Machine learning method, device and equipment for multi-cancer classification and storage medium

Abstract

The invention relates to the field of bioinformatics data processing, in particular to a machine learning method, device and equipment for multi-cancer classification and a storage medium, and the machine learning method for multi-cancer classification comprises the following steps: preprocessing a sample; the samples in the training set are divided into a plurality of sub-groups, the samples in each sub-group are classified, the AUC value of each binary classification model is calculated, and sub-optimal models of the sub-groups are obtained through screening; the hyper-parameter of the sub-optimal model of each sub-group is optimized, and the prediction probability of the category to which the cancer species of each sub-group belongs is output; calculating the AUC value of each multi-classification model, and selecting the multi-classification model with the highest AUC value as a total optimal model; the hyper-parameters of the total optimal model are optimized; and calculating a model evaluation quantitative index. The method has the effects of integrating different types of clinical data and carrying out combined diagnosis on a plurality of cancer species.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics data processing, and in particular to a machine learning method, device, equipment and storage medium for multi-cancer classification. Background Art

[0002] Digestive tract cancer is one of the most common malignant tumors that seriously endangers human health. Early diagnosis of cancer is one of the effective means to improve the survival rate of cancer patients. Currently, the screening and diagnosis of digestive tract cancer in clinical practice are mainly through imaging methods such as endoscopy or ultrasound, as well as pathological testing. However, endoscopic examinations for colorectal cancer and gastric cancer are expensive, and patient compliance is poor. Although ultrasonic examinations for liver cancer are simple, convenient and inexpensive, they have a high rate of missed diagnosis due to limited imaging clarity and the inability to perform tissue biopsies immediately.

[0003] In the patent application with publication number CN106399570A, an auxiliary diagnosis system for early colorectal cancer is provided, including a PCR detection kit for the methylation level of BMP3 and NDRG4 genes in stool samples, a PCR detection kit for KRAS gene mutations in stool samples, and a fecal occult blood detection kit. The diagnosis result is obtained by substituting multiple test results into a mathematical model based on logistic regression, and comparing the score obtained by the model with a preset threshold, thereby achieving the diagnosis of colorectal cancer. This method has achieved a sensitivity of 76.4% and a specificity of 88.10% in clinical practice.

[0004] In the patent application with the announcement number CN105200157B, a method for diagnosing gastric cancer is provided, which includes a kit for quantitative detection of 5 microRNAs in plasma. The method quantitatively analyzes the CT value obtained by PCR, substitutes it into a mathematical model also based on logistic regression, and compares the score obtained by the model with the preset threshold, thereby realizing the diagnosis of colorectal cancer. The method has achieved a sensitivity of 84.0% and a specificity of 84.0% in clinical practice.

[0005] In the patent application with publication number CN113611404A, a mathematical model for liver cancer diagnosis is provided. The method performs binary logistic regression analysis on the results of gender, age, and liver cancer triple test (including alpha-fetoprotein allotypes, alpha-fetoprotein, and abnormal prothrombin), and on this basis, a multi-factor logistic regression model is established. After the diagnostic result is brought into the model, a score is obtained, and the score is compared with the preset threshold to achieve the diagnosis of liver cancer. The diagnostic sensitivity of this method for primary hepatocellular carcinoma is 86.9%, and the specificity is 80.2%.

[0006] It can be seen that the current diagnosis of single cancer types almost reaches a sensitivity of more than 80% and a specificity of more than 80% based on the detection of clinical samples by detection kits and the mathematical model analysis of the detection results. However, these existing technologies require the use of different kits to perform multiple tests on the test subject, which not only increases the cost of testing, but also brings inconvenience to the test subject due to the increase in the use of clinical samples. Therefore, it is difficult to conduct digestive tract cancer screening for non-cancer high-risk populations and trace the detected cancers through these methods, which will greatly limit the application of cancer screening tests. Summary of the invention

[0007] In order to integrate different types of clinical data and perform joint diagnosis of multiple cancer types, the present invention provides a machine learning method, device, equipment and storage medium for multi-cancer classification.

[0008] The present invention provides a machine learning method for multi-cancer classification, which adopts the following technical solution:

[0009] A machine learning method for multi-cancer classification comprises the following steps:

[0010] Preprocess the samples and then divide them into training set and validation set;

[0011] Input the samples in the training set into multiple pre-divided subgroups, each subgroup has two categories, classify the samples in each subgroup using multiple binary classification models, calculate the AUC value of each binary classification model, compare the AUC values ​​of each binary classification model, and select the binary classification model with the largest AUC value as the optimal model for the subgroup; output the predicted probability of the category to which the cancer type in each subgroup belongs;

[0012] The predicted probability is used as the input of the multi-classification model, and the samples in each sub-group are classified by the multi-classification model. The AUC value of each multi-classification model is calculated, and the AUC value of each multi-classification model is compared. The multi-classification model with the highest AUC value is selected as the overall optimal model.

[0013] In a specific implementation scheme, different samples are subjected to different pretreatment methods:

[0014] If the sample is a tumor marker test result, the ratio of the measured value to the reference value is used, a ratio less than or equal to 1 is a negative result, and a ratio greater than 1 is a positive result;

[0015] If the sample is a CT value obtained by gene methylation qPCR detection, use the internal reference gene for ΔCT conversion or ΔΔCT conversion to calculate ΔCT or ΔΔCT;

[0016] By comparing ΔCT with a preset first threshold, if ΔCT is greater than the first threshold, it is a positive result, and if ΔCT is less than or equal to the first threshold, it is a negative result;

[0017] By comparing ΔΔCT with a preset second threshold, if ΔΔCT is greater than the second threshold, it is a positive result, and if ΔΔCT is less than or equal to the second threshold, it is a negative result.

[0018] In a specific feasible implementation scheme, after selecting the sub-optimal model of each sub-grouping, the hyperparameters of each sub-optimal model are optimized; after selecting the optimal model of the multi-classification model, the hyperparameters of the overall optimal model are optimized.

[0019] In a specific implementation scheme, the binary classification model uses algorithms including logistic regression, k-nearest neighbor, naive Bayes, support vector machine, random forest, and an integrated algorithm of random forest and logistic regression.

[0020] In a specific embodiment, the multi-classification model uses algorithms including logistic regression, k-nearest neighbor, naive Bayes, support vector machine, decision tree, random forest and neural network.

[0021] In a specific implementation scheme, the method of dividing the subgroups is:

[0022] The patients were divided based on different pathological diagnoses, and similar pathological diagnoses were merged according to the time series of tumor development.

[0023] In a specific feasible implementation scheme, after obtaining the overall optimal model, the overall optimal model is used to predict the samples in the validation set, the prediction results are compared with the pathological diagnosis, and the evaluation quantitative indicators of the overall optimal model are calculated by constructing a confusion matrix.

[0024] The present invention also provides a machine learning device for multi-cancer classification, which adopts the following technical solution:

[0025] A machine learning device for multi-cancer classification, comprising

[0026] one or more processors; memory;

[0027] and one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs include instructions, which, when executed by the processor, enable the machine learning device for multi-cancer classification to perform the above-mentioned machine learning method for multi-cancer classification.

[0028] The present invention also provides an electronic device, which adopts the following technical solution:

[0029] An electronic device comprises: at least one processor; and a memory connected to the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned machine learning method for multi-cancer classification.

[0030] The present invention also provides a non-transitory computer-readable storage medium storing computer instructions, which adopts the following technical solution:

[0031] A non-transitory computer-readable storage medium storing computer instructions, characterized in that the computer instructions are used to enable a computer to execute the above-mentioned machine learning method for classification of multiple cancer types.

[0032] In summary, the present invention includes at least one of the following beneficial technical effects:

[0033] 1. Obtain different clinical data, and automatically select the optimal algorithm model to analyze and predict the clinical data based on the clinical data, which can accurately analyze and predict different types of cancer. At the same time, it integrates multiple algorithm models, which can automatically adjust the algorithm according to the different clinical data obtained, without the need for repeated testing of patients, and facilitates the analysis and prediction of various types of cancer based on different clinical data.

[0034] 2. The traditional stacking method is to use different algorithms for the same clinical data and then stack the algorithm results. In the present invention, subgroups are first established based on different pathological diagnoses, and then the classification results of each subgroup are stacked, which amplifies the differences between different pathological diagnoses and improves the accuracy of cancer analysis and prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flowchart of the machine learning method for multi-cancer classification.

[0036] Figure 2 is the confusion matrix of the embodiment. DETAILED DESCRIPTION

[0037] The machine learning method for multi-cancer classification provided by the present invention can be applied on a server or a terminal. The server can be a physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (Content Delivery Network, CDN), and basic cloud computing services such as big data and artificial intelligence platforms. The terminal can be a user equipment (User Equipment, UE) such as a mobile phone, smart phone, laptop computer, digital broadcast receiver, personal digital assistant (PDA), tablet computer (PAD) with strong computing power, handheld device, vehicle-mounted device, wearable device, computing device or other processing equipment connected to a wireless modem, mobile station (Mobile Station, MS), mobile terminal (Mobile Terminal), etc., and the present invention is not limited here.

[0038] In order to facilitate the understanding of the technical solution of the present invention, some professional terms are explained below.

[0039] DNA: Deoxyribonucleic acid, which stands for deoxyribonucleic acid. cfDNA refers to the DNA that exists freely in peripheral blood and is not contained in intact cell structures.

[0040] PCR: Polymerase Chain Reaction, stands for polymerase chain reaction. qPCR refers to fluorescent quantitative PCR, which is a method of measuring the total amount of products after each polymerase chain reaction (PCR) cycle in a DNA amplification reaction using fluorescent chemicals. It is a method of quantitatively analyzing specific DNA sequences in the sample to be tested using internal or external reference methods. During the PCR amplification process, the PCR process is detected in real time through fluorescent signals. Since there is a linear relationship between the CT value of the template and the starting copy number of the template during the exponential period of PCR amplification, it becomes the basis for quantification.

[0041] CT: Cycle Threshold, stands for cycle threshold. ΔCT is an indicator to measure gene expression level, which indicates the difference in CT value between the target gene and the reference gene. ΔΔCT is the difference in ΔCT value of the target gene in different samples, which is used to compare the changes in target gene expression level between different samples or different experimental groups.

[0042] ROC: Receiver operating characteristic curve represents ROC curve, or receiver operating characteristic curve.

[0043] The following is combined with Figure 1-2 The present invention is described in further detail.

[0044] Reference Figure 1 , the machine learning method for multi-cancer classification includes the following steps:

[0045] S100, preprocess the samples and then divide the samples into a training set and a validation set.

[0046] Obtain clinical data as samples. The clinical data can be genomic data such as cancer-related gene mutations, methylation, expression levels, or test results of commonly used tumor markers in clinical practice.

[0047] For different clinical data, the test items are used as features, the measured values ​​of the test items are used as variables, and different preprocessing methods are used. For example: if the sample is a tumor marker test result, then a tumor marker test result is a sample, the marker in the tumor marker test result is a feature, and the concentration of the tumor marker is a variable. When the sample is a tumor marker test result, the ratio of the measured value to the reference value can be used, where the reference value is the reference value of the tumor marker in a normal human body. If the ratio of the measured value to the reference value is greater than 1, it is a positive result, indicating that the concentration of the corresponding tumor marker in the patient's body exceeds the standard; if the ratio of the measured value to the reference value is less than or equal to 1, it is a negative result, indicating that the concentration of the corresponding tumor marker in the patient's body is at a normal level.

[0048] The ratio of the measured value to the reference value can be further processed in binary mode. In this case, the output results are only 0 and 1. When 0 is output, it is a negative result, and when 1 is output, it is a positive result.

[0049] The sample is the CT value obtained by qPCR detection of gene methylation, with the gene methylation state as the feature and the CT value as the variable. The internal reference gene can be used for ΔCT conversion, ΔCT is calculated, and ΔCT is compared with the preset first threshold value, which is the methylation level of the gene. If ΔCT is greater than the first threshold value, it indicates that the methylation level of the gene in the sample is high, which is a positive result. If ΔCT is less than or equal to the first threshold value, it is a negative result.

[0050] The calculation formula of ΔCT is:

[0051]

[0052] In the formula, Indicates the CT value of the target gene in the sample to be tested; Indicates the CT value of the reference gene in the sample to be tested.

[0053] When the sample is a CT value obtained by qPCR detection of gene methylation, an internal reference gene can also be used for ΔΔCT conversion, ΔΔCT is calculated, and ΔΔCT is compared with a preset second threshold value, where the second threshold value is the methylation level of the gene. If ΔΔCT is greater than the second threshold value, it indicates that the methylation level of the gene in the sample is high, which is a positive result; if ΔΔCT is less than or equal to the second threshold value, it is a negative result.

[0054] The calculation formula of ΔΔCT is:

[0055]

[0056] In the formula, It indicates the CT value of the target gene in the positive reference sample; Represents the CT value of the reference gene in the positive reference sample.

[0057] After the sample completes the above transformation, the missing values ​​of the sample are filled by the k-nearest neighbor algorithm. If the missing value of a variable of a feature in the sample accounts for more than 20%, the missing value of the feature will no longer be filled and the feature will be deleted. The remaining variables are standardized by the Z-Score algorithm, and the mean of the processed variables is 0 and the variance is 1. For ease of understanding, taking the sample as a tumor marker test result as an example, one sample is a tumor marker test result, and the features can be CA72-4 (carbohydrate antigen 72-4), AFP (alpha-fetoprotein), CEA (carcinoembryonic antigen), CA12-5 (carbohydrate antigen 12-5), CA19-9 (carbohydrate antigen 19-9) test items in the tumor marker test results, and the variables are the concentrations of CA72-4, AFP, CEA, CA12-5, and CA19-9. There are cases where variable data is missing in the sample. The missing value is the variable that is missing a certain feature, that is, there is a lack of concentration data in CA72-4, AFP, CEA, CA12-5, and CA19-9. The missing value is the missing concentration of CA72-4, AFP, CEA, CA12-5, and CA19-9. The missing values ​​in the sample can be filled by the k-nearest neighbor algorithm. When the missing values ​​of a certain feature account for more than 20% of the total number of the feature variables, the feature is deleted. The variables of the remaining features are standardized by the Z-Score algorithm.

[0058] S200, storing samples in the training set into a plurality of preset sub-groups, and importing the samples in the sub-groups into a binary classification model for model training.

[0059] Subgroups are distinguished in advance according to different pathological diagnoses, and each subgroup is divided into two categories, A and B. For example: if there are three pathological diagnoses in the sample, namely gastric cancer, liver cancer and health, gastric cancer and health will be included in the same group, with health included in category A and gastric cancer included in category B. Similarly, liver cancer and health will be included in the same group, with health included in category A and liver cancer included in category B. At the same time, in order to improve the applicability of sample size, sample feature distribution, etc. during model training, and thus improve the stability of the model, when distinguishing subgroups, refer to the time series of tumor development and merge similar pathological diagnoses.

[0060] It should be noted that when there are healthy people in the subgroups, Class A is always healthy and Class B is the corresponding cancer type; when all subgroups are cancer types, similar pathological diagnoses are merged. For example, if the subgroups are liver cancer, gastric cancer and colorectal cancer, liver cancer is included in Class A, and gastric cancer and colorectal cancer are included in Class B, which can facilitate subsequent processing of the data.

[0061] The samples in the training set are input into each subgroup respectively, and the samples in each subgroup are used to train the binary classification model. The binary classification model classifies the samples in the subgroup and classifies the samples into the corresponding subgroup A and B. During the model training process, the classification results of the binary classification model are verified by repeating the k-fold cross validation algorithm, and the ROC curve is output, and then the area under the ROC curve, i.e., the AUC value, is calculated. The binary classification model uses algorithms including logistic regression, k-nearest neighbor, naive Bayes, support vector machine, random forest, and an integrated algorithm integrating random forest and logistic regression. By classifying the samples in the subgroups with different algorithms, the ROC curves of different algorithms can be obtained, thereby calculating different AUC values. The larger the AUC value, the better the model performance. The model with the largest AUC value is selected as the sub-optimal model of the subgroup. The hyperparameters of each sub-optimal model are optimized by the grid search algorithm. After the optimization, the classification results of the B class in the subgroup are quantified into the predicted probability output in the form of probability.

[0062] It should be noted that the positive results and negative results determined in step S100 are the positive results and negative results of the features in the corresponding samples. A single positive or negative feature cannot be used to judge the tumor condition. The binary classification model performs comprehensive statistical calculations on the positive results and negative results in the samples, and then classifies the samples in the corresponding sub-groups.

[0063] S300, using the predicted probability as the input of the multi-classification model for model training to obtain the overall optimal model.

[0064] The predicted probability of each subgroup for class B was used as the input of the multi-classification model. The multi-classification model predicted the cancer type for the samples in the training set based on the predicted probabilities of all subgroups. During the training of the multi-classification model, the classification results of the multi-classification model were verified by repeating the k-fold cross-validation algorithm, and the AUC value was calculated. The multi-classification model used algorithms including logistic regression, k-nearest neighbor, naive Bayes, support vector machine, decision tree, random forest and neural network. By comparing the AUC values ​​of different multi-classification models, the multi-classification model with the highest AUC value was selected as the overall optimal model. The hyperparameters of the overall optimal model were optimized by the grid search algorithm.

[0065] S400, model evaluation.

[0066] The samples in the validation set are used as the input of the sub-optimal model of each sub-group that completes the model training, and the predicted probability of the samples in each sub-group corresponding to class B is obtained. The predicted probability is used as the input of the overall optimal model to predict the samples. The prediction results are compared with the actual pathological diagnosis structure to construct a confusion matrix. The quantitative indicators of the model evaluation of the overall optimal model are calculated through the confusion matrix, and then the quantitative indicator report is output. The quantitative indicators include accuracy, area under the ROC curve, sensitivity, specificity, positive predictive value, and negative predictive value.

[0067] The accuracy calculation formula is:

[0068]

[0069] In the formula, ACC represents the accuracy rate;

[0070] TP means that the predicted results are consistent with the actual pathological diagnosis, and both patients have tumors;

[0071] TNM means that the predicted results are consistent with the actual pathological diagnosis, and both patients are healthy;

[0072] FP means that the predicted result is that the patient may have a tumor, but the actual pathological diagnosis result is healthy;

[0073] FN means that the predicted result is that the patient is healthy, but the actual pathological diagnosis result is that the patient may have a tumor.

[0074] The sensitivity calculation formula is:

[0075]

[0076] Where TPR stands for sensitivity.

[0077] The specificity calculation formula is:

[0078]

[0079] Wherein, TNR stands for specificity.

[0080] The positive predictive value was calculated as follows:

[0081]

[0082] Here, PPV stands for positive predictive value.

[0083] The negative predictive value was calculated as follows:

[0084]

[0085] Where NPV stands for positive predictive value.

[0086] Compare the quantitative index with the preset evaluation index. If the quantitative index is higher than the evaluation index or the same as the evaluation index, it means that the overall optimal model training has met the requirements and can be put into use. Otherwise, it means that the overall optimal model needs further training. If the overall optimal model needs further optimization, one or more of the following solutions will be adopted:

[0087] Further optimizing the hyperparameters in step S200 and step S300;

[0088] In step S200 and step S300, more algorithms are introduced for training;

[0089] In step S100, the number of samples is increased and the model training is performed again.

[0090] The present invention also discloses a machine learning device for multi-cancer classification, comprising one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs include instructions, which, when executed by the processor, enable the machine learning device for multi-cancer classification to perform the above-mentioned machine learning method for multi-cancer classification.

[0091] The present invention also discloses an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned machine learning method for multi-cancer classification.

[0092] The present invention also discloses a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable a computer to execute the above-mentioned machine learning method for classification of multiple cancer types.

[0093] Example:

[0094] In order to facilitate the understanding of the overall plan, the methylation PCR test results of 5 target genes in plasma and the test results of 5 tumor markers are used as samples for further explanation. The samples include precancerous lesions of colorectal cancer, gastric cancer and hepatocellular carcinoma, different tumor stages, interference samples of other cancers, and samples from healthy people. The specific number of samples is determined based on the number of AUC evaluation cases that meet statistical requirements, and the sample size is calculated based on the sample size estimation method in the diagnostic reagent research, as shown in Table 1.

[0095] Table 1

[0096]

[0097]

[0098] S100, for the methylation qPCR test results, taking into account the differences in cfDNA content of different samples and the differences between PCR experiment batches, the standardized fluorescence curve of the qPCR machine is used, that is, the standardized fluorescence value measured in each cycle. By analyzing the fluorescence curve, the CT value in the methylation PCR test results of the target gene is obtained, and then ΔCT conversion is performed, ΔCT is calculated, and ΔCT is compared with the preset first threshold. If ΔCT is greater than the first threshold, it is a positive result, and ΔCT is less than or equal to the first threshold, it is a negative result.

[0099] For the test results of five tumor markers, namely CA72-4, AFP, CEA, CA12-5 and CA19-9, taking into account the different detection standards of different tumor markers, the ratio between the measured value and the reference value was used, where a ratio less than or equal to 1 was a negative result, and a ratio greater than 1 was a positive result.

[0100] The missing values ​​of the samples were filled by the k-nearest neighbor algorithm, and the features with missing values ​​accounting for more than 20% of the total variables were deleted. The remaining variables were then standardized by the Z-Score algorithm, and the mean of the processed variables was 0 and the variance was 1.

[0101] S200, the subgroups are distinguished according to different pathological diagnoses, and when distinguishing the subgroups, similar pathological diagnoses are merged with reference to the time series of tumor development, and the samples in the training set are divided into 6 subgroups. Each subgroup is further divided into two categories, A and B. The specific groups and categories are shown in Table 2.

[0102] Table 2

[0103] Group Category A Category B 1 healthy Cancer (stomach, colorectal, and liver) 2 healthy Digestive tract cancer (gastric cancer and colorectal cancer) 3 healthy Liver cancer 4 healthy Colorectal cancer 5 healthy Gastric cancer 6 Liver cancer Digestive tract cancer (gastric cancer and colorectal cancer)

[0104] Different types of binary classification models including logistic regression, k-nearest neighbor, naive Bayes, support vector machine, random forest, and integrated algorithms integrating random forest and logistic regression were used to divide the samples in the corresponding subgroups into class A and class B for model training. During the model training process, the classification results of the binary classification model were verified by using a 5-fold cross-validation algorithm repeated 5 times, the ROC curve was output, and the AUC value was further calculated, as shown in Table 3.

[0105] Table 3

[0106]

[0107] The binary classification model with the largest AUC value in each sub-group is selected as the optimal model for the sub-group. Grid search is used to optimize the hyperparameters. The hyperparameter optimization results are shown in Table 4.

[0108] Table 4

[0109]

[0110] The binary classification algorithm that completes the hyperparameter optimization quantifies the classification results of class B in the corresponding subgroup and outputs them in the form of probability.

[0111] S300, the predicted probability of each subgroup for class B is used as the input of the multi-classification model, and the multi-classification model predicts the cancer type for the samples in the training set according to the predicted probabilities of all subgroups. During the multi-classification model training process, the classification results of the multi-classification model are verified by repeating the 5-fold cross validation algorithm 5 times, and the AUC value is calculated. The AUC value of the multi-classification model is shown in Table 5 below.

[0112] Table 5

[0113]

[0114]

[0115] The multi-classification model with the highest AUC value is selected as the sub-optimal model. The grid search algorithm is used to optimize the hyperparameters of the sub-optimal model. The optimized hyperparameters are shown in Table 6.

[0116] Table 6

[0117] Optimal multi-classification model Optimized hyperparameters Logistic Regression C=1,penalty='l2',solver='lbfgs'

[0118] Reference Figure 2, S400, the samples in the validation set are used as the input of each sub-optimal model to complete the model training, and the predicted probability of the samples in each sub-group corresponding to class B is obtained. The predicted probability is used as the input of the total optimal model to predict the samples. A confusion matrix is ​​constructed based on the prediction results of the model and the actual pathological diagnosis, and then the quantitative indicators including accuracy, area under the ROC curve, sensitivity and specificity are calculated through the confusion matrix. The specific quantitative indicators are shown in Table 7 below.

[0119] Table 7

[0120] Quantitative indicators Numeric Accuracy 0.858 Area under the ROC curve 0.847 Sensitivity 0.835 Specificity 0.878

[0121] After further analysis of the overall optimal model, it was found that most prediction errors were due to confusion in the diagnostic results between healthy people and precancerous lesions. If the results output by the overall optimal model were simplified to cancer and non-cancer, the accuracy of the overall optimal model could be further improved. The sensitivity of the overall optimal model could reach 85.1%, the specificity could reach 88.5%, the positive predictive value could reach 94.0%, and the negative predictive value could reach 75%.

[0122] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A machine learning method for multi-cancer classification, characterized by: The steps include: Preprocess the samples and then divide them into training set and validation set; Input the samples in the training set into multiple pre-divided subgroups, each subgroup has two categories, classify the samples in each subgroup using multiple binary classification models, calculate the AUC value of each binary classification model, compare the AUC values ​​of each binary classification model, and select the binary classification model with the largest AUC value as the optimal model for the subgroup; output the predicted probability of the category to which the cancer type in each subgroup belongs; The predicted probability is used as the input of the multi-classification model, and the samples in each sub-group are classified by the multi-classification model. The AUC value of each multi-classification model is calculated, and the AUC value of each multi-classification model is compared. The multi-classification model with the highest AUC value is selected as the overall optimal model.

2. The machine learning method for multi-cancer classification according to claim 1, characterized in that: Different samples use different preprocessing methods: If the sample is a tumor marker test result, the ratio of the measured value to the reference value is used, a ratio less than or equal to 1 is a negative result, and a ratio greater than 1 is a positive result; If the sample is a CT value obtained by gene methylation qPCR detection, use the internal reference gene for ΔCT conversion or ΔΔCT conversion to calculate ΔCT or ΔΔCT; By comparing ΔCT with a preset first threshold, if ΔCT is greater than the first threshold, it is a positive result, and if ΔCT is less than or equal to the first threshold, it is a negative result; By comparing ΔΔCT with a preset second threshold, if ΔΔCT is greater than the second threshold, it is a positive result, and if ΔΔCT is less than or equal to the second threshold, it is a negative result.

3. The machine learning method for multi-cancer classification according to claim 1, characterized in that: After selecting the sub-optimal model of each sub-group, optimize the hyperparameters of each sub-optimal model; After selecting the overall optimal model of the multi-classification model, the hyperparameters of the overall optimal model are optimized.

4. The machine learning method for multi-cancer classification according to claim 1, characterized in that: The binary classification model used algorithms including logistic regression, k-nearest neighbor, naive Bayes, support vector machine, random forest, and an integrated algorithm of random forest and logistic regression.

5. The machine learning method for multi-cancer classification according to claim 1, characterized in that: The multi-classification models used algorithms including logistic regression, k-nearest neighbors, naive Bayes, support vector machines, decision trees, random forests, and neural networks.

6. The machine learning method for multi-cancer classification according to claim 1, characterized in that: The method of dividing the subgroups is: The patients were divided based on different pathological diagnoses, and similar pathological diagnoses were merged according to the time series of tumor development.

7. The machine learning method for multi-cancer classification according to claim 1, characterized in that: After obtaining the overall optimal model, the samples in the validation set are predicted using the overall optimal model, the prediction results are compared with the pathological diagnosis, and the evaluation quantitative indicators of the overall optimal model are calculated by constructing a confusion matrix.

8. A machine learning device for multi-cancer classification, characterized in that: include one or more processors; Memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs include instructions, which, when executed by the processor, enable the machine learning device for multi-cancer classification to perform the machine learning method for multi-cancer classification as described in any one of claims 1-7.

9. An electronic device, characterized in that: include: at least one processor; and a memory communicatively connected to at least one processor; wherein the memory stores instructions executable by at least one processor, and the instructions are executed by at least one processor so that at least one processor can execute the machine learning method for multi-cancer classification described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to enable a computer to execute the machine learning method for multi-cancer classification according to any one of claims 1 to 7.

Images