screening variables was screened from the baseline data by using at least one of univariate Cox regression analysis, full subset regression analysis, LASSO regression and cross-validation analysis;

[0011] At least one set of initial screening variables was input into the multi-factor Cox model for re-screening, and one set of initial screening variables was determined as an optional factor combination with the minimum AIC value;

[0012] Construct multiple machine learning combined prediction models based on optional factor combinations, and determine the optimal prediction model based on the calculated average C-index value of each prediction model;

[0013] According to the optimal prediction model, the optional factor combination is screened and evaluated for importance to obtain the optimal variable combination. A risk assessment item risk score is introduced as a new signature to perform survival analysis on the variable combination in the optimal variable combination, including Kaplan-Meier survival curve analysis, univariate Cox analysis, multivariate Cox analysis and establishment of a nomogram for N-year survival rate prediction analysis.

[0014] Among them, the variables in the optimal variable combination are all independent variables related to prognosis, including surgical variables as protective factors and age, tumor grade, tumor size, M stage and lung metastasis variables as risk variables.

[0015] Optionally, the data of osteosarcoma patients obtained from the SEER database are screened, filled, and continuous variables are discretized to obtain baseline data containing multiple variables including:

[0016] According to the set inclusion criteria, the data of osteosarcoma patients in the SEER database were retrieved, and the data of patients who were lost to follow-up or missing were deleted;

[0017] The continuous variables representing the size of osteosarcoma in the data of osteosarcoma patients were divided into several groups of categorical variables according to the cut-off value;

[0018] According to the data of osteosarcoma patients after deletion and several groups of categorical variables divided by continuous variables representing the size of osteosarcoma, the baseline data containing multiple variables were obtained.

[0019] Optionally, at least one set of preliminary screening variables is screened out from the baseline data by at least one of univariate Cox regression analysis, full subset regression analysis, and LASSO regression and cross-validation analysis, including:

[0020] In the univariate Cox regression analysis, 17 variables were screened with p < 0.05 as the criterion: age, marital status, tumor size, primary site, laterality, tumor stage and grade, tumor grouping, T stage, N stage, M stage, surgery, chemotherapy, radiotherapy, bone metastasis, brain metastasis, liver metastasis, and lung metastasis;

[0021] In the full subset regression analysis, 7 variables were screened using the optimal subset regression model evaluation criteria: age, primary site, tumor grade, tumor size, M stage, surgery, and lung metastasis; the optimal subset regression model evaluation criteria included adjusting R 2 maximum, Malleus CP minimum, and Bayesian Information Criterion minimum;

[0022] In LASSO regression and cross validation, seven variables were selected based on the tuning coefficient λ value corresponding to the minimum mean square error: marital status, primary site, tumor grade, tumor stage, surgery, radiotherapy, and lung metastasis.

[0023] Optionally, at least one set of initial screening variables is input into a multi-factor Cox model for re-screening, and one set of initial screening variables is determined as an optional factor combination with a minimum AIC value, including:

[0024] Input at least one set of preliminary screening variables obtained from univariate Cox regression analysis, full subset regression analysis, LASSO regression and cross-validation into a multivariate Cox model;

[0025] After multivariate Cox regression analysis of each group of initial screening variables was performed using a multivariate Cox model, one group of initial screening variables was determined as an optional factor combination based on the maximum AUC value and the minimum AIC value;

[0026] Among them, the set of variables obtained by the full subset regression analysis for the optional factor combination is: age, primary site, tumor grade, tumor size, M stage, surgery and lung metastasis.

[0027] Optionally, constructing multiple machine learning combined prediction models based on the optional factor combination, and determining the optimal prediction model according to the calculated average C-index value of each prediction model includes:

[0028] Build more than 100 machine learning combination prediction models based on optional factor combinations and more than 100 machine learning models;

[0029] Randomly divide the baseline data of multiple variables into several sub-samples containing the same number of patient instances;

[0030] For each prediction model, select any sub-sample as the validation set, select other sub-samples outside the validation set as the training set, repeat the validation set and training set selection process multiple times, and obtain an intermediate training model containing the same number of training sets and validation sets as the number of sub-samples;

[0031] Each prediction model was trained and tested based on the intermediate training model, and the LASSO+random survival forest model was determined to be the optimal prediction model by comparing the average C-index value of each prediction model among more than 100 models.

[0032] Optionally, the optional factor combination is screened and evaluated for importance according to the optimal prediction model to obtain the optimal variable combination, and a risk assessment item risk score as a new signature is introduced to perform survival analysis on the variable combination in the optimal variable combination, including Kaplan-Meier survival curve analysis, univariate Cox analysis, multivariate Cox analysis and establishment of a nomogram for N-year survival rate prediction analysis, including:

[0033] The LASSO part of the LASSO+random survival forest model was used to perform a ten-fold cross-validation on the optional factor combination to find the optimal parameter lambda.min under the cross-validation. After obtaining the optimal parameter, the lasso regression model was trained using the training set to obtain the variables and coefficients of the lasso regression model. The variables with a coefficient of 0 were eliminated to obtain the optimal variable combination containing 6 variables that were independent factors related to prognosis;

[0034] The random survival forest part of the LASSO+random survival forest model is used to permutate and combine the six variables in the optimal variable combination selected from the LASSO part to obtain 2 6-1 = 63 prediction models, 63 models were evaluated, and age and lung metastasis were determined as the priority analysis objects according to the AUC value on the drawn ROC curve and the importance ranking of each variable in the optimal variable combination;

[0035] A new risk assessment item risk score of Signature was introduced, and a risk item model was constructed based on the risk assessment item risk score for age and lung metastasis variables as priority analysis objects;

[0036] Performing univariate Cox analysis and / or multivariate Cox analysis on the risk assessment item risk score and the primary site, tumor grade, tumor size and surgery in the optimal variable combination except for age and lung metastasis variables to obtain univariate Cox analysis results or multivariate Cox analysis results;

[0037] The Kaplan-Meier survival curves of risk score, primary site, tumor grade, tumor size and surgical procedure were drawn and the log-rank test was performed.

[0038] Quantify and visualize the results of univariate Cox analysis or multivariate Cox analysis to obtain a nomogram to achieve N-year survival rate prediction analysis;

[0039] Output survival analysis report based on Kaplan-Meier survival curve, univariate Cox analysis, multivariate Cox analysis and establishment of nomogram N-year survival rate prediction analysis results;

[0040] in,

[0041] The risk term model is:

[0042] Risk score (Signature) = 0.0245156974684657·Age+1.19326626353926·lungmetastas, Age is age, lung metastas is lung metastasis.

[0043] Optionally, according to the optimal prediction model, the optional factor combination is subjected to variable screening and importance evaluation to obtain the optimal variable combination, and a risk assessment item risk score as a new Signature is introduced to perform survival analysis including Kaplan-Meier survival curve analysis, univariate Cox analysis, multivariate Cox analysis and establishment of a nomogram N-year survival rate prediction analysis on the variable combination in the optimal variable combination, and further comprising:

[0044] Based on the information of the nomogram and the patient information, a 3D visualization disease progression model of the patient's osteosarcoma-related part is constructed with the help of corresponding 3D simulation components for doctors and patients to view;

[0045] Among them, patient information includes personal age, gender, body shape, medical history, and medical examination images of the affected area.

[0046] In a second aspect, an embodiment of the present invention provides an osteosarcoma survival prediction system based on a machine learning model, comprising:

[0047] The baseline data providing module is used to screen, fill in and discretize the data of osteosarcoma patients in the SEER database to obtain the baseline data containing multiple variables;

[0048] A preliminary screening variable output module is used to screen out at least one set of preliminary screening variables from the data baseline information by using at least one of univariate Cox regression analysis, full subset regression analysis, and LASSO regression and cross-validation analysis;

[0049] An optional factor output module is used to input at least one set of preliminary screening variables into a multi-factor Cox model for re-screening, and to determine one set of preliminary screening variables as an optional factor combination with a minimum AIC value;

[0050] The optimal model determination module is used to construct a variety of machine learning combined prediction models based on the optional factor combination, and determine the optimal prediction model according to the calculated average C-index value of each prediction model;

[0051] The survival prediction evaluation module is used to screen and evaluate the importance of optional factor combinations according to the optimal prediction model to obtain the optimal variable combination, introduce a risk assessment item risk score as a new signature, and perform survival analysis on the variable combination in the optimal variable combination, including KM survival curve analysis, univariate Cox analysis, multivariate Cox analysis, and establishment of a nomogram for N-year survival rate prediction analysis;

[0052] Among them, the variables in the optimal variable combination are all independent variables related to prognosis, including surgical variables as protective factors and age, tumor grade, tumor size, M stage and lung metastasis variables as risk variables.

[0053] In a third aspect, an embodiment of the present invention provides an osteosarcoma survival prediction device based on a machine learning model, comprising:

[0054] server;

[0055] communicating with the server to connect to a database; and,

[0056] a memory in communication with the database;

[0057] Among them, the server is deployed with more than one hundred machine learning models, and the memory stores instructions that can be executed by the at least one database, and the instructions are executed by the at least one database so that the at least one database can execute the osteosarcoma survival prediction method based on the machine learning model as described above.

[0058] In a fourth aspect, an embodiment of the present invention provides a computer-readable medium having computer executable instructions stored thereon, which, when executed by a processor, implements the osteosarcoma survival prediction method based on the machine learning model as described above.

[0059] (III) Beneficial effects

[0060] The beneficial effects of the present invention are as follows: starting from the data of osteosarcoma patients in the SEER database, the present invention obtains the optimal combination of factors that affect the overall survival of osteosarcoma patients through multiple screenings, thereby solving the problem that the existing clinical characteristics are difficult to evaluate the actual situation of osteosarcoma patients, and then determining the optimal prediction model based on the optimal factor combination, providing osteosarcoma patients with a suitable high-precision model with the strongest predictive ability that can incorporate multiple prognosis-related changing factors. Therefore, the present invention lays a solid foundation for formulating more refined treatment strategies and improving the survival rate and quality of life of patients, and has clinical practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A schematic diagram of a specific process of a method for predicting osteosarcoma survival based on a machine learning model provided by an embodiment of the present invention;

[0062] Figure 2 A specific flow chart of step S1 of a method for predicting osteosarcoma survival based on a machine learning model provided in an embodiment of the present invention;

[0063] Figure 3 A specific flow chart of step S2 of a method for predicting osteosarcoma survival based on a machine learning model provided in an embodiment of the present invention;

[0064] Figure 4 A forest plot obtained by univariate Cox regression analysis of a method for predicting osteosarcoma survival based on a machine learning model provided in an embodiment of the present invention;

[0065] Figure 5 In the full subset regression of a method for predicting osteosarcoma survival based on a machine learning model provided in an embodiment of the present invention, R is adjusted 2 Schematic diagram of combinations of variables for standard determination;

[0066] Figure 6A and B are respectively a LASSO coefficient curve of a method for predicting osteosarcoma survival based on a machine learning model provided by an embodiment of the present invention and a cross-validation schematic diagram for adjusting parameter selection in the LASSO model;

[0067] Figure 7 A specific flow chart of step S3 of a method for predicting osteosarcoma survival based on a machine learning model provided in an embodiment of the present invention;

[0068] Figure 8 A specific flow chart of step S4 of a method for predicting osteosarcoma survival based on a machine learning model provided in an embodiment of the present invention;

[0069] Fig. 9 C-index values ​​of 117 machine learning models of a method for predicting osteosarcoma survival based on a machine learning model provided in an embodiment of the present invention;

[0070] Fig.10 A specific flow chart of step S5 of a method for predicting osteosarcoma survival based on a machine learning model provided in an embodiment of the present invention;

[0071] Fig.11 A univariate Cox forest diagram of a method for predicting osteosarcoma survival based on a machine learning model provided in an embodiment of the present invention;

[0072] Fig.12 A schematic diagram of a multi-factor Cox forest for a training set of a method for predicting osteosarcoma survival based on a machine learning model provided in an embodiment of the present invention;

[0073] Fig.13 A schematic diagram of a multi-factor Cox forest for a validation set of a method for predicting osteosarcoma survival based on a machine learning model provided in an embodiment of the present invention;

[0074] Fig.14 A training set Kaplan-Meier survival curve of an osteosarcoma survival prediction method based on a machine learning model is provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0075] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation modes in conjunction with the accompanying drawings.

[0076] like Figure 1As shown, an osteosarcoma survival prediction method based on a machine learning model proposed in an embodiment of the present invention includes: performing data screening, data filling and continuous variable discretization on the data of osteosarcoma patients obtained from the SEER database to obtain data baseline data containing multiple variables; screening at least one group of preliminary screening variables from the data baseline data by at least one of univariate Cox regression analysis, full subset regression analysis, LASSO regression and cross-validation analysis; inputting at least one group of preliminary screening variables into a multifactor Cox model for re-screening, and determining one group of preliminary screening variables as an optional factor combination with a minimum AIC value; constructing a plurality of machine learning combination prediction models based on the optional factor combination, and determining the optimal prediction model according to the calculated average C-index value of each prediction model; performing variable screening and importance evaluation on the optional factor combination according to the optimal prediction model to obtain the optimal variable combination, and introducing a risk assessment item risk as a new signature score performed survival analysis on the variable combination in the optimal variable combination, including Kaplan-Meier survival curve analysis, univariate Cox analysis, multivariate Cox analysis and establishment of a nomogram for N-year survival rate prediction analysis; among them, the variables in the optimal variable combination are all independent variables related to prognosis, including surgical variables as protective factors and age, tumor grade, tumor size, M stage and lung metastasis variables as risk variables.

[0077] The present invention starts from the data of osteosarcoma patients in the SEER database, and obtains the optimal combination of factors that affect the overall survival of osteosarcoma patients through multiple screenings, which solves the problem that the existing clinical characteristics are difficult to evaluate the actual situation of osteosarcoma patients, and then determines the optimal prediction model based on the optimal factor combination, providing osteosarcoma patients with a suitable high-precision model with the strongest predictive ability that can incorporate multiple prognosis-related changing factors. Therefore, the present invention lays a solid foundation for formulating more refined treatment strategies and improving the survival rate and quality of life of patients, and has clinical practical value.

[0078] In order to better understand the above technical solution, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a clearer and more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0079] Specifically, the present invention provides a method for predicting osteosarcoma survival based on a machine learning model, characterized in that it includes:

[0080] S1. Data screening, data filling and continuous variable discretization are performed on the data of osteosarcoma patients obtained from the SEER database to obtain baseline data containing multiple variables.

[0081] Furthermore, if Figure 2 As shown, step S1 includes:

[0082] S11. According to the set inclusion criteria, the data of osteosarcoma patients in the SEER database were retrieved, and the data of patients who were lost or missing were deleted to make the subsequent model more accurate.

[0083] The SEER database is an authoritative cancer statistics database in the United States, which can provide a basis for clinical research. At present, although there have been relevant studies on osteosarcoma based on the SEER database, they all have low ability to predict survival rates (AUC is basically lower than 0.8). This directly affects the judgment of the patient's prognosis; these models have no defects in external data verification. External validation is a very important step, which can help understand the predictive performance of the model in different data sets or populations, thereby testing the robustness and generalization ability of the model. The lack of external validation may limit the practical value of these prediction models and challenge their application in actual clinical work. At the same time, the present invention collects osteosarcoma patient data from medical centers in 4 different regions of China as a validation set to externally validate the prediction model and increase the credibility of the model.

[0084] In a specific embodiment, the present invention is a retrospective cohort study, and a total of 895 osteosarcoma patients were included in the SEER database from 2010 to 2018. According to the ratio of 7:3, the training set has 627 people and the validation set has 268 samples.

[0085] Therefore, the inclusion criteria of SEER data are as follows: (1) Osteosarcoma is a primary malignant tumor (Primary Site-labeled), and the extracted record content is 9180, 9181, 9182, 9183, 9184, 9185, 9186, 9187, 9192, 9193, 9194, or 9200 according to the International Classification of Diseases for Oncology ICD-O coding standard. (2) The SEER database after 2010 includes relevant information on metastatic sites, so patients diagnosed between 2010 and 2018 are included. (3) Make sure that osteosarcoma is the first and only primary malignant tumor (First malignant primary indicator), and the extracted record content is YES. (4) The tumor is confirmed to be accompanied by lung metastasis, bone metastasis, brain metastasis, and liver metastasis by imaging or pathological examination. (5) Complete clinical information, including age at diagnosis, gender, marital status, race, primary site, tumor size, tumor stage and grade, metastatic site, surgery, and whether radiotherapy and chemotherapy were performed; (6) Complete follow-up information; (7) Known cause of death and survival time after diagnosis.

[0086] Exclusion criteria for SEER data: (1) incomplete clinicopathological and survival information; (2) unknown tumor size, stage, and race; (3) data lost to follow-up; and (4) missing data.

[0087] Therefore, according to the data obtained in the above steps, the dependent variables: survival status, survival time. The missing data in the data are supplemented, and the data with a survival time of 0 means that the patient died within 1 month, which is also of research significance in medicine, so it is retained.

[0088] S12. Divide the continuous variable representing the size of osteosarcoma in the data of osteosarcoma patients into several groups of categorical variables according to the cut-off value.

[0089] Tumor size is usually expressed as volume or diameter, and tumor size is best divided into the following three groups: less than or equal to 38 mm, 38-160 mm, and greater than 160 mm. Therefore, the continuous variable of tumor size was converted into a categorical variable of "≤38, 38-160, >160". Other patients with tumor size coded as "999" were assigned to "Unable to evaluate". Therefore, the continuous variable tumor size was converted into a categorical variable of four groups of "≤38, 38-160, >160, Unable to evaluate".

[0090] S13. A data baseline containing multiple variables is obtained based on the data of osteosarcoma patients that have been deleted and the continuous variables representing the size of osteosarcoma, which are divided into several groups of categorical variables.

[0091] The final data baseline containing multiple variables is:

[0092] Table 1 Baseline data of osteosarcoma patients in the training group and validation group from 2010 to 2019

[0093]

[0094]

[0095]

[0096] Table 1 shows the demographic, clinical pathological and treatment characteristics of osteosarcoma patients in the SEER database. There was no significant statistical difference in all clinical indicators between the training set and the validation set, and the training set and the validation set were similar in clinical indicators, which indicated that the grouping was reasonable.

[0097] S2. Screen out at least one set of initial screening variables from the baseline data by at least one of univariate Cox regression analysis, full subset regression analysis, LASSO regression and cross-validation analysis.

[0098] Furthermore, if Figure 3 As shown, step S2 includes:

[0099] S22, Reference Figure 4 In the univariate Cox regression analysis, 17 variables were screened with p < 0.05 as the criterion: age, marital status, tumor size, primary site, lateralization, tumor stage and grade, tumor grouping, T stage, N stage, M stage, surgery, chemotherapy, radiotherapy, bone metastasis, brain metastasis, liver metastasis and lung metastasis.

[0100] S23, Reference Figure 5 In the full subset regression analysis, 7 variables were screened using the optimal subset regression model evaluation criteria: age, primary site, tumor grade, tumor size, M stage, surgery, and lung metastasis; among them, the optimal subset regression model evaluation criteria included adjusting R 2 Maximum, Malleus CP minimum, and Bayesian Information Criterion minimum.

[0101] S24. In LASSO regression and cross validation, seven variables were selected based on the tuning coefficient λ value corresponding to the minimum mean square error: marital status, primary site, tumor grade, tumor stage, surgery, radiotherapy and lung metastasis.

[0102] To find the best model, LASSO introduces the variable λ (lambda, also called shrinkage operator, model coefficient ratio, tuning coefficient or penalty value). Figure 6As shown in the figure: As λ increases, the regression coefficient β of each variable decreases, and some even become 0, indicating that the variable contributes little to the model at this time and can be eliminated. Figure 6 In A, a line represents the change in the regression coefficient β value of a variable, the number below the x-axis is the penalty value (tuning coefficient), and the number above the x-axis is the number of variables remaining under this value. LASSO regression compresses the regression coefficients of the variables in the regression model by generating a penalty function to prevent overfitting and solve the problem of serious collinearity. Therefore, the λ value determines which variables can make the model optimal, and cross-validation can be used to find the optimal λ value. When the mean square error (MSE) is minimized, the corresponding λ value determines the variables included in the model. The smaller the MSE value, the better the accuracy of the prediction model.

[0103] Figure 6 B shows the partial likelihood deviation curve as Log(λ) changes. Figure 6 B gives two penalty values ​​(tuning coefficients) λ: one is the λ value when the MSE (mean square error) is the smallest, that is, lambda.min; the other is the λ value of the simplest model obtained within a variance range of the lambda.min value. This value gives a model with good performance and the least number of independent variables, so this value is generally selected. Finally, LASSO regression screened out 6 variables: marital status, primary site, tumor grade, tumor stage, surgery, radiotherapy, and lung metastasis.

[0104] S3. Input at least one set of initial screening variables into the multi-factor Cox model for re-screening, and determine one set of initial screening variables as an optional factor combination with the minimum AIC value.

[0105] Furthermore, if Figure 7 As shown, step S3 includes:

[0106] S31. Input at least one set of preliminary screening variables obtained from univariate Cox regression analysis, full subset regression analysis, LASSO regression and cross-validation into a multivariate Cox model.

[0107] S32. After performing multivariate Cox regression analysis on each group of initial screening variables using a multivariate Cox model, one group of initial screening variables was determined as an optional factor combination based on the maximum AUC value and the minimum AIC value. Among them, the optional factor combination was a group of variables obtained by full subset regression analysis: age, tumor grade, tumor size, M stage, surgery, and lung metastasis.

[0108] The initial screening variables screened by the above three methods were respectively included in the multivariate Cox model, and the final models of the three methods were determined by the minimum AIC value. Finally, the ROC curves of the three models in 1-year, 3-year and 5-year overall survival were drawn to evaluate the performance of the three models, and the best model was evaluated by the AUC value. The AIC of the three models were: 2733.897 for univariate Cox, 2688.516 for optimal subset regression, and 2720.877 for Lasso. After comparison, the model constructed using the variables screened by BSR in both the training set and the validation set was better, with the largest AUC value and the smallest AIC value. Therefore, the optimal subset was finally selected to return to 7 variables (age, primary site, tumor grade, tumor size, M stage, surgery and lung metastasis) to construct a prediction model.

[0109] S4. Construct multiple machine learning combined prediction models based on optional factor combinations, and determine the optimal prediction model based on the calculated average C-index value of each prediction model.

[0110] Furthermore, if Figure 8 As shown, step S4 includes:

[0111] S41. Construct more than 100 machine learning combination prediction models based on optional factor combinations and more than 100 machine learning models, specifically 117 machine learning models.

[0112] S42. Randomly divide the baseline data of multiple variables into a number of sub-samples containing the same number of patient instances.

[0113] S43. For each prediction model, select any sub-sample as the validation set, select other sub-samples outside the validation set as the training set, repeat the validation set and training set selection process multiple times, and obtain an intermediate training model containing the same number of training sets and validation sets as the number of sub-samples.

[0114] S44. Train and test each prediction model based on the intermediate training model, and determine that the LASSO+random survival forest model is the optimal prediction model by comparing the average C-index value of each prediction model among more than one hundred models.

[0115] To compare the performance of 117 different machine learning models and their combinations in predicting prognosis, the average C-index value of each model was calculated. Fig. 9It was found that the C-index value of LASSO+Random Survival Forest (RSF) was the highest in both the training set (C-index, 0.877), the validation set (C-index, 0.795) and the average of the training and validation sets (C-index, 0.836). In the future, variable screening and importance scoring will be performed based on the optimal prediction model, and risk score will be formulated as an important new signature for the prognosis evaluation of osteosarcoma.

[0116] S5. According to the optimal prediction model, the optional factor combinations are screened and evaluated for importance to obtain the optimal variable combination. A risk assessment item risk score is introduced as a new signature to perform survival analysis on the variable combination in the optimal variable combination, including Kaplan-Meier survival curve analysis, univariate Cox analysis, multivariate Cox analysis and establishment of a nomogram for N-year survival rate prediction analysis.

[0117] Furthermore, if Fig.10 As shown, step S5 includes:

[0118] S51. Use the LASSO part of the LASSO+random survival forest model to perform ten-fold cross-validation on the optional factor combination to find the optimal parameter lambda.min under cross-validation. After obtaining the optimal parameter, use the training set to train the lasso regression model to obtain the variables and coefficients of the lasso regression model, eliminate the variables with a coefficient of 0, and obtain the optimal variable combination containing 6 variables that are independent factors related to prognosis.

[0119] S52. Use the random survival forest part of the LASSO + random survival forest model to permute and combine the 6 variables in the optimal variable combination selected from the LASSO part to obtain 2 6 -1=63 prediction models, 63 models were evaluated, and age and lung metastasis were determined as priority analysis objects based on the AUC value on the drawn ROC curve and the importance ranking of each variable in the optimal variable combination.

[0120] S53. Introduce a new risk assessment item risk score of Signature, and build a risk item model for age and lung metastasis variables as priority analysis objects based on the risk assessment item riskscore.

[0121] Then, a random survival forest was performed on the six variables, and the optimal number of variables was selected based on the AUC. When the number of variables was 2, the AUC was close to being relatively large. At this time, the corresponding variables were age and lung metastasis. The risk term model was established as follows:

[0122] Risk score(Signature)=0.0245156974684657·Age+1.19326626353926·lungmestas.

[0123] The index of this risk item model can be said to predict the risk value of patient survival. The higher the value, the greater the risk value. Although it can be predicted independently, clinical practice often hopes to obtain 1, 3, and 5-year survival rates, and combined with other indicators to predict with higher accuracy. Therefore, further establishing a nomogram can predict 1, 3, and 5-year survival rates.

[0124] S54. Perform univariate Cox analysis and / or multivariate Cox analysis on the risk assessment item risk score and the primary site, tumor grade, tumor size and surgery in the optimal variable combination except age and lung metastasis variables to obtain univariate Cox analysis results or multivariate Cox analysis results.

[0125] S55. Draw the Kaplan-Meier survival curve for the risk assessment items risk score, primary site, tumor grade, tumor size and surgery, and perform a log-rank test.

[0126] S56. Quantify and visualize the results of univariate Cox analysis or multivariate Cox analysis to obtain a nomogram to achieve N-year survival rate prediction analysis.

[0127] S57. Output a survival analysis report based on the Kaplan-Meier survival curve, univariate Cox analysis, multivariate Cox analysis and the establishment of a nomogram to predict the N-year survival rate.

[0128] Univariate Cox analysis of Signatur revealed that its HR was greater than 1 (HR = 3.06 (2.59-3.61)). The univariate Cox analysis of Signatur and primary site, tumor grade, tumor size, and surgery is shown in Table 2 and Fig.11 At the same time, these indicators were subjected to multi-factor Cox analysis of the training set, and a multi-factor Cox forest plot of the training set was drawn ( Fig.12 ). Kaplan-Meier survival curves were drawn for the variables and log-rank tests were performed ( Fig.14, where A is divided into high and low sig groups. B tumor size: ≤38. C tumor size: 38-160. D tumor size: >160. E tumor size: Unable to evaluate. F primary site upper limb. G primary site: lower limb. H primary site: spine, mandible, etc. I tumor grade: well differentiated. J tumor grade: moderately differentiated. K tumor grade: poorly differentiated. L tumor grade: undifferentiated. M tumor grade: unknown. N surgery: yes. O surgery: no). At the same time, the SEER data were subjected to log-rank test to identify whether there was a difference in survival rate between the two groups of patients. Multifactorial Cox forest plot ( Fig.12 ) The results showed that all five parameters were independent risk factors affecting the overall survival of osteosarcoma patients, and the protective factors (surgery) and risk factors (signature, primary site, tumor grade and tumor size) could be clearly distinguished. In particular, signature was an independent risk factor for the overall survival of osteosarcoma patients. The log-rank test results showed that the P values ​​of the five categorical variables were all less than 0.05, with statistically significant differences, and the validation set results were also consistent (Table 4 and Fig.13 ). P>0.05, no statistical difference between the two groups. Therefore, it is believed that there is no significant difference in the survival rate of patients in the training set and the validation set.

[0129] In the KM curve training set, except for tumor size ≤ 38 and Grade I, the rest are meaningful. In the validation set, tumor size ≤ 38, Grade I, primary site 1, and Grade I are meaningless, and the rest are meaningful, which is consistent overall.

[0130] Table 2 Univariate Cox

[0131]

[0132]

[0133] Table 3 Multi-factor Cox regression analysis of training set

[0134]

[0135] Table 4 Multi-factor Cox regression analysis of validation set

[0136]

[0137] Furthermore, after step S5, it also includes: based on the information of the nomogram and the patient information, using the corresponding three-dimensional simulation components to construct a three-dimensional visual disease progression model of the patient's osteosarcoma-related part for medical and patient viewing; wherein the patient information includes personal age, gender, body shape, previous medical history and medical examination images of the affected area.

[0138] Nomogram is a method that can quantify and visualize Cox regression results and is a common form of predictive model expression. Nomogram standardizes the regression coefficients and then displays them as risk scores on the axis. There are currently two common methods for using Nomogram: 1) Each variable is listed separately, and each subvariable is quantified into a specific score (points). Then, the cumulative scores of all variables are matched with the result scale to obtain the predicted probability.

[0139] In addition, an embodiment of the present invention also provides an osteosarcoma survival prediction system based on a machine learning model, comprising:

[0140] The baseline data providing module is used to screen, fill in and discretize the data of osteosarcoma patients obtained from the SEER database to obtain baseline data containing multiple variables.

[0141] The initial screening variable output module is used to screen out at least one set of initial screening variables from the data baseline information by at least one of univariate Cox regression analysis, full subset regression analysis, and LASSO regression and cross-validation analysis.

[0142] The optional factor output module is used to input at least one set of preliminary screening variables into the multi-factor Cox model for re-screening, and determine one set of preliminary screening variables as an optional factor combination with the minimum AIC value.

[0143] The optimal model determination module is used to construct multiple machine learning combined prediction models based on optional factor combinations, and determine the optimal prediction model based on the calculated average C-index value of each prediction model.

[0144] The survival prediction evaluation module is used to screen variables and evaluate the importance of optional factor combinations according to the optimal prediction model to obtain the optimal variable combination. A risk assessment item risk score is introduced as a new signature to perform survival analysis on the variable combination in the optimal variable combination, including KM survival curve analysis, univariate Cox analysis, multivariate Cox analysis and establishment of a nomogram for N-year survival rate prediction analysis.

[0145] In addition, an embodiment of the present invention further provides an osteosarcoma survival prediction device based on a machine learning model, comprising: a server; a database connected to the server in communication; and a memory connected to the database in communication; wherein the server is deployed with more than one hundred machine learning models, and the memory stores instructions that can be executed by at least one database, and the instructions are executed by the at least one database so that the at least one database can execute the osteosarcoma survival prediction method based on the machine learning model as described above.

[0146] Furthermore, an embodiment of the present invention also provides a computer-readable medium having computer executable instructions stored thereon, and when the executable instructions are executed by a processor, the above-mentioned osteosarcoma survival prediction method based on the machine learning model is implemented.

[0147] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0148] It should be noted that the word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention may be implemented by means of hardware comprising several distinct components and by means of a suitably programmed computer. In a claim enumerating several means, several of these means may be embodied by one and the same hardware. The use of the words first, second, third, etc., is for convenience only and does not imply any order. These words may be understood as part of the name of the component.

[0149] Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention should also include these modifications and variations.

Claims

1. A method for predicting osteosarcoma survival based on a machine learning model, characterized in that: include: Data screening, data filling and continuous variable discretization were performed on the data of osteosarcoma patients obtained from the SEER database to obtain baseline data containing multiple variables; At least one set of preliminary screening variables is screened from the baseline data by using at least one of univariate Cox regression analysis, full subset regression analysis, LASSO regression and cross-validation analysis; At least one set of initial screening variables was input into the multi-factor Cox model for re-screening, and one set of initial screening variables was determined as an optional factor combination with the minimum AIC value; Construct multiple machine learning combined prediction models based on optional factor combinations, and determine the optimal prediction model based on the calculated average C-index value of each prediction model; According to the optimal prediction model, the optional factor combination is screened and evaluated for importance to obtain the optimal variable combination. A risk assessment item risk score is introduced as a new signature to perform survival analysis on the variable combination in the optimal variable combination, including Kaplan-Meier survival curve analysis, univariate Cox analysis, multivariate Cox analysis and establishment of a nomogram for N-year survival rate prediction analysis. According to the optimal prediction model, the optional factor combination is screened and evaluated for importance to obtain the optimal variable combination. A risk assessment item risk score is introduced as a new signature to perform survival analysis on the variable combination in the optimal variable combination, including Kaplan-Meier survival curve analysis, univariate Cox analysis, multivariate Cox analysis and establishment of a nomogram for N-year survival rate prediction analysis. The survival analysis includes: The LASSO part of the LASSO+random survival forest model was used to perform a ten-fold cross-validation on the optional factor combination to find the optimal parameter lambda.min under the cross-validation. After obtaining the optimal parameter, the LASSO regression model was trained using the training set to obtain the variables and coefficients of the LASSO regression model. The variables with a coefficient of 0 were eliminated to obtain the optimal variable combination containing 6 variables that were independent factors related to prognosis. The random survival forest part of the LASSO+random survival forest model is used to permutate and combine the six variables in the optimal variable combination selected from the LASSO part to obtain 2 6 -1 = 63 prediction models, 63 models were evaluated, and age and lung metastasis were determined as the priority analysis objects according to the AUC value on the drawn ROC curve and the importance ranking of each variable in the optimal variable combination; A new risk assessment item risk score of Signature was introduced, and a risk item model was constructed based on the risk assessment item risk score for age and lung metastasis variables, which were the priority analysis objects; Performing univariate Cox analysis and / or multivariate Cox analysis on the risk assessment item risk score and the primary site, tumor grade, tumor size and surgery in the optimal variable combination except for age and lung metastasis variables to obtain univariate Cox analysis results or multivariate Cox analysis results; The Kaplan-Meier survival curves of risk score, primary site, tumor grade, tumor size and surgical procedure were drawn and the log-rank test was performed. Quantify and visualize the results of univariate Cox analysis or multivariate Cox analysis to obtain a nomogram to achieve N-year survival rate prediction analysis; Output survival analysis report based on Kaplan-Meier survival curve, univariate Cox analysis, multivariate Cox analysis and establishment of nomogram N-year survival rate prediction analysis results; in, The risk term model is: Risk score (Signature) = 0.0245156974684657·Age+1.19326626353926·lungmetastas, Age is age, lung metastas is lung metastasis; Patient information includes personal age, gender, body type, medical history, and medical examination images of the affected area; The variables in the optimal variable combination were all independent variables related to prognosis, including surgical variables as protective factors and age, tumor grade, tumor size, M stage, and lung metastasis variables as risk variables.

2. The osteosarcoma survival prediction method based on a machine learning model according to claim 1, characterized in that: The data of osteosarcoma patients in the SEER database were screened, filled, and continuous variables were discretized to obtain the baseline data containing multiple variables, including: According to the set inclusion criteria, the data of osteosarcoma patients in the SEER database were retrieved, and the data of patients who were lost to follow-up or missing were deleted; The continuous variables representing the size of osteosarcoma in the data of osteosarcoma patients were divided into several groups of categorical variables according to the cut-off value; According to the data of osteosarcoma patients after deletion and several groups of categorical variables divided by continuous variables representing the size of osteosarcoma, the baseline data containing multiple variables were obtained.

3. The osteosarcoma survival prediction method based on a machine learning model according to claim 1, characterized in that: At least one set of initial screening variables was screened from the baseline data by using at least one of univariate Cox regression analysis, full subset regression analysis, LASSO regression and cross-validation analysis, including: In the univariate Cox regression analysis, 17 variables were screened with p < 0.05 as the criterion: age, marital status, tumor size, primary site, laterality, tumor stage and grade, tumor grouping, T stage, N stage, M stage, surgery, chemotherapy, radiotherapy, bone metastasis, brain metastasis, liver metastasis, and lung metastasis; In the full subset regression analysis, 7 variables were screened using the optimal subset regression model evaluation criteria: age, primary site, tumor grade, tumor size, M stage, surgery, and lung metastasis; the optimal subset regression model evaluation criteria included adjusting R 2 maximum, Malleus CP minimum, and Bayesian Information Criterion minimum; In LASSO regression and cross validation, seven variables were selected based on the tuning coefficient λ value corresponding to the minimum mean square error: marital status, primary site, tumor grade, tumor stage, surgery, radiotherapy, and lung metastasis.

4. The osteosarcoma survival prediction method based on a machine learning model according to claim 1, characterized in that: At least one set of initial screening variables was input into the multi-factor Cox model for re-screening, and one set of initial screening variables was determined as an optional factor combination with the minimum AIC value, including: Input at least one set of preliminary screening variables obtained from univariate Cox regression analysis, full subset regression analysis, LASSO regression and cross-validation into a multivariate Cox model; After multivariate Cox regression analysis of each group of initial screening variables was performed using a multivariate Cox model, one group of initial screening variables was determined as an optional factor combination based on the maximum AUC value and the minimum AIC value; Among them, the set of variables obtained by the full subset regression analysis for the optional factor combination is: age, primary site, tumor grade, tumor size, M stage, surgery and lung metastasis.

5. The osteosarcoma survival prediction method based on a machine learning model according to claim 1, characterized in that: Based on the optional factor combination, a variety of machine learning combined prediction models are constructed, and the optimal prediction model is determined based on the calculated average C-index value of each prediction model, including: Build more than 100 machine learning combination prediction models based on optional factor combinations and more than 100 machine learning models; Randomly divide the baseline data of multiple variables into several sub-samples containing the same number of patient instances; For each prediction model, select any sub-sample as the validation set, select other sub-samples outside the validation set as the training set, repeat the validation set and training set selection process multiple times, and obtain an intermediate training model containing the same number of training sets and validation sets as the number of sub-samples; Each prediction model was trained and tested based on the intermediate training model, and the LASSO+random survival forest model was determined to be the optimal prediction model by comparing the average C-index value of each prediction model among more than 100 models.

6. The osteosarcoma survival prediction method based on a machine learning model according to any one of claims 1 to 5, characterized in that: According to the optimal prediction model, the optional factor combination is screened and evaluated for importance to obtain the optimal variable combination. A risk assessment item risk score is introduced as a new signature to perform survival analysis on the variable combination in the optimal variable combination, including Kaplan-Meier survival curve analysis, univariate Cox analysis, multivariate Cox analysis and establishment of a nomogram for N-year survival rate prediction analysis. The following also includes: Based on the information of the nomogram and the patient information, a 3D visualization disease progression model of the patient's osteosarcoma-related part is constructed with the help of corresponding 3D simulation components for doctors and patients to view; Among them, patient information includes personal age, gender, body shape, medical history, and medical examination images of the affected area.

7. An osteosarcoma survival prediction system based on a machine learning model, characterized in that: include: The baseline data providing module is used to screen, fill in and discretize the data of osteosarcoma patients in the SEER database to obtain the baseline data containing multiple variables; A preliminary screening variable output module is used to screen out at least one set of preliminary screening variables from the data baseline information by using at least one of univariate Cox regression analysis, full subset regression analysis, and LASSO regression and cross-validation analysis; An optional factor output module is used to input at least one set of preliminary screening variables into a multi-factor Cox model for re-screening, and to determine one set of preliminary screening variables as an optional factor combination with a minimum AIC value; The optimal model determination module is used to construct a variety of machine learning combined prediction models based on the optional factor combination, and determine the optimal prediction model according to the calculated average C-index value of each prediction model; The survival prediction evaluation module is used to screen and evaluate the importance of optional factor combinations according to the optimal prediction model to obtain the optimal variable combination, introduce a risk assessment item risk score as a new signature, and perform survival analysis on the variable combination in the optimal variable combination, including KM survival curve analysis, univariate Cox analysis, multivariate Cox analysis, and establishment of a nomogram for N-year survival rate prediction analysis; According to the optimal prediction model, the optional factor combination is screened and evaluated for importance to obtain the optimal variable combination. A risk assessment item risk score is introduced as a new signature to perform survival analysis on the variable combination in the optimal variable combination, including Kaplan-Meier survival curve analysis, univariate Cox analysis, multivariate Cox analysis and establishment of a nomogram for N-year survival rate prediction analysis. The survival analysis includes: The LASSO part of the LASSO+random survival forest model was used to perform a ten-fold cross-validation on the optional factor combination to find the optimal parameter lambda.min under the cross-validation. After obtaining the optimal parameter, the LASSO regression model was trained using the training set to obtain the variables and coefficients of the LASSO regression model. The variables with a coefficient of 0 were eliminated to obtain the optimal variable combination containing 6 variables that were independent factors related to prognosis. The random survival forest part of the LASSO+random survival forest model is used to permutate and combine the six variables in the optimal variable combination selected from the LASSO part to obtain 2 6 -1 = 63 prediction models, 63 models were evaluated, and age and lung metastasis were determined as the priority analysis objects according to the AUC value on the drawn ROC curve and the importance ranking of each variable in the optimal variable combination; A new risk assessment item risk score of Signature was introduced, and a risk item model was constructed based on the risk assessment item risk score for age and lung metastasis variables, which were the priority analysis objects; Performing univariate Cox analysis and / or multivariate Cox analysis on the risk assessment item risk score and the primary site, tumor grade, tumor size and surgery in the optimal variable combination except for age and lung metastasis variables to obtain univariate Cox analysis results or multivariate Cox analysis results; The Kaplan-Meier survival curves of risk score, primary site, tumor grade, tumor size and surgical procedure were drawn and the log-rank test was performed. Quantify and visualize the results of univariate Cox analysis or multivariate Cox analysis to obtain a nomogram to achieve N-year survival rate prediction analysis; Output survival analysis report based on Kaplan-Meier survival curve, univariate Cox analysis, multivariate Cox analysis and establishment of nomogram N-year survival rate prediction analysis results; in, The risk term model is: Risk score (Signature) = 0.0245156974684657·Age+1.19326626353926·lungmetastas, Age is age, lung metastas is lung metastasis; Patient information includes personal age, gender, body type, medical history, and medical examination images of the affected area; The variables in the optimal variable combination were all independent variables related to prognosis, including surgical variables as protective factors and age, tumor grade, tumor size, M stage, and lung metastasis variables as risk variables.

8. An osteosarcoma survival prediction device based on a machine learning model, characterized in that: include: server; communicating with the server to connect to a database; and, a memory in communication with the database; Among them, the server is deployed with more than one hundred machine learning models, and the memory stores instructions that can be executed by the at least one database, and the instructions are executed by the at least one database so that the at least one database can execute the osteosarcoma survival prediction method based on the machine learning model as described in any one of claims 1-6.

9. A computer-readable medium having computer-executable instructions stored thereon, characterized in that: When the executable instructions are executed by the processor, the osteosarcoma survival prediction method based on the machine learning model as described in any one of claims 1 to 6 is implemented.

Images