CN110895969B - Atrial fibrillation prediction decision tree and its pruning method - Google Patents

Abstract

A atrial fibrillation prediction decision tree and a pruning method thereof belong to the field of data processing, and in order to solve the problem of constructing a decision tree to mine out indexes influencing atrial fibrillation prediction, a root node A peak in the decision tree is the peak of information gain, the attribute is the maximum information gain, the normal range of the peak is 41 to 87, the first branch of the decision tree is the value of the A peak when a < = 0, and the atrial fibrillation of a patient is judged as the atrial fibrillation of the patient because the data has no non-0 number, namely when a = 0; when a >0, the ef attribute needs to be considered continuously, when the ef value is smaller than 58, the patient is judged to be normal, and the decision tree has important guiding significance for determining some important reference indexes for influencing atrial fibrillation.

Description

Atrial fibrillation prediction decision tree and its pruning method

Technical field

The invention belongs to the field of data processing and relates to a method for constructing a decision tree for predicting atrial fibrillation and the decision tree.

Background technique

Atrial fibrillation is a supraventricular tachyarrhythmia characterized by rapid, disordered atrial electrical activity. The main manifestations of atrial fibrillation on the electrocardiogram are: P wave disappears and is replaced by irregular atrial fibrillation waves; RR interval is absolutely irregular (when atrioventricular conduction is present). This is also the main basis used to judge atrial fibrillation in medicine and other aspects. Medically, atrial fibrillation is mainly divided into paroxysmalAF, persistent atrial fibrillation (persistentAF), long-standing persistentAF and permanent atrial fibrillation (permanentAF) based on the duration of the atrial fibrillation episode. ). The specific classification is shown in Table 1.

Table 1.1 Detailed classification of atrial fibrillation in medicine

Atrial fibrillation is a very common arrhythmia clinically, with an incidence rate of 0.5%-1% in my country. The probability of onset increases with age. The risk of atrial fibrillation in patients with hypertension is 1.7 times higher than that in people with normal blood pressure. Currently, 33% of patients with atrial fibrillation are attributable to hypertension. In view of the high incidence of atrial fibrillation in patients with hypertension, some people even believe that atrial fibrillation is another manifestation of target organ damage in hypertension. However, there is currently no good clinical indicator to predict the occurrence of AF in patients with hypertension. In addition, some patients with atrial fibrillation have no obvious clinical symptoms, which results in these patients being unconsciously exposed to the risk of various critical illnesses. When clinical symptoms appear or a disease breaks out, it often leads to organic cardiovascular disease, thus It greatly affects the patient's health and even threatens his life. Therefore, it is particularly important to study the probability of atrial fibrillation among patients with hypertension.

At present, there are many methods to predict atrial fibrillation. In the medical field, we start with the treatment of atrial fibrillation. Although internationally there are CHA ₂ DS ₂ -VASc score (hypertension, age, diabetes, stroke, vascular disease, gender, congestive heart failure) and HATCH score (hypertension, age, cerebral ischemic attack, COPD, Heart failure) is used to predict atrial fibrillation, but both scores have various limitations, making the prediction methods irregular and the prediction results inaccurate. In the computer field, everyone generally determines whether a patient has atrial fibrillation based on the patient's electrocardiogram, judging P waves, analyzing the change pattern of RR interval distribution over time and other factors. The algorithms used include statistical algorithms. Machine learning. There are also smart watches that can be used to test some characteristic indicators of the human body for prediction, and smartphones can be used to scan faces to predict the person's face. Sometimes, even for asymptomatic patients, medical equipment can be directly used to test the patient's Holter heart rate for prediction. However, these still lack normativeness and no specific standards.

Contents of the invention

In order to solve the problem of building a decision tree to dig out indicators that affect the prediction of atrial fibrillation, the present invention proposes the following technical solutions:

A kind of atrial fibrillation prediction decision tree, the root node A peak in the decision tree, this attribute has the largest information gain rate, its normal range is 41 to 87, the first branch of the decision tree, when a<=0, a refers to represents the value of peak A. The patient has atrial fibrillation. Since there are no non-zero numbers in the data, when a=0, it is judged that the patient has atrial fibrillation; when a>0, the ef attribute needs to be continued to be considered. When When the ef value is less than 58, the patient is judged to be normal.

Another atrial fibrillation prediction decision tree. The root node in the decision tree is XGN. When the XGN level is greater than 1, the patient is judged to have atrial fibrillation. When the XGN level is less than or equal to 1, continue to consider peak A. When peak A is 0, continue to consider FS. When FS is greater than 0, the patient is judged to have atrial fibrillation. Otherwise, continue to consider FJB. When FJB is less than or equal to 0, consider LVPWD. When LVPWD is less than or equal to 9, continue to consider the value of EF. When EF is less than or equal to 57, then The patient is judged to be normal, otherwise it is atrial fibrillation; continue to trace back to the right branch of LVPWD. When LVPWD is greater than 9, the value of FDMB1 is considered. When the value is less than or equal to 101, the patient is judged to be atrial fibrillation, otherwise LAD is considered. When LAD is less than or equal to 50, the patient is judged to have atrial fibrillation, otherwise the patient is judged to be normal; continue to trace back to the right branch of FJB. When FJB is greater than 0, consider GXB. When GXB is less than or equal to 2, the patient is judged to be normal. Otherwise, the patient is judged to have atrial fibrillation; continue to trace back to the right branch of FS, when FS is greater than 0, the patient is judged to have atrial fibrillation; continue to trace back to the right branch of peak A, when A is greater than 0, consider TNB, when TNB is less than or equal to 0 When , the patient is judged to be normal, otherwise FDMB is considered. When FDMB is greater than 0, the patient is judged to be normal. Otherwise, the E value is considered. When E is greater than 72, the patient is judged to have atrial fibrillation. Otherwise, the MCHC value is considered. When MCHC is less than or equal to 338 , then the patient is judged to have atrial fibrillation, otherwise it is normal, and the entire decision tree is traversed.

A method for pruning an atrial fibrillation prediction decision tree, including:

1) Calculate the number of three types of predicted misclassified samples respectively: Calculate the sum of the number of predicted misclassified samples of all leaf nodes of the subtree Tv, recorded as E1; Calculate the number of predicted misclassified samples when the subtree Tv is pruned and replaced by leaf nodes. , recorded as E2; calculate the maximum number of branch prediction misclassification samples of subtree Tv, recorded as E3;

2) Compare: when E1 is the smallest, no pruning is performed; when E2 is the smallest, pruning is performed and a leaf node is used to replace the subtree Tv; when E3 is the smallest, the largest branch is used to replace the subtree Tv.

Beneficial effects: The present invention provides a predictive decision tree, and provides that the decision tree has important guiding significance for determining some important reference indicators that affect atrial fibrillation. The pruning method is suitable for the decision tree, and the construction efficiency and accuracy are improved Better.

Description of drawings

Figure 1 is a schematic diagram of the decision tree structure;

Figure 2 is a schematic diagram of the medical data manuscript;

Figure 3 is a schematic diagram of exporting to Excel table;

The picture is a schematic diagram of 4 cardiac ultrasound properties;

Figure 5 is a schematic diagram of the 4weka operation interface;

Figure 6 is a schematic diagram of the decision tree using default values;

Figure 7 is a schematic diagram of the accuracy of the decision tree;

Figure 8 is a schematic diagram of the decision tree of 154 factors;

Figure 9 is a schematic diagram of the decision tree accuracy.

Detailed ways

Example 1:

In order to solve the problem of building a decision tree for atrial fibrillation prediction, the present invention proposes the following technical solution: a method for building a decision tree for atrial fibrillation prediction, including:

Step 1: If the data set S belongs to the same category, create a leaf node, mark the corresponding class label, and stop building the tree; otherwise, proceed to step 2;

Step 2: Calculate the information gain rate Gain-rate(A) of all attributes in the data set S;

Step 3: Select attribute A with the maximum information gain rate;

Step 4: Establish attribute A as the root node of decision tree T, where T is the decision tree to be constructed;

Step 5: Divide the data set into multiple subsets according to different values of attribute A, perform steps 1-4 cyclically on the subset Sv, and construct a subtree Tv. Sv is a sample subset with the value v of attribute A;

Step 6: Add subtree Tv to the corresponding branch of decision tree T;

Step 7: The loop ends and the decision tree T is obtained.

Further, the data processing method is: if the class label is missing, delete the information directly; if the attribute value is missing, merge the value into the most common category or replace it with the most commonly used value; when processing continuous values, we must first Sort multiple data, divide the data set using each data as a threshold, calculate the information gain of each division, select the threshold based on the maximum gain, and use the threshold to divide the data set.

Further, perform pruning operation on the decision tree:

1) Calculate the number of three types of predicted misclassified samples respectively: Calculate the sum of the number of predicted misclassified samples of all leaf nodes of the subtree Tv, recorded as E1; Calculate the number of predicted misclassified samples when the subtree Tv is pruned and replaced by leaf nodes. , recorded as E2; calculate the maximum number of branch prediction misclassification samples of subtree Tv, recorded as E3;

2) Compare: when E1 is the smallest, no pruning is performed; when E2 is the smallest, pruning is performed and a leaf node is used to replace the subtree Tv; when E3 is the smallest, the largest branch is used to replace the subtree Tv.

Further, the splitting attribute is selected based on the information gain rate:

The formula of information entropy is:

Info_Gain(A)＝H(S)-H(A)

Where S represents the data set, c _i represents the i-th category of the data set, and p(c _i ) represents the probability that category c _i is selected;

When dividing the decision tree, what is generally calculated is the information entropy of a certain feature attribute. Assuming that the feature attribute A has n different values, the feature attribute A divides the data set S into n small data sets, represented by s _i , The probability of each small data set being selected is p(s _i ). According to formula (1), the information entropy of each small data set s _i is H(s _i ). The information entropy calculation formula of feature attribute A is:

The information gain calculation formula is:

Info_Gain(A)＝H(S)-H(A) (3)

The information gain rate calculation formula is:

Furthermore, by changing the parameters of the decision tree algorithm, the constructed decision tree is continuously adjusted, so that the accuracy and branch attribute values of the constructed decision tree are optimized: there are 11 modifiable parameters of the J48 algorithm, including binarySplits, debug, and saveInstance. , subtreeRaising, unpruned, and useLaplace all use default values. Modify and verify the five parameters ConfidenceFactor, minNumObj, numFolds, seed, and ReduceErrorPruning to continuously approach the accurate values of medical data; put the data files after data processing into the weka software , select the algorithm and modify the corresponding parameters of the algorithm, run the results, conduct experiments on possible values of various parameters, and finally select the optimal experimental results;

The experiment is divided into two branches:

The first branch of the experiment conducts experiments on several attributes of cardiac ultrasound indicators. The last column is the class label, f is atrial fibrillation, z is normal, and all parameters of the algorithm use default values; according to the decision tree, it can be seen that in the cardiac ultrasound Among the attributes, the three attributes that have the greatest impact on atrial fibrillation are A peak, ef and lasd. Specifically, as for the root node A peak in the decision tree, this attribute has the largest information gain rate. Its normal range is 41 to 87. The decision tree's The first branch, when a<=0, a refers to the value of peak A, and the patient develops atrial fibrillation. Since there are no non-zero numbers in the data, when a=0, it is judged that the patient develops atrial fibrillation. ;When a>0, you need to continue to consider the ef attribute. When the ef value is less than 58, the patient is judged to be normal;

The second branch of the experiment collects the patient's characteristic indicators, including blood routine, thyroid function, coagulation image, liver function, blood lipids, and cardiac ultrasound index detection items as attribute columns. The last column is the class label, f is atrial fibrillation, and z is normal. , all parameters of the algorithm use default values. According to the decision tree, XGN (cardiac function grade), A peak (cardiac ultrasound index), FS (rheumatic valvular heart disease), FJB (interstitial Lung disease), LVPWD (cardiac ultrasound index), EF (cardiac ultrasound index), FDMB1 (pulmonic valve blood flow velocity), FDMB (pulmonic valve), LAD (cardiac ultrasound index), GXB (coronary heart disease), TNB (diabetes) , MCHC (hemoglobin concentration), E peak (cardiac ultrasound index), specifically in the decision tree, the root node is , when A peak is 0, continue to consider FS. When FS is greater than 0, it is judged that the patient has atrial fibrillation. Otherwise, continue to consider FJB. When FJB is less than or equal to 0, consider LVPWD. When LVPWD is less than or equal to 9, continue to consider EF. value (that is, EF1 in the decision tree). When EF is less than or equal to 57, the patient is judged to be normal, otherwise it is atrial fibrillation; continue to trace back to the right branch of LVPWD. When LVPWD is greater than 9, consider the value of FDMB1. When the When the value is less than or equal to 101, the patient is judged to have atrial fibrillation, otherwise LAD is considered. When LAD is less than or equal to 50, the patient is judged to have atrial fibrillation, otherwise the patient is judged to be normal; continue to trace back to the right branch of FJB, when FJB is greater than When 0, consider GXB. When GXB is less than or equal to 2, the patient is judged to be normal. Otherwise, the patient is judged to have atrial fibrillation. Continue to trace back to the right branch of FS. When FS is greater than 0, the patient is judged to have atrial fibrillation. Continue to trace back to peak A. The right branch of , then the patient is judged to have atrial fibrillation, otherwise the MCHC value is considered. When MCHC is less than or equal to 338, the patient is judged to have atrial fibrillation, otherwise it is normal, and the entire decision tree is traversed.

Example 2:

This disclosure uses data mining methods to establish a standardized decision tree model for medical reference.

Explanation of standard terms involved:

Data Mining (DM) is to extract valuable patterns, connections and knowledge that humans have already recognized in the face of massive data from multiple sources and accumulated over a long period of time. It is to mine data and discover knowledge without any prior assumptions. Data mining is a technology that analyzes each data to find patterns in a large amount of data. It mainly consists of three steps: data preparation, pattern search and pattern representation. Data mining tasks include correlation analysis, cluster analysis, classification analysis, anomaly analysis, specific group analysis and evolution analysis, etc. What the present invention does is classification analysis to analyze whether patients with hypertension suffer from atrial fibrillation.

The decision tree algorithm is a typical algorithm used for classification prediction in the field of data mining. It has low computational complexity and intuitive output results. The present invention introduces a decision tree algorithm into predicting the probability of atrial fibrillation in patients with hypertension.

Decision tree is a basic classification and regression method. This invention mainly uses classification decision tree. The decision tree model has a tree structure, and in classification problems, represents the process of classifying instances based on features. Compared with Naive Bayesian classification, the advantage of decision trees is that they do not require any domain knowledge or parameter settings during the construction process. Therefore, in practical applications, decision trees are more suitable for exploratory knowledge discovery. Decision tree algorithms include ID3 algorithm, C4.5 algorithm and CART algorithm. This invention uses the C4.5 algorithm to conduct experiments. C4.5 is mainly improved on the basis of ID3. In the ID3 algorithm, information gain tends to select attributes with more values when selecting attributes. In order to solve this problem, the information gain rate is used instead of information gain in the C4.5 algorithm. A decision tree is a tree structure consisting of a root node, a series of internal nodes and leaf nodes. Each node has only one parent node and two or more child nodes, and the nodes are connected through branches. Each internal node of the decision tree corresponds to a non-category attribute or combination of attributes, each edge corresponds to each possible value of the attribute, and each leaf node corresponds to a categorical attribute value. An example of a decision tree structure is shown in Figure 1.

In view of the well-known meaning of the above decision tree, the present invention classifies indicators suitable for atrial fibrillation prediction. The method is as follows:

C4.5 algorithm flow

Step 1: If the data set S belongs to the same category, create a leaf node, mark the corresponding class label, and stop building the tree; otherwise, proceed to step 2;

Step 2: Calculate the information gain rate Gain-rate(A) of all attributes in the data set S;

Step 3: Select attribute A with the maximum information gain rate;

Step 4: Establish attribute A as the root node of decision tree T, where T is the decision tree to be constructed;

Step 5: Divide the data set into multiple subsets according to different values of attribute A, perform steps 1-4 cyclically on the subset Sv, and construct a subtree Tv. Sv is a sample subset with the value v of attribute A;

Step 6: Add subtree Tv to the corresponding branch of decision tree T;

Step 7: The loop ends and the decision tree T is obtained.

Numerical processing: Ability to process training data with missing attribute values. For missing class labels, delete the information directly; for missing attribute values, merge these values into one of the most common categories or replace them with the most commonly used values. Can handle continuous valued attributes. To process continuous values, we must first sort multiple data, divide the data set using each data as a threshold, calculate the information gain of each division, select the threshold based on the maximum gain, and use the threshold to divide the data set.

Pruning: Through the above decision tree generation process, we can build a decision tree based on the training data set, but the accuracy of the decision tree and other performances require us to evaluate the work that should be done on this tree. Because the decision tree we obtained is purely based on the training data set, there will be some overfitting problems. In order to solve this problem, we need to prune the decision tree. The basic idea of decision tree pruning is to remove some trees (subtrees) that are not helpful for the classification accuracy of unknown test samples. There are two improved recursive branching methods to generate simple and easier-to-understand trees: pre-pruning and post-pruning. .

Pre-pruning: Making decisions before branches prevents the data set from generating too many branches. Pruning is performed while constructing the decision tree.

Post-pruning: Mainly to reduce the impact of noise and prune excess branches.

Because the J48 algorithm used in this invention is post-pruning, the post-pruning method is introduced in detail here. Post-pruning methods include: REP (Reduced Error Pruning), PEP (Pessimistic Error Pruning), MEP (MinimumError Pruning), CCP (Cost-Complexity Pruning), etc. The default pruning method of the C4.5 algorithm is the REP pruning method. The basic idea is:

1) Calculate the number of three types of predicted misclassified samples respectively: Calculate the sum of the number of predicted misclassified samples of all leaf nodes of the subtree Tv, recorded as E1; Calculate the number of predicted misclassified samples when the subtree Tv is pruned and replaced by leaf nodes. , recorded as E2; calculate the maximum number of branch prediction misclassification samples of subtree Tv, recorded as E3.

2) Compare. When E1 is the smallest, no pruning is performed; when E2 is the smallest, pruning is performed, replacing the subtree Tv with a leaf node; when E3 is the smallest, the "grafting" strategy is adopted, that is, the largest branch is used to replace the subtree Tv.

Split attribute selection: The criterion for split attribute selection is the fundamental difference between decision tree algorithms. As mentioned above, ID3 selects split attributes through information gain, and C4.5 selects split attributes through information gain rate. Information entropy is the expected value of information. For a data set, information entropy expresses the degree of disorder of the data set. The more categories the data set contains, the greater the corresponding information entropy. The formula is:

Where S represents the data set, c _i represents the i-th category of the data set, and p(c _i ) represents the probability that category c _i is selected;

When dividing the decision tree, what is generally calculated is the information entropy of a certain feature attribute. Assuming that the feature attribute A has n different values, the feature attribute A divides the data set S into n small data sets, represented by s _i , The probability of each small data set being selected is p(s _i ). According to formula (1), the information entropy of each small data set s _i is H(s _i ). The information entropy calculation formula of feature attribute A is:

The information gain calculation formula is:

Info_Gain(A)＝H(S)-H(A) (3)

The information gain rate calculation formula is:

Algorithm application

Data description: The data used in this invention are provided by a hospital in Dalian City and generated by actual measurements of hypertensive patients, with a total of 360 copies. The laboratory report mainly includes white blood cell count (WBC), absolute granulocyte count (Neu#), NT-proBNP, EF (ejection fraction), LVEF (left ventricular ejection fraction), hypertension grade, whether you have atrial fibrillation, etc. . Figure 2 shows some of the original data items.

Data preprocessing: The data file type running on the Weka platform is .csv file, and our data file is Excel table data, so the first step is to convert the data file into a .csv file. Other indicators not considered by the present invention are filtered out from the data provided by the hospital, leaving only the research objects. Abnormal data is deleted, and the J48 algorithm for blank value attributes is automatically processed. Since each value of the 154-dimensional data has a large order of magnitude, the present invention will use relevant medical standards to targetedly extract the 11-dimensional data, that is, cardiac ultrasound indicators, to conduct more specific experiments. Such as ef (ejection fraction), a peak, e peak, etc. Simplified as shown in Figure 3.

Operating environment: Waikato Environment for KnowledgeAnalysis (WEKA) is a free, non-commercial open source machine learning and data mining software based on JAVA environment. The main developer is from New Zealand. The official website is: http://weka.wikispaces.com/. As an open data mining working platform, WEKA integrates a large number of machine learning algorithms that can undertake data mining tasks, including data preprocessing, correlation analysis, classification, and regression. , clustering and visualization on the new interactive interface, embed WEKA into MyEclipse to facilitate secondary development of WEKA; modify or add some of the latest data mining algorithms, and display the mining results in various forms , enabling users to find the required knowledge conveniently and clearly. Before mining, you need to configure JDBC and load the database driver. The Weka control platform and operation interface are shown in Figure 4. If you use the weka software, you need to open the control platform first, select the first option Explorer to start the experiment, the opened interface is as shown in the operation interface diagram, the first step is to select the experiment you want to conduct through the openfile option, and then according to your own Experimental requirements include conducting different experiments on data, such as data preprocessing, classification algorithms, clustering algorithms, association rules and other optional parameters. According to the experimental requirements, the present invention selects the J48 algorithm among the classification algorithms for experiments. The software operation interface is shown in Figure 5.

Decision tree construction: The construction of decision tree is not unique. Unfortunately, the construction of optimal decision tree is an NP problem. Therefore, how to construct a good decision tree is the focus of research. The present invention continuously adjusts the constructed decision tree by changing the parameters of the decision tree algorithm, so that the accuracy and branch attribute values of the constructed decision tree are optimized. There are 11 modifiable parameters of the J48 algorithm. Among them, binarySplits, debug, saveInstance, subtreeRaising, unpruned, and useLaplace all use default values. Five parameters, ConfidenceFactor, minNumObj, numFolds, seed, and ReduceErrorPruning, are modified. The experiments of this invention mainly modify and verify the remaining six parameters to continuously approach the accurate values of the medical data, making the decision tree more accurate and feasible. The weka software is similar to a black box. As long as you put the processed data files into weka, select the algorithm you want and modify the corresponding parameters of the algorithm, you can run the results. Experiments were conducted on the possible values of various parameters, and the optimal experimental results were finally selected as follows. The experiment is generally divided into two branches. One part is to experiment on 11 attributes of cardiac ultrasound. The last column is the class label, f is atrial fibrillation, and z is normal. The experimental data contains a total of 360 individuals, including 186 males and 174 females. There were 178 people with atrial fibrillation and 182 people with normal conditions (normal people here refer to patients with pure hypertension). All parameters of the algorithm use default values, and the experimental results are shown in Figure 6.

From the above decision tree, we can know that among the attributes of cardiac ultrasound, the three attributes that have the greatest impact on atrial fibrillation are peak A, ef, and lasd. Specific to the root node A peak in the decision tree (when it disappears, it means that atrial fibrillation has occurred.) This attribute has the largest information gain rate, and its normal range is 41 to 87. In the first branch, we can see that when a <= 0, the patient develops atrial fibrillation. Since there are no non-zero numbers in the data, when a = 0, it can be determined that the patient develops atrial fibrillation. When a>0, the ef attribute needs to be continued to be considered. When the ef value is less than 58, the patient is judged to be normal. Analyze the decision tree by analogy. Decision tree accuracy screenshots include accuracy rate, error rate, Kappa value, etc. These factors can be used to evaluate the quality of the algorithm. This invention mainly bases the judgment on the accuracy rate. It can be seen from Figure 7 that the accuracy is 83.0556%.

The second part of the experimental data contains a total of 308 data. 154 patient characteristic indicators, including blood routine, thyroid function, coagulation image, liver function, blood lipids, cardiac ultrasound and other indicator detection items are used as attribute columns. The last column is the class label, f is atrial fibrillation, z is normal. There are 162 males and 146 females in the data. There were 128 patients with atrial fibrillation and 180 normal patients. Similar to the above, all parameters of the algorithm use default values. The experimental results are shown in Figure 8.

Through the decision tree, we can see that among the 154 attributes that play a role in atrial fibrillation, they include ), LVPWD (cardiac ultrasound index), EF (cardiac ultrasound index), FDMB1 (pulmonic valve blood flow velocity), FDMB (pulmonic valve), LAD (cardiac ultrasound index), GXB (coronary heart disease), TNB (diabetes), MCHC (hemoglobin concentration), E peak (cardiac ultrasound index). Some of these 13 indicators have not attracted enough attention in medicine. For example, the effect of hemoglobin concentration on atrial fibrillation.

Specific to the decision tree, the root node is When FS is greater than 0, it is judged that the patient has atrial fibrillation, otherwise FJB is continued to be considered. When FJB is less than or equal to 0, LVPWD is considered. When LVPWD is less than or equal to 9, the value of EF (that is, EF1 in the decision tree) is continued to be considered. When EF is less than When it is equal to 57, the patient is judged to be normal, otherwise it is atrial fibrillation; continue to trace back to the right branch of LVPWD. When LVPWD is greater than 9, the value of FDMB1 is considered. When the value is less than or equal to 101, the patient is judged to have atrial fibrillation. , otherwise LAD is considered. When LAD is less than or equal to 50, the patient is judged to have atrial fibrillation, otherwise, the patient is judged to be normal; continue to trace back to the right branch of FJB, when FJB is greater than 0, consider GXB, when GXB is less than or equal to 2, then The patient is judged to be normal, otherwise the patient is judged to have atrial fibrillation; continue to trace back to the right branch of FS, when FS is greater than 0, the patient is judged to have atrial fibrillation; continue to trace back to the right branch of peak A, when A is greater than 0, consider TNB, When TNB is less than or equal to 0, the patient is judged to be normal, otherwise FDMB is considered. When FDMB is greater than 0, the patient is judged to be normal, otherwise the E value is considered. When E is greater than 72, the patient is judged to have atrial fibrillation, otherwise the MCHC value is considered. When MCHC is less than or equal to 338, the patient is judged to have atrial fibrillation, otherwise it is normal, and so on, traversing the entire decision tree. The accuracy of this model is 85.0649%.

Through the above different experiments, taking the decision tree and accuracy into consideration, the present invention will select Figure 8 as the final model. This model considers many factors and is relatively comprehensive. For medical workers, it is more concise and generous. The model also has medical approval.

Aiming at the problem in the medical community that there is no uniformly standardized model for predicting atrial fibrillation and that patients with hypertension have a higher probability of suffering from atrial fibrillation than ordinary people, the present invention refers to a medical review of atrial fibrillation prediction and proposes a decision-tree-based model for predicting atrial fibrillation. tremor prediction method to solve this problem. Through this method, an intuitive and concise decision tree is established for medical research reference. This model combines a large amount of real medical data to ensure the accuracy of the model as comprehensively as possible. The accuracy of the model is 85.0649%. During the model building process, not only can the potential relationships between various medical indicators in patients with hypertension be unearthed, but also which indicators are more likely to cause atrial fibrillation, as well as some indicators that were not paid in-depth attention in medicine at the beginning. In the next work, the first point will be to increase the amount of data to make the model more generalizable and prevent over-fitting. The second point is to use machine learning algorithms to make larger and better classifications and establish a practical and standardized decision tree.

Example 3:

In order to solve the problem of selecting indicators that more accurately reflect atrial fibrillation, the present invention constructs a method for selecting indicators for atrial fibrillation artificial intelligence experiments:

S1. Build a decision tree;

S2. Adjust parameters to optimize the decision tree;

S3. Experiment on the possible values of various parameters, and finally select the optimal experimental result, which is used as the main indicator for decision tree prediction.

Further, the main indicators are the three attributes of cardiac ultrasound: A peak, ef and lasd.

Further, the main indicators are XGN (cardiac function grade), A peak (cardiac ultrasound indicator), FS (rheumatic valvular heart disease), FJB (interstitial lung disease), LVPWD (cardiac ultrasound indicator), EF (cardiac ultrasound index), FDMB1 (pulmonic valve blood flow velocity), FDMB (pulmonic valve), LAD (cardiac ultrasound index), GXB (coronary heart disease), TNB (diabetes), MCHC (hemoglobin concentration), E peak (cardiac ultrasound index).

The method of constructing a decision tree is as described in Examples 1 and 2.

The invention also relates to the application of a prediction decision tree in predicting atrial fibrillation.

The present invention uses artificial intelligence and big data processing to make a more reasonable selection of atrial fibrillation prediction indicators. The indicators are obtained through big data processing and can more accurately reflect atrial fibrillation. These indicators are used to evaluate the ability of atrial fibrillation. To reduce the missed detection of atrial fibrillation, the present invention also provides a method for constructing a decision tree that can be used for atrial fibrillation prediction, and completely explains the establishment process of the decision tree, so that a standard decision tree can be established in the field of atrial fibrillation prediction. A decision tree model was constructed using mathematical methods, and the decision tree was given, which has important guiding significance for determining some important reference indicators that affect atrial fibrillation.

Example 4:

Atrial fibrillation (AF, referred to as atrial fibrillation) is one of the most common clinical arrhythmias, with a prevalence of approximately 0.4%-1.0% in the overall population, and the prevalence increases with age. , research shows that the prevalence rate in people <55 years old is only 0.1%, while the prevalence rate in people >80 years old is as high as 9%. A common clinical complication of atrial fibrillation is systemic thromboembolism. Stroke is the main embolic event caused by atrial fibrillation. It is also the complication with the highest disability rate in patients with atrial fibrillation. Compared with patients without atrial fibrillation, the incidence of stroke is less The mortality rate increases by 5 times, and the mortality rate increases by 2 times. Ischemic stroke is the main reason for the increase in mortality, and atrial fibrillation is an independent risk factor for ischemic stroke, and its incidence rate increases with age. Other hazards of atrial fibrillation include: loss of atrial auxiliary pump function leading to heart failure, electrical disorders leading to sudden death, irregular and fast ventricular rates leading to physical and mental disorders.

Accurately predicting the occurrence of atrial fibrillation and applying effective preventive measures is an important part of the treatment process of atrial fibrillation. At present, the diagnosis of atrial fibrillation is mainly based on electrocardiogram and electrocardiogram extensions such as dynamic electrocardiogram, monitored electrocardiogram and implanted long-term electrocardiogram. In recent years, great achievements have been made in integrating electrocardiogram technology and artificial intelligence. However, the accuracy of diagnosing atrial fibrillation based on the traditional electrocardiogram technology of more than 100 years is high, but the missed diagnosis rate is also high, especially for paroxysmal atrial fibrillation and infrequent attacks. Symptomatic atrial fibrillation is no less harmful than symptomatic atrial fibrillation. This technology is based on clinical big data combined with artificial intelligence (AI) to develop a new atrial fibrillation diagnosis system, with a view to replacing the traditional electrocardiogram diagnostic technology, at least as a screening diagnostic system for high-risk patients with atrial fibrillation before electrocardiogram examination, and as an important supplement to the classic electrocardiogram examination .

Methods and techniques: This study uses the information integration platform of the applicant's affiliated hospital - Zhongshan Hospital Affiliated to Dalian University to analyze all clinical, imaging, examination and other data of patients with hypertension, using big data processing methods, as described in Example 3 Use the decision tree method described above to create an automatic intelligent diagnosis model, such as a decision tree model. On this basis, use this model to diagnose and analyze patients with hypertension combined with atrial fibrillation, and further use the AI system deep learning method to further analyze the model. Revise and continuously improve, and finally develop a complete artificial intelligence diagnosis system for atrial fibrillation. This invention closely combines clinical big data with AI, and through big data processing and AI self-learning, it will surely open up a new breakthrough for predicting the occurrence of AF and provide an important diagnostic means for atrial fibrillation prevention and treatment strategies.

AI model production: Using the information integration platform of the applicant's affiliated hospital - Zhongshan Hospital Affiliated to Dalian University, the clinical data (medical history, physical examination, physical and chemical examinations) of hypertensive patients registered in our hospital from January 2010 to December 2017 etc.) perform big data processing and establish a primary diagnostic model.

AI model verification: Use the primary AI model to input the relevant parameter data of patients diagnosed with hypertension in our hospital into the computer to test the diagnostic ability of the AI model (including prediction sensitivity, specificity, compliance rate and prediction efficiency)

AI model improvement: Through the ability of AI self-deep learning, the model is constantly revised and improved, and gradually developed and improved.

The above are only preferred embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person familiar with the technical field can, within the technical scope disclosed by the present invention, proceed according to the present invention. Any equivalent replacement or modification of the created technical solution and its inventive concept shall be covered by the protection scope of the invention.

The above are only preferred embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person familiar with the technical field can, within the technical scope disclosed by the present invention, proceed according to the present invention. Any equivalent replacement or modification of the created technical solution and its inventive concept shall be covered by the protection scope of the invention.

Claims (2)

1. A judgment method based on atrial fibrillation prediction decision tree, which is characterized by: The method for constructing a decision tree for predicting atrial fibrillation includes: Step 1: If the data set S belongs to the same category, create a leaf node, mark the corresponding class label, and stop building the tree; otherwise, proceed to step 2; Step 2: Calculate the information gain rate Gain-rate(A) of all attributes in the data set S; Step 3: Select attribute A with the maximum information gain rate; Step 4: Establish attribute A as the root node of decision tree T, where T is the decision tree to be constructed; Step 5: Divide the data set into multiple subsets according to different values of attribute A, perform steps 1-4 cyclically on the subset Sv, and construct a subtree Tv. Sv is a sample subset with the value v of attribute A; Step 6: Add subtree Tv to the corresponding branch of decision tree T; Step 7: The loop ends and the decision tree T is obtained; The atrial fibrillation prediction decision tree is a computer processing decision-making model. Its processing and decision-making process specifically includes: If the root node A in the decision tree has a peak, this attribute has the largest information gain rate, and its normal range is 41 to 87. The decision tree The first branch of Tremor; when a>0, the ef attribute needs to be continued to be considered. When the ef value is less than 58, the patient is judged to be normal; If the root node in the decision tree is XGN, when the XGN level is greater than 1, the patient is judged to have atrial fibrillation. When the XGN level is less than or equal to 1, continue to consider peak A. When peak A is 0, continue to consider FS. When FS is greater than 0, It is judged that the patient has atrial fibrillation, otherwise continue to consider FJB. When FJB is less than or equal to 0, consider LVPWD. When LVPWD is less than or equal to 9, continue to consider the value of EF. When EF is less than or equal to 57, the patient is judged to be normal, otherwise Atrial fibrillation; continue to trace back to the right branch of LVPWD. When LVPWD is greater than 9, the value of FDMB1 is considered. When the value is less than or equal to 101, the patient is judged to have atrial fibrillation. Otherwise, LAD is considered. When LAD is less than or equal to 50, then The patient is judged to have atrial fibrillation, otherwise the patient is judged to be normal; continue to trace back to the right branch of FJB, when FJB is greater than 0, consider GXB, when GXB is less than or equal to 2, the patient is judged to be normal, otherwise the patient is judged to be atrial fibrillation; continue Backtrack to the right branch of FS. When FS is greater than 0, the patient is judged to have atrial fibrillation. Continue to trace back to the right branch of peak A. When A is greater than 0, TNB is considered. When TNB is less than or equal to 0, the patient is judged to be normal. Otherwise, consider FDMB. When FDMB is greater than 0, the patient is judged to be normal. Otherwise, the E value is considered. When E is greater than 72, the patient is judged to have atrial fibrillation. Otherwise, the MCHC value is considered. When MCHC is less than or equal to 338, the patient is judged to have atrial fibrillation. Otherwise, it is normal and the entire decision tree is traversed; Among them, the splitting attribute is selected according to the information gain rate: The formula of information entropy is: Info_Gain(A)＝H(S)-H(A) Where S represents the data set, c _i represents the i-th category of the data set, and p(c _i ) represents the probability that category c _i is selected; When dividing the decision tree, what is generally calculated is the information entropy of a certain feature attribute. Assuming that the feature attribute A has n different values, the feature attribute A divides the data set S into n small data sets, represented by s _i , The probability of each small data set being selected is p(s _i ). According to formula (1), the information entropy of each small data set s _i is H(s _i ). The information entropy calculation formula of feature attribute A is: The information gain calculation formula is: Info_Gain(A)＝H(S)-H(A) (3) The information gain rate calculation formula is: 2. The judgment method based on atrial fibrillation prediction decision tree according to claim 1, characterized in that: The method of pruning the atrial fibrillation prediction decision tree includes: 1) Calculate the number of three types of predicted misclassified samples respectively: Calculate the sum of the number of predicted misclassified samples of all leaf nodes of the subtree Tv, recorded as E1; Calculate the number of predicted misclassified samples when the subtree Tv is pruned and replaced by leaf nodes. , recorded as E2; calculate the maximum number of branch prediction misclassification samples of subtree Tv, recorded as E3; 2) Compare: when E1 is the smallest, no pruning is performed; when E2 is the smallest, pruning is performed and a leaf node is used to replace the subtree Tv; when E3 is the smallest, the largest branch is used to replace the subtree Tv.