A bladder tumor prognosis prediction method based on machine learning

Abstract

The present invention relates to the field of data prediction technology, and in particular to a bladder tumor prognosis prediction method based on machine learning. The method comprises: obtaining a medical record category and the medical record samples contained therein and their recurrence rates; constructing a number of base classifiers according to the recurrence rates; obtaining the misclassification rate of each base classifier for each medical record category, and then obtaining the prognosis prediction accuracy contribution coefficient of the medical record category; obtaining the classification convergence speed parameter of the medical record sample according to the prediction value of the medical record sample in the base classifier; obtaining the weight correction value of the medical record sample and the training weight of the medical record category according to the classification convergence speed parameter and the prognosis prediction accuracy contribution coefficient; constructing a bladder tumor patient prognosis prediction model according to the weight correction value and the training weight to obtain the prediction result. The present invention aims to solve the problem of different reference values ​​of different samples, reduce the misclassification rate of samples in the prediction model, and enhance the accuracy of the results.

Description

Technical Field

[0001] The present invention relates to the technical field of data prediction, and in particular to a bladder tumor prognosis prediction method based on machine learning. Background Art

[0002] Bladder tumors have the possibility of recurrence in their prognosis. Predicting the recurrence of bladder tumors through the medical records of bladder tumor patients has an important impact on the treatment of bladder tumor patients.

[0003] Existing prognosis predictions for bladder cancer patients are often based on machine learning algorithms, which use the classification logic of the machine learning algorithm itself to predict the prognosis of samples. However, due to the individual differences among different bladder cancer patients, the prognosis performance varies. When the medical records of bladder cancer patients are used as samples, the machine learning algorithm ignores the multidimensional feature connection between different samples and the prediction results, resulting in a high misclassification rate during classification. As a result, the trained machine learning prediction model is difficult to effectively classify samples, resulting in insufficient accuracy in predicting the tumor recurrence rate of bladder cancer patients by the machine learning prediction model. Summary of the invention

[0004] In order to solve the technical problems in the related art that the prediction model is not accurate enough, the classification effect of the base classifier on the medical record samples is poor, the base classifier training iteration in the integrated prediction model has a high misclassification rate for the medical record samples, and the iterative convergence speed of the base classifier in the integrated prediction model is slow, the present invention provides a bladder tumor prognosis prediction method based on machine learning, and the technical scheme adopted is as follows:

[0005] The present invention proposes a bladder tumor prognosis prediction method based on machine learning, the method comprising:

[0006] Acquire a plurality of medical condition file categories according to the medical condition file information of the bladder tumor patient, wherein each medical condition file category includes a plurality of medical condition file samples and the recurrence rate of each medical condition file sample;

[0007] According to the recurrence rate of each medical record sample, a label value of each medical record sample is obtained; according to the label value of each medical record sample in all medical record categories, a base classifier of a prognosis prediction model for several bladder tumor patients is obtained;

[0008] According to the preset initialized sample weight of each medical record sample, combined with the base classifier of each bladder tumor patient prognosis prediction model, the misclassification rate of each base classifier for each medical record category is obtained;

[0009] According to the misclassification rate of each medical record category by all base classifiers, the prognosis prediction accuracy contribution coefficient of each medical record category is obtained; according to the label value of each medical record sample and the predicted value predicted in each base classifier, the classification convergence speed parameter of each medical record sample is obtained; according to the classification convergence speed parameter of each medical record sample and the prognosis prediction accuracy contribution coefficient of the medical record category to which the medical record sample belongs, the weight correction value of each medical record sample is obtained; according to the weight correction values ​​of all medical record samples in each medical record category, the training weight of each medical record category is obtained;

[0010] According to the weight correction values ​​of all medical archive samples and the training weights of all medical archive samples, a bladder tumor patient prognosis prediction model is constructed, and bladder tumor patients are predicted based on the bladder tumor patient prognosis prediction model to obtain prediction results;

[0011] Furthermore, the specific method for obtaining the recurrence rate of the medical record sample is:

[0012] Any medical record category is recorded as the target medical record category, any medical record sample in the target medical record category is recorded as the target medical record sample, the number of medical record samples that are identical to the data in the target medical record sample in the medical record information of all bladder tumor patients is recorded as the number of samples with the same physical signs as the target medical record samples, the number of medical record samples of bladder tumor recurrence that is identical to the data in the target medical record sample is recorded as the number of recurrence samples of the target medical record samples, and the ratio of the number of recurrence samples to the number of samples with the same physical signs is recorded as the recurrence rate of the target medical record cases.

[0013] Furthermore, the specific method of obtaining the tag value of the medical record sample is:

[0014] The recurrence rate of each medical record sample in each medical record category is used as the label value of each medical record sample to obtain the label values ​​of all medical record samples in each medical record category;

[0015] Arrange the label values ​​of each medical record sample in each medical record category in ascending order to obtain a label value sequence for each medical record category;

[0016] Adopting the multi-threshold segmentation algorithm otsu to adaptively segment the label value sequence of each medical record category into multiple segments, and obtaining several segmentation sequences of the label value sequence of each medical record category;

[0017] The maximum value of the label value in each segmentation sequence of each medical record category is used as the label value of each medical record sample in the segmentation sequence, and the label value of each medical record sample under each medical record category in the medical record information of all bladder tumor patients is obtained.

[0018] Furthermore, the base classifiers of several bladder tumor patient prognosis prediction models are obtained according to the label value of each medical record sample in all medical record categories, including:

[0019] All medical record samples in all medical record categories are used as first input data and input into the decision tree algorithm;

[0020] The first input data is divided into a training set and a validation set according to a preset ratio, the label value of each medical record sample is used as a leaf node of a decision tree, and the decision tree is used for training and classification to obtain a base classifier.

[0021] Furthermore, the specific method for obtaining the misclassification rate of each base classifier for each medical record category is as follows:

[0022] Record any medical record category as the target medical record category, and record any base classifier as the target base classifier;

[0023] The numerical values ​​of all the medical record samples in the target medical record category and the initialization sample weights are used as the second input data, and input into the target base classifier of the bladder tumor patient prognosis prediction model for prediction, so as to obtain the recurrence rate prediction value of the target base classifier for each medical record sample in the target medical record category;

[0024] The absolute value of the difference between the recurrence rate prediction value of the target base classifier for each medical record sample in the target medical record category and the label value of the corresponding medical record sample is recorded as the prediction difference value of the target base classifier for each medical record sample in the target medical record category;

[0025] Preset a misclassification threshold; obtain the number of medical record samples that meet the prediction difference value greater than the misclassification threshold among all medical record samples of the target medical record category of the target base classifier, and record it as the number of misclassified samples of the target medical record category by the target base classifier; record the ratio of the number of misclassified samples of the target medical record category by the target base classifier to the total number of medical record samples of the target medical record category as the misclassification rate of the target base classifier for the target medical record category.

[0026] Furthermore, the prognosis prediction accuracy contribution coefficient of each medical record category is obtained according to the misclassification rate of each medical record category, including:

[0027] The ratio of the mean misclassification rate of all base classifiers for the target medical record category to the mean sum of the misclassification rates of all base classifiers for all medical record categories is recorded as the contribution coefficient of the prognosis prediction accuracy of the target medical record category.

[0028] Furthermore, the method of obtaining the classification convergence speed parameter of each medical condition file sample according to the predicted value and the marked value of each medical condition file sample in each base classifier includes:

[0029] The ratio of the predicted difference value of the target medical record sample in the target medical record category in the target base classifier to the maximum predicted difference value of all medical record samples in the target medical record category in the target base classifier is recorded as the misclassification degree of the target medical record sample in the target medical record category in the target base classifier;

[0030] Arrange the misclassification degrees of the target medical condition file samples in the target medical condition file category in all base classifiers according to the update order in the base classifier training process to obtain a misclassification degree sequence of the target medical condition file samples in the target medical condition file category;

[0031] Construct a two-dimensional dimensionality reduction space of target medical condition file samples in the target medical condition file category, wherein the horizontal axis is the order of the misclassification degree sequence, and the vertical axis is the value of the misclassification degree;

[0032] Using a principal component analysis algorithm to reduce the dimension of the two-dimensional reduced space of the target medical condition file samples in the target medical condition file category, obtaining a number of principal component directions of the target medical condition file samples in the target medical condition file category, and recording the projection vector of the principal component direction with the largest eigenvalue among the principal component directions as the convergence vector of the target medical condition file samples in the target medical condition file category;

[0033] The modulus of the convergence vector of the target medical file sample in the target medical file category and the extreme difference of the misclassification degree of the target medical file sample in the target medical file category in all base classifiers are recorded as the classification convergence speed parameter of the target medical file sample in the target medical file category.

[0034] Furthermore, the specific method of obtaining the weight correction value of the medical record sample is as follows:

[0035] The product of the prognosis prediction accuracy contribution coefficient of the medical record category to which the target medical record sample belongs and the classification convergence speed parameter of the target medical record sample is recorded as the weight correction value of the target medical record sample.

[0036] Furthermore, the specific method of obtaining the training weight of the medical record category is:

[0037] The median value of the weight correction values ​​of all medical record samples in each medical record category is used as the training weight of each medical record category.

[0038] Furthermore, the specific method of obtaining the prognosis prediction model for bladder tumor patients is:

[0039] All base classifiers were integrated using the AdaBoost iterative algorithm to obtain the initial model for predicting the prognosis of bladder tumor patients;

[0040] Dividing all medical record samples into a model training set and a model validation set according to a preset ratio, the model training set and the model validation set constitute the third input data of the initial model for predicting the prognosis of bladder tumor patients;

[0041] The training weights of the medical record categories are used as the category weights in the iterative algorithm AdaBoost, and the weight correction values ​​of the medical record samples are used as the sample weights of the medical record samples in the AdaBoost algorithm to obtain the iterative parameters of the initial model for predicting the prognosis of bladder tumor patients.

[0042] The third input data of the initial model for prognosis prediction of bladder tumor patients is input into the initial model for prognosis prediction of bladder tumor patients, and training is performed in combination with the iterative parameters of the initial model for prognosis prediction of bladder tumor patients. The training is stopped until the classification accuracy of the initial model for prognosis prediction of bladder tumor patients reaches the preset classification accuracy, so as to obtain the prognosis prediction model for bladder tumor patients.

[0043] The present invention has the following beneficial effects:

[0044] The present invention obtains several medical record categories and several medical record samples and their recurrence rates according to the medical record information of bladder tumor patients; obtains the label value of each medical record sample according to the recurrence rate of each medical record sample; obtains the base classifiers of several bladder tumor patient prognosis prediction models according to the label value of each medical record sample in all medical record categories, and avoids the problem of too many leaf nodes in the base classifier leading to reduced construction efficiency of the base classifier by segmenting and fusing the recurrence rate; obtains the misclassification rate of each medical record category by each base classifier in combination with the base classifier of each bladder tumor patient prognosis prediction model according to the preset initialized sample weight of each medical record sample; obtains the prognosis prediction accuracy contribution coefficient of each medical record category according to the misclassification rate of all base classifiers for each medical record category, which is used to measure the association between each medical record category and the prognosis disease; obtains the classification convergence speed parameter of each medical record sample according to the label value of each medical record sample and the predicted value predicted in each base classifier, and obtains the influence of the items contained in the sample on the classification decision speed when each medical record sample is classified; The convergence speed parameter and the contribution coefficient of the prognosis prediction accuracy of the medical record category to which the medical record sample belongs are obtained, the weight correction value of each medical record sample is obtained, and the influence weight of each medical record sample on the construction of the prediction model during the training process of the prognosis prediction model for bladder tumor patients is obtained; according to the weight correction value of all medical record samples in each medical record category, the training weight of each medical record category is obtained, and the influence weight of each medical record category on the construction of the prediction model during the training process of the prognosis prediction model for bladder tumor patients is obtained; according to the weight correction value of all medical record samples and the training weight of all medical record samples, a prognosis prediction model for bladder tumor patients is constructed, and bladder tumor patients are predicted based on the prognosis prediction model for bladder tumor patients to obtain prediction results. The present invention aims to screen out samples with little reference significance for the construction of the prognosis prediction model in all medical record samples through the weight correction value, and to apply different reference weights to different medical record categories through the training weights, so that after the prognosis prediction model for bladder tumor patients is constructed, the medical record information of different patients is quickly decided, and a more accurate prediction of the prognosis recurrence rate of bladder tumor patients is obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0046] Figure 1 A flow chart of a bladder tumor prognosis prediction method based on machine learning provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the specific implementation, structure, features and effects of a bladder tumor prognosis prediction method based on machine learning proposed by the present invention in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0048] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0049] The following is a detailed description of a specific scheme of a bladder tumor prognosis prediction method based on machine learning provided by the present invention in conjunction with the accompanying drawings.

[0050] See also Figure 1 , which shows a flow chart of a bladder tumor prognosis prediction method based on machine learning provided by an embodiment of the present invention, the method comprising:

[0051] S101: obtaining a plurality of medical condition file categories according to the medical condition file information of the bladder tumor patient, wherein each medical condition file category includes a plurality of medical condition file samples and the recurrence rate of each medical condition file sample;

[0052] This embodiment aims to construct a prognosis prediction model for bladder tumor patients using the medical record information of bladder tumor patients, so the medical record information of bladder tumor patients is collected. The medical record information of bladder tumor patients is a multi-category information containing all treatment-related data of bladder tumor patients before and after treatment, such as clinical data, genomic data, pathological image data, etc., so when constructing a prognosis prediction model for bladder tumor patients using the medical record information of bladder tumor patients, the corresponding medical record information of bladder tumor patients can be selected according to actual needs.

[0053] As a preferred example, the medical condition file information includes several medical condition file categories, including a clinical data category, a genomic data category, a pathological image data category, and a follow-up data category, wherein each medical condition file category includes data of several items.

[0054] As an example, the clinical data category is usually in the form of electronic medical records or clinical research report forms (CRFs), recording the basic information, medical history, treatment records, laboratory test results and other items of bladder cancer patients. The genomic data category is usually in FASTQ format, GTF or GFF format, storing bladder cancer patients' gene sequences, gene expression data, copy number variations, single nucleotide polymorphisms (SNPs) and other items. The pathology image data category is usually in TIFF, SVS, CZI, NDPI and other formats, storing bladder cancer patients' Whole Slide Images (WSI), tissue slice images, cell images and other items. The follow-up data category is usually in the XLSX database, storing bladder cancer patients' tracking data after treatment, such as survival status, recurrence, quality of life and other items.

[0055] Furthermore, the data of each item under each medical record category in the medical record information of all bladder tumor patients is used as an item of a medical record sample. The medical record information of all bladder tumor patients includes several medical record categories, and each medical record category includes several medical record samples. As an example: a medical record sample in any medical record category of a bladder tumor patient includes the following items: the patient's age, gender, weight, blood pressure, blood sugar, and drug usage.

[0056] Further, any one of the medical record categories is recorded as the target medical record category, any one of the medical record samples in the target medical record category is recorded as the target medical record sample, the number of medical record samples with the same data as the target medical record sample in the medical record information of all bladder tumor patients is recorded as the number of samples with the same signs of the target medical record sample, the number of medical record samples with the same data as the target medical record sample that have recurred bladder tumors is recorded as the number of recurrence samples of the target medical record sample, and the ratio of the number of recurrence samples to the number of samples with the same signs is recorded as the recurrence rate of the target medical record case. It should be noted that since the number of medical record samples with the same data in the same medical record category is the same, the recurrence rates of medical record samples with the same data are the same.

[0057] As an example, the information of any medical record sample in all medical record information is: age 45 years old, gender male, weight 90KG, blood pressure 100mmHg, blood sugar 6.5mmol / L, and drug usage 10mg. There are 100 medical record samples with the same numerical values ​​as the medical record sample among all bladder tumor patients, that is, the number of samples with the same physical signs as the medical record sample is 100, among which 10 medical record samples relapse, that is, the number of recurrence samples of the medical record sample is 10, then the recurrence rate of all medical record samples with the same numerical values ​​as the medical record sample, including the medical record sample, is 10%.

[0058] S102: obtaining a label value of each medical record sample according to the recurrence rate of each medical record sample; obtaining a base classifier of a plurality of bladder tumor patient prognosis prediction models according to the label value of each medical record sample in all medical record categories;

[0059] In this embodiment, in the process of building a prognosis prediction model for bladder tumor patients by using the medical record categories and the medical record samples therein, thereby obtaining the predicted value of the prognosis recurrence rate of the predicted bladder tumor patients, it is first necessary to construct multiple base classifiers of the prediction model. The existing prediction model generally uses the recurrence rate of each medical record sample as the label value to construct several base classifiers for the leaf nodes of the decision tree. This construction method will result in too many leaf nodes of the decision tree when the number of medical record samples is large, affecting the construction efficiency of the base classifier of the prediction model. In this embodiment, the recurrence rate of each medical record sample is segmented and fused to obtain the label value of each medical record sample, and then the label value of each medical record sample is used as the construction basis of the leaf node of the decision tree to establish the base classifier of the prognosis prediction model for bladder tumor patients.

[0060] Preferably, as an example, the method for obtaining the tag value of each medical record sample under each medical record category in the medical record information of all bladder tumor patients is as follows:

[0061] The recurrence rate of each medical record sample in each medical record category is used as the label value of each medical record sample to obtain the label values ​​of all medical record samples in each medical record category;

[0062] Arrange the label values ​​of each medical record sample in each medical record category in ascending order to obtain a label value sequence for each medical record category;

[0063] The label value sequence of each medical record category is adaptively segmented into multiple segments using the multi-threshold segmentation algorithm otsu to obtain several segmentation sequences of the label value sequence of each medical record category. The multi-threshold segmentation algorithm is a well-known technology and will not be described in detail in this embodiment.

[0064] The maximum value of the label value in each segmentation sequence of each medical record category is used as the label value of each medical record sample in the segmentation sequence, and the label value of each medical record sample under each medical record category in the medical record information of all bladder tumor patients is obtained.

[0065] It should be noted that when the amount of data of the medical record samples in the same medical record category is large, all the medical record samples have a large number of different recurrence rates. Constructing a decision tree with a large number of different recurrence rates as label values ​​will result in too many leaf nodes in the decision tree, so that the base classifier constructed by the decision tree needs to provide a large amount of computing resources when classifying the medical record samples. However, in the items corresponding to the same medical record samples in the same medical record category of actual bladder tumor patients, since the manifestations of the disease are simple and the same when refined to a single item, there are a large number of medical record samples with similar data and similar recurrence rates. The multi-threshold segmentation algorithm can divide the data according to different data categories as much as possible, so as to maximize the difference between the division results. Therefore, in this embodiment, among the medical record information of all bladder tumor patients, the medical record samples with similar recurrence rates in the same medical record category are sorted and segmented with the recurrence rate as the label value using the multi-threshold segmentation algorithm. The maximum value of the label values ​​of multiple medical record samples in the same segmentation sequence is used as the label value of all medical record samples in the segmentation sequence. The leaf nodes of the decision tree are constructed by replacing the label value of each medical record sample with the maximum value of the segmentation sequence, which greatly reduces the generation of decision tree leaf nodes when constructing the decision tree, and establishes the decision tree more efficiently.

[0066] Furthermore, the number of base classifiers constructed in this embodiment is described by taking 100 as an example, and other embodiments may set other numbers of base classifiers. As an example, the construction method of each base classifier of the bladder tumor patient prognosis prediction model is as follows:

[0067] All medical record samples in all medical record categories are used as first input data and input into the decision tree algorithm;

[0068] The first input data is divided into a training set and a validation set according to a preset ratio, and the label value of each medical record sample is used as a leaf node of a decision tree. The decision tree is used for training and classification to obtain a base classifier. It should be noted that the preset ratio of the training set and the validation set described in this embodiment is described using four to one as an example. The preset ratios of other embodiments may be set to other values, which are not specifically limited in this embodiment.

[0069] Similarly, the base classifier of the prognosis prediction model for all bladder tumor patients is obtained.

[0070] Furthermore, all base classifiers were integrated using the AdaBoost iterative algorithm to obtain an initial prognosis prediction model for bladder tumor patients, in which the construction method of each base classifier was the same.

[0071] S103: According to the preset initialized sample weight of each medical record sample, combined with the base classifier of each bladder tumor patient prognosis prediction model, the misclassification rate of each base classifier for each medical record category is obtained.

[0072] It should be noted that after obtaining the initial bladder tumor patient prognosis prediction model, it is necessary to perform machine learning training on the initial bladder tumor patient prognosis prediction model, that is, after multiple iterations of classification of the base classifier, obtain the weight of the influence of each base classifier on the output result of the final bladder tumor patient prognosis prediction model. In the training process of each base classifier, the sample weight of the medical record sample is used to distinguish the accuracy of the sample, that is, the sample weight of the normal sample participating in the training is larger, and the sample weight of the noise sample is smaller, thereby determining that the training result of the base classifier is closer to the data value of the normal sample, and the misclassification rate is used to describe the classification error of the base classifier during the training process. Therefore, this embodiment presets the initialization sample weight for the medical record samples in the same medical record category participating in the training of the base classifier, and then trains the medical record samples participating in the training of the base classifier according to the initialization sample weight, obtains the recurrence rate prediction value of each medical record sample in the same medical record category for the base classifier, and then obtains the misclassification rate of the base classifier for the medical record category according to the difference between the recurrence rate prediction value and the actual value of the medical record sample in the same medical record category.

[0073] Preferably, this embodiment takes any base classifier and any medical record category as an example to calculate the misclassification rate of each base classifier for each medical record category. The method for calculating the misclassification rate of other base classifiers and medical record categories is the same as the example method. Then the steps for obtaining the misclassification rate of each base classifier for each medical record category are as follows:

[0074] It should be further explained that before using the medical record information of bladder tumor patients to predict the prognosis of bladder tumor patients, each medical record sample in all medical record categories obtained may be related to the bladder tumor prognosis status of bladder tumor patients; therefore, this embodiment presets the initialization sample weights of all medical record samples in the medical record information of bladder tumor patients to be equal, so that each medical record sample has an equal influence on the prognosis status of bladder tumor at the beginning of prediction, and ensures that each medical record sample contributes evenly to the training process in the initial stage of the training process.

[0075] Specifically, the initialization sample weight of each medical record sample is obtained as follows:

[0076] The reciprocal of the number of all medical record samples in all medical record categories is recorded as the initialization sample weight of each medical record sample.

[0077] It should be noted that this embodiment only provides a method for presetting the weights of initialized samples for medical record samples, and other embodiments may select other presetting methods. This embodiment does not specifically limit the method for presetting the weights of initialized samples.

[0078] Furthermore, the misclassification rate of each base classifier for each medical record category is obtained as follows:

[0079] Denote any base classifier as the target base classifier;

[0080] The numerical values ​​of all the medical record samples in the target medical record category and the initialization sample weights are used as the second input data, and input into the target base classifier of the bladder tumor patient prognosis prediction model for prediction, so as to obtain the recurrence rate prediction value of the target base classifier for each medical record sample in the target medical record category;

[0081] The absolute value of the difference between the recurrence rate prediction value of the target base classifier for each medical record sample in the target medical record category and the label value of the corresponding medical record sample is recorded as the prediction difference value of the target base classifier for each medical record sample in the target medical record category;

[0082] A misclassification threshold is preset. The misclassification threshold of this embodiment is described by taking 0.05 as an example. Other embodiments set other values, which are not specifically limited in this embodiment. The number of medical record samples satisfying the predicted difference value greater than the misclassification threshold among all medical record samples of the target medical record category of the target base classifier is obtained, and recorded as the number of misclassified samples of the target medical record category of the target base classifier. The ratio of the number of misclassified samples of the target medical record category of the target base classifier to the total number of medical record samples of the target medical record category is recorded as the misclassification rate of the target base classifier for the target medical record category.

[0083] Similarly, the misclassification rate of each base classifier for each medical record category is obtained.

[0084] It should be noted that when the base classifier classifies each medical record sample in different medical record categories, since the correlation between the medical record samples in different medical record categories and the prognostic condition is not completely similar, and the correlation between the medical record samples and the prognostic condition is only a vague connection, the base classifier may misclassify the medical record samples. The higher the misclassification rate of the base classifier for the medical record samples in different medical record categories, the more vague the correlation between the corresponding medical record category and the prognostic condition.

[0085] S104: Obtain the prognosis prediction accuracy contribution coefficient of each medical record category according to the misclassification rate of all base classifiers for each medical record category; obtain the classification convergence speed parameter of each medical record sample according to the label value of each medical record sample and the predicted value predicted in each base classifier; obtain the weight correction value of each medical record sample according to the classification convergence speed parameter of each medical record sample and the prognosis prediction accuracy contribution coefficient of the medical record category to which the medical record sample belongs; obtain the training weight of each medical record category according to the weight correction values ​​of all medical record samples in each medical record category;

[0086] It should be noted that the use of machine learning to predict the prognosis of bladder tumors is a process of integrating the prediction model using all base classifiers. In the existing use of the base classifier integrated prediction model, there are great differences in the information in the medical record categories during the input and update process of the medical record samples, that is, the prediction references provided by all medical record categories when predicting the prognosis of patients are different, and all items in the same medical record sample in the same medical record category have the same meaning for the classification result in the same base classifier, but in different base classifiers, there are errors in the classification results due to the differences in the base classifiers, that is, it does not take into account that the items contained in the medical record samples are multidimensional information, which leads to errors in the medical record samples. points, and the actual prognosis of bladder tumor patients cannot be accurately predicted based on the medical record samples; therefore, this embodiment first obtains the prognosis prediction accuracy contribution coefficient of each medical record category according to the misclassification rate of each medical record category in all base classifiers, and obtains the classification convergence speed parameter of each medical record sample according to the difference in prediction results of all medical record samples in the same medical record category in all base classifiers, which is used to measure the classification speed of the medical record sample. The prognosis prediction accuracy contribution coefficient of the medical record category is combined with the classification convergence speed parameter of all medical record samples therein to obtain the training weight of each medical record category, which represents the reference value of each medical record category to the training result.

[0087] Preferably, as an example, the method of obtaining the prognosis prediction accuracy contribution coefficient of each medical record category is as follows:

[0088] The ratio of the mean of the misclassification rates of all base classifiers for the target medical record category to the sum of the mean of the misclassification rates of all base classifiers for all medical record categories is recorded as the prognosis prediction accuracy contribution coefficient of the target medical record category. Similarly, the prognosis prediction accuracy contribution coefficient of each medical record category is obtained.

[0089] It should be noted that different medical record categories describe the condition of bladder tumor patients from different perspectives, and the accuracy and information value of the descriptions vary. Therefore, in this embodiment, the error rate of the medical record category in all base classifiers is used to represent the accuracy of the description of the condition of bladder tumor patients. The medical record category with a larger error rate contributes less to the prediction accuracy when predicting the prognosis of bladder tumors.

[0090] Preferably, as an example, the classification convergence speed parameter of each medical record sample is obtained as follows:

[0091] The ratio of the predicted difference value of the target medical file sample in the target medical file category in the target base classifier to the maximum predicted difference value of all medical file samples in the target medical file category in the target base classifier is recorded as the misclassification degree of the target medical file sample in the target medical file category in the target base classifier.

[0092] Furthermore, the misclassification degrees of the target medical file samples in the target medical file category in all base classifiers are arranged according to the update order during the base classifier training process to obtain a sequence of the misclassification degrees of the target medical file samples in the target medical file category. It should be noted that the update order of the base classifiers described in this embodiment is generated by the AdaBoost algorithm.

[0093] Since the sequence is a two-dimensional sequence, the principal component analysis algorithm can reduce the dimensionality of multidimensional data while retaining the data information. Therefore, a two-dimensional dimensionality reduction space of the target medical record samples in the target medical record category is constructed, where the horizontal axis is the order of the misclassification degree sequence and the vertical axis is the value of the misclassification degree.

[0094] Furthermore, the principal component analysis algorithm is used to reduce the dimensionality of the two-dimensional reduced space of the target medical file samples in the target medical file category to obtain several principal component directions of the target medical file samples in the target medical file category, and the projection vector of the principal component direction with the largest eigenvalue among the principal component directions is recorded as the convergence vector of the target medical file samples in the target medical file category.

[0095] The modulus of the convergence vector of the target medical record sample in the target medical record category and the extreme difference of the misclassification degree of the target medical record sample in the target medical record category in all base classifiers are recorded as the classification convergence speed parameter of the target medical record sample in the target medical record category. Similarly, the classification convergence speed parameter of each medical record sample is obtained.

[0096] It should be noted that different medical record samples have different degrees of correlation with the results of bladder tumor prognosis. When the degree of correlation between the medical record sample and the results of bladder tumor prognosis is high, the medical record sample has a lower misclassification rate and a higher convergence speed in all base classifiers. In this case, when the base classifier is iteratively trained, the medical record sample should be given a higher weight, so that the prediction result of the prediction model is closer to the classification result of the medical record sample, thereby improving the accuracy of the prediction model and the efficiency of the calculation speed. When the degree of correlation between the medical record sample and the results of bladder tumor prognosis is low, the medical record sample must have a higher misclassification rate and a lower convergence speed in all base classifiers. In this case, when the base classifier is iteratively trained, the medical record sample should be given a lower weight, so that the prediction result of the prediction model is less close to the classification result of the medical record sample, thereby reducing the inaccuracy of the prediction model and the high calculation delay.

[0097] As an example, the weight correction value of each medical record sample is obtained as follows:

[0098] The product of the prognosis prediction accuracy contribution coefficient of the medical record category to which the target medical record sample belongs and the classification convergence speed parameter of the target medical record sample is recorded as the weight correction value of the target medical record sample.

[0099] It should be noted that the faster the classification convergence speed of any medical record sample, the higher the contribution coefficient of the medical record category to the prognosis prediction accuracy of the medical record sample, which means that the medical record category to which the medical record sample belongs provides more training support for the prognosis prediction accuracy, and the medical record sample is more closely related to the prognosis result of bladder tumor. Therefore, it is necessary to give the medical record sample a higher weight in the subsequent base classifier classification training to improve the prediction accuracy and computational efficiency of the machine learning model.

[0100] As another example, the training weights for each category of medical records are obtained as follows:

[0101] The median value of the weight correction values ​​of all medical record samples in each medical record category is used as the training weight of each medical record category.

[0102] It should be noted that different categories of medical record categories contribute differently to the accuracy of bladder tumor prediction results. The reason for this phenomenon is determined by the medical record samples in the medical record category. The median of all medical record samples can more accurately reflect the relationship between the medical record category and the accuracy of bladder tumor prediction results. Therefore, the median of the weight correction values ​​of all medical record samples in any category of medical record category is used as the training weight of the medical record category of that category, which is used for the weight of the medical record category when the machine learning prediction model is updated.

[0103] S105: constructing a bladder tumor patient prognosis prediction model according to the weight correction values ​​of all medical record samples and the training weights of all medical record samples, and performing prediction on bladder tumor patients based on the bladder tumor patient prognosis prediction model to obtain prediction results.

[0104] It should be noted that after obtaining the weight correction values ​​of the medical record samples and the training weights of the medical record categories in the base classifier training process, the prognosis prediction model for bladder tumor patients is obtained according to the weight correction values ​​of the medical record samples and the training weights of the medical record categories, and then the prognosis prediction results for bladder tumor patients are obtained according to the bladder tumor patient prognosis prediction model.

[0105] As an example, the specific process of constructing a prognosis prediction model for bladder cancer patients is as follows:

[0106] All base classifiers were integrated using the AdaBoost iterative algorithm to obtain the initial model for predicting the prognosis of bladder tumor patients;

[0107] According to the preset ratio, all medical record samples are divided into a model training set and a model verification set, and the model training set and the model verification set constitute the third input data of the initial model for predicting the prognosis of bladder tumor patients; it should be noted that the preset ratio of the model training set and the model verification set described in this embodiment is described using four to one as an example, and the preset ratios in other embodiments may be set to other values, which are not specifically limited in this embodiment.

[0108] Furthermore, the AdaBoost algorithm includes category weights of data categories and sample weights of samples. In this embodiment, the training weights of the medical record categories are used as category weights in the iterative algorithm AdaBoost, and the weight correction values ​​of the medical record samples are used as sample weights of the medical record samples in the AdaBoost algorithm to obtain iterative parameters of the initial model for predicting the prognosis of bladder tumor patients.

[0109] The third input data of the initial model for prognosis prediction of bladder tumor patients is input into the initial model for prognosis prediction of bladder tumor patients, and training is performed in combination with the iterative parameters of the initial model for prognosis prediction of bladder tumor patients. The training is stopped until the classification accuracy of the initial model for prognosis prediction of bladder tumor patients reaches a preset classification accuracy of 0.001, and the prognosis prediction model for bladder tumor patients is obtained.

[0110] As another example, the specific process of obtaining the results of prognosis prediction for bladder cancer patients is as follows:

[0111] The medical record information of the bladder tumor patients required for bladder tumor prognosis prediction is used as input and input into the bladder tumor patient prognosis prediction model to obtain the bladder tumor patient prognosis prediction result.

[0112] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0113] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

Claims

1. A bladder tumor prognosis prediction method based on machine learning, characterized in that: The method comprises: Acquire a plurality of medical condition file categories according to the medical condition file information of the bladder tumor patient, wherein each medical condition file category includes a plurality of medical condition file samples and the recurrence rate of each medical condition file sample; According to the recurrence rate of each medical record sample, a label value of each medical record sample is obtained; according to the label value of each medical record sample in all medical record categories, a base classifier of a prognosis prediction model for several bladder tumor patients is obtained; According to the preset initialized sample weight of each medical record sample, combined with the base classifier of each bladder tumor patient prognosis prediction model, the misclassification rate of each base classifier for each medical record category is obtained; According to the misclassification rate of all base classifiers for each medical record category, the contribution coefficient of the prognosis prediction accuracy of each medical record category is obtained; According to the label value of each medical condition file sample and the predicted value predicted in each base classifier, the classification convergence speed parameter of each medical condition file sample is obtained; the classification convergence speed parameter of each medical condition file sample is obtained according to the predicted value and label value of each medical condition file sample in each base classifier, including: The ratio of the predicted difference value of the target medical record sample in the target medical record category in the target base classifier to the maximum predicted difference value of all medical record samples in the target medical record category in the target base classifier is recorded as the misclassification degree of the target medical record sample in the target medical record category in the target base classifier; Arrange the misclassification degrees of the target medical condition file samples in the target medical condition file category in all base classifiers according to the update order in the base classifier training process to obtain a misclassification degree sequence of the target medical condition file samples in the target medical condition file category; Construct a two-dimensional dimensionality reduction space of target medical condition file samples in the target medical condition file category, wherein the horizontal axis is the order of the misclassification degree sequence, and the vertical axis is the value of the misclassification degree; Using a principal component analysis algorithm to reduce the dimension of the two-dimensional reduced space of the target medical condition file samples in the target medical condition file category, obtaining a number of principal component directions of the target medical condition file samples in the target medical condition file category, and recording the projection vector of the principal component direction with the largest eigenvalue among the principal component directions as the convergence vector of the target medical condition file samples in the target medical condition file category; The modulus of the convergence vector of the target medical condition file sample in the target medical condition file category and the range of the misclassification degree of the target medical condition file sample in the target medical condition file category in all base classifiers are recorded as the classification convergence speed parameter of the target medical condition file sample in the target medical condition file category; According to the classification convergence speed parameter of each medical record sample and the prognosis prediction accuracy contribution coefficient of the medical record category to which the medical record sample belongs, the weight correction value of each medical record sample is obtained; the specific method for obtaining the weight correction value of the medical record sample is: The product of the prognosis prediction accuracy contribution coefficient of the medical record category to which the target medical record sample belongs and the classification convergence speed parameter of the target medical record sample is recorded as the weight correction value of the target medical record sample; According to the weight correction values ​​of all medical record samples in each medical record category, the training weight of each medical record category is obtained; the specific method of obtaining the training weight of the medical record category is: The median value of the weight correction values ​​of all medical record samples in each medical record category is used as the training weight of each medical record category; According to the weight correction values ​​of all medical record samples and the training weights of all medical record categories, a prognosis prediction model for bladder tumor patients is constructed, and bladder tumor patients are predicted based on the prognosis prediction model for bladder tumor patients to obtain prediction results.

2. The method for predicting bladder tumor prognosis based on machine learning according to claim 1, characterized in that: The specific method for obtaining the recurrence rate of the medical record sample is as follows: Any medical record category is recorded as the target medical record category, any medical record sample in the target medical record category is recorded as the target medical record sample, the number of medical record samples that are identical to the data in the target medical record sample in the medical record information of all bladder tumor patients is recorded as the number of samples with the same physical signs as the target medical record samples, the number of medical record samples of bladder tumor recurrence that is identical to the data in the target medical record sample is recorded as the number of recurrence samples of the target medical record samples, and the ratio of the number of recurrence samples to the number of samples with the same physical signs is recorded as the recurrence rate of the target medical record cases.

3. The method for predicting bladder tumor prognosis based on machine learning according to claim 1, characterized in that: The specific method of obtaining the tag value of the medical record sample is as follows: The recurrence rate of each medical record sample in each medical record category is used as the label value of each medical record sample to obtain the label values ​​of all medical record samples in each medical record category; Arrange the label values ​​of each medical record sample in each medical record category in ascending order to obtain a label value sequence for each medical record category; Adopting the multi-threshold segmentation algorithm otsu to adaptively segment the label value sequence of each medical record category into multiple segments, and obtaining several segmentation sequences of the label value sequence of each medical record category; The maximum value of the label value in each segmentation sequence of each medical record category is used as the label value of each medical record sample in the segmentation sequence, and the label value of each medical record sample under each medical record category in the medical record information of all bladder tumor patients is obtained.

4. The method for predicting bladder tumor prognosis based on machine learning according to claim 1, characterized in that: The base classifiers of several bladder tumor patient prognosis prediction models are obtained according to the label value of each medical record sample in all medical record categories, including: All medical record samples in all medical record categories are used as first input data and input into the decision tree algorithm; The first input data is divided into a training set and a validation set according to a preset ratio, the label value of each medical record sample is used as a leaf node of a decision tree, and the decision tree is used for training and classification to obtain a base classifier.

5. The method for predicting bladder tumor prognosis based on machine learning according to claim 1, characterized in that: The specific method for obtaining the misclassification rate of each base classifier for each medical record category is as follows: Record any medical record category as the target medical record category, and record any base classifier as the target base classifier; The numerical values ​​of all the medical record samples in the target medical record category and the initialization sample weights are used as the second input data, and input into the target base classifier of the bladder tumor patient prognosis prediction model for prediction, so as to obtain the recurrence rate prediction value of the target base classifier for each medical record sample in the target medical record category; The absolute value of the difference between the recurrence rate prediction value of the target base classifier for each medical record sample in the target medical record category and the label value of the corresponding medical record sample is recorded as the prediction difference value of the target base classifier for each medical record sample in the target medical record category; Preset a misclassification threshold; obtain the number of medical record samples that meet the prediction difference value greater than the misclassification threshold among all medical record samples of the target medical record category of the target base classifier, and record it as the number of misclassified samples of the target medical record category by the target base classifier; record the ratio of the number of misclassified samples of the target medical record category by the target base classifier to the total number of medical record samples of the target medical record category as the misclassification rate of the target base classifier for the target medical record category.

6. The method for predicting bladder tumor prognosis based on machine learning according to claim 1, characterized in that: The method of obtaining the prognosis prediction accuracy contribution coefficient of each medical condition file category according to the misclassification rate of each medical condition file category includes: The ratio of the mean misclassification rate of all base classifiers for the target medical record category to the mean sum of the misclassification rates of all base classifiers for all medical record categories is recorded as the contribution coefficient of the prognosis prediction accuracy of the target medical record category.

7. The method for predicting bladder tumor prognosis based on machine learning according to claim 1, characterized in that: The specific method of obtaining the prognosis prediction model for bladder tumor patients is as follows: All base classifiers were integrated using the AdaBoost iterative algorithm to obtain the initial model for prognosis prediction of bladder tumor patients; Dividing all medical record samples into a model training set and a model validation set according to a preset ratio, the model training set and the model validation set constitute the third input data of the initial model for predicting the prognosis of bladder tumor patients; The training weights of the medical record categories are used as the category weights in the iterative algorithm AdaBoost, and the weight correction values ​​of the medical record samples are used as the sample weights of the medical record samples in the AdaBoost algorithm to obtain the iterative parameters of the initial model for predicting the prognosis of bladder tumor patients. The third input data of the initial model for prognosis prediction of bladder tumor patients is input into the initial model for prognosis prediction of bladder tumor patients, and training is performed in combination with the iterative parameters of the initial model for prognosis prediction of bladder tumor patients. The training is stopped until the classification accuracy of the initial model for prognosis prediction of bladder tumor patients reaches the preset classification accuracy, so as to obtain the prognosis prediction model for bladder tumor patients.

Images