TWI678709B - Disease prediction method through a big database formed by data mining of neural network - Google Patents

Abstract

A method for assisting neural network from data exploration to building a big data database and then performing disease prediction includes the following steps. First, input disease-related data for data conversion, extract features from disease data, and perform data Encoding to obtain the data to be trained, then calculate the similarity between the features using the Euclidean distance, and group the data to obtain the model to be trained. Finally, enter the disease-related data without data conversion and the The training model uses Euclidean distance to perform data clustering to obtain a training model.

Description

Neural network to assist data exploration and big data Identify methods for disease prediction

The invention relates to a disease prediction method, in particular to a method for assisting neural network from data exploration to establishing a database of big data, and then performing disease prediction.

In recent years, the application of artificial intelligence (AI) in medical care has been extensively researched and discussed. More and more studies show that AI will play a key role in the future application of human health.

The accuracy of a machine learning model depends heavily on the amount of data used to train the model. Machine learning is further divided into supervised learning, unsupervised learning, and reinforcement learning. Regardless of the learning strategy adopted, the goal is to find appropriate parameters for each neuron in the network for future use.

When training machine learning models, there are technically important parts such as Data Cleaning, Feature Extraction and Feature Selection, and model selection. In order to train the machine, data samples are collected as training data ( Training Data). Features from the training data help us determine the target.

Many medical data are currently presented in text, which cannot be directly used as a feature input training model, but the text data is very important for the prediction of disease and treatment. A large amount of medical text data is manually written by one person Marking will waste human resources too much and cause an increase in the error rate. How to use neural network to help directly convert a large amount of text data into features that can be input into the training model under the need of the most economical medical human resources is the current direction of efforts.

Features can affect the results obtained. Some features are not very important. There is not enough clues to know the importance of each feature when clustering. Therefore, it is easy to bias some of the distribution characteristics and cause meaningless clustering results. If there are too many required features, it will cause difficulties in practical application. How to select real and useful influence parameters to improve the accuracy of neural network prediction is a very worthwhile goal, and how to define and compare the prediction accuracy of models is another important subject to be developed.

An object of the present invention is to provide a method for predicting diseases by using neural network to assist data exploration and data identification of big data, including the following steps.

First, input disease-related data for data conversion, extract features in the disease data, and encode the data to obtain the data to be trained. Then, use the Euclidean distance to calculate the similarity between the features and group the data. In order to obtain the desired training model, finally, input disease-related data without data conversion and the desired training model uses Euclidean distance to perform data clustering to obtain the training model.

Another technical means of the present invention is that the above-mentioned word-vector conversion is used to convert each word in the disease data to a vector. All the words converted by a word can represent a vector space, and the space between words can be calculated in this space. Vector distance.

Another technical means of the present invention is that the foregoing is to give proportional weights with different characteristics, and then perform data clustering.

Another technical means of the present invention is the above-mentioned resources. Data grouping is the classification of disease data.

Another technical means of the present invention is that the above method further validates the training model, which includes 3 sub-steps, first input the training model, then take 1/4 of the total number of cases of the training model for model training, and then Take 1/4 of the total number of cases of the training model for model verification, and finally take 1/2 of the total number of cases of the training model for final model training. The trained model then takes the total number of cases of the training model. The other 1/2 is tested to confirm the accuracy of the model verification.

Another technical means of the present invention is to perform the next step after confirming that curve fitting is possible.

Another technical means of the present invention is that the above is a case of randomly taking the training model.

Another technical means of the present invention is a method of parameter adjustment during model training to improve the accuracy and stability of the model.

Another technical means of the present invention is that the accuracy of the verification model verification is verified through a 95% confidence interval.

Another technical means of the present invention is to repeatedly and repeatedly delete feature weights of the input layer until one of the following three situations occurs: stop: summing up all input layer feature weights is greater than 95% and must include all features; delete The 95% confidence interval of the model error rate generated by the five random samplings after the feature cannot have five simultaneous overlapping ranges; the 95% confidence interval of the error rate of the five models overlapping each other does not include 0. After stopping, the last remaining The input layer features are used as the input layer for the final model training.

關係式，其中，SD(standard deviation)為 訓練模型之資料庫中所有真實數值的標準差，mean為資料庫中所有真實數值之平均值，MAPE為預測模型之平均絕 對值誤差率。 Another technical means of the present invention is to establish the benefits of the model and use it as a method of comparing the accuracy between different models to satisfy the Model

The relational expression, where SD (standard deviation) is the standard deviation of all true values in the database of the training model, mean is the average of all true values in the database, and MAPE is the average absolute value error rate of the prediction model.

The beneficial effect of the present invention is that first, supervised learning is used, then unsupervised learning is used, and neural network applications such as reinforcement learning are used to predict the disease, so as to improve the accuracy and validity of the training model.

11 ~ 14‧‧‧step

FIG. 1 is a schematic flow chart illustrating a preferred embodiment of a method for disease prediction by using neural network to assist data exploration and big data data identification in the present invention.

The features and technical contents of the related patent application of the present invention will be clearly presented in the following detailed description of the preferred embodiments with reference to the drawings.

Refer to FIG. 1, which is a preferred embodiment of a method for disease prediction by using neural network to assist data exploration and big data data identification in the present invention. Here, supervised learning is used first, then unsupervised learning is used, and then reinforcement Methods for predicting diseases using neural networks, such as learning, include the following steps.

Neural network for disease prediction medical institutions in the first stage needs to convert text report data into numbers. For example, X-rays and electrocardiograms are text reports for the use of subsequent prediction models.

First, step 11 is performed to train disease-related data through neural network to perform data conversion, extract features in the disease data, and perform data coding to obtain data to be trained.

In step 11, word vector conversion is used to convert each word in the disease data to a vector. All the converted words can represent a vector space, and the vector distance between the words can be calculated in this space. For example: Count the number of times a word appears in a file, and then calculate the probability of two words appearing together to determine their similarity The shorter the vector distance between words, the more similar.

The technique of word vector assigns each word to a vector, which is used to represent the relative characteristics of this word in various dimensions, and the vector is used to distinguish the similarity between each word and other words.

For example, there are 3,000 X-ray reports. If the content does not exceed 200 words, the fill-in setting is a total of 200 words. Each word is a vector representation. Each vector is represented by 100 dimensions. Each report is equivalent to 20,000. Features determine whether there is a fracture. Next, manual labelling is started on 3,000 X-ray reports. For example, 1 indicates that there is a fracture, and 0 indicates that there is no fracture. If the predicted feature is completely classified as 1, and none is classified as 0.

Further, in step 11, data clustering is performed according to the proportional weights with different characteristics. 20,000 features can be regarded as neurons with 20,000 input layers. For example: the vector sets 100 important signals, how much weight each signal occupies, the middle hidden layer performs calculations, and the output layer is 10 with or without fracture. 01. By adding proportional weights to the input layer, the discrimination accuracy is improved.

Basically, the space vector of each thing has different importance. When there is noise in some data, the clustering becomes very important. For example: the report writes a slight collapse, or the text arrangement is a little different, which can be overall. It seems that the text is similar, so weight is given. For example, some words only write a fragment, but the fracture may be a hip fracture or a fracture in other parts. Therefore, it is not so important to identify the spine fracture.

Next, step 12 is performed to calculate the similarity between the features based on the Euclidean distance for the data to be trained, and perform data clustering to obtain the model to be trained. The Euclidean distance calculates the semantic similarity between the two word vectors, and the data is grouped Data were classified for disease.

In the X-ray report, different proportional weights are given to specific features. The common words in fracture are collapse, compress, and fracture. Examples are given below: the system regards it as the most important inside, and the ones near him follow It is very important, because in the past, there were a lot of words in the process of judging the text. Some words were considered unimportant and no anchor points were given. First, a weight was defined to inform the importance of the anchor points. For example, no bony collapse is very important. The key is not to be ignored. The next step is supervised learning, which refers to labeling each word.

After training the features, you will find that it is divided into several blocks and distinguished whether there are fractures. It is particularly noted that the number of set dimensions depends on the number of words reported by the physician. If the doctor types different lengths, the set words are relatively different. For example, if the bone is normal, osteoporosis, and bone loss in between, the drawing is divided into 3 blocks.

Then, step 13 is performed to input disease-related data without data conversion and the model to be trained to perform data clustering by using Euclidean distance to obtain a training model.

After the system recognizes the vector, I want to train the model, but only use 3,000 copies that are not stable or not. I use unsupervised learning here, and then enter 30,000 copies of disease-related data without data conversion. In the previous step, Take out the feature vector, and use 30,000 important vectors to classify the data based on the above 3,000 copies using the European distance to automatically encode the 30,000 copies. No manual coding is required. Use vectors to calculate the same clustering features. Together, such 30,000 copies will be coded. After that, 30,000 copies will be added to the original artificial 3,000 copies. A total of 33,000 will generate a new model, which is also a model for subsequent training. This training model can be significantly Saving the time actually used by the clinician to build the Model is also the biggest focus of the present invention. To build a training model, it is necessary to make the Label correct at the beginning, improve the efficiency of the Label with unsupervised learning, and then use more models to train.

Later, using reinforcement learning, in the clinical work, it is necessary to return data to tell the system which is right or wrong. The process of user use continuously loses data and returns to the system for the model to be continuously revised. The following are methods to improve and quickly build the model. To improve accuracy.

Finally, step 14 is performed to verify the training model. Here, a disease prediction model is established and a bone density prediction model is taken as an example, which includes the following three sub-steps.

1. First, a training model is input, and the training model is a database containing detection values of a dual energy X-ray absorptiometry (DXA) and the aforementioned coding completed by the aforementioned method of this patent. 2. Next, randomly take 1/4 of the total number of cases of the training model for model training, and then randomly take 1/4 of the total number of cases of the training model for model validation. After curve fitting, proceed to the next step. 3. Finally, randomly take 1/2 of the total number of cases of the training model for the final model training, and then take the other 1/2 of the database for testing to confirm the effectiveness of the model verification. Among them, in the third step, confirm that the correctness of the model verification is verified through the 95% confidence interval. Before starting to train the model, you need to verify the correlation between the input layer factors (factors also refer to features), and you can view all the factors. The correlation between linear and non-linear, if the correlation is greater than 0.8, then the two factors are not suitable for coexisting in the input layer, and only one of them can be selected and put into the model.

In actual implementation, assuming a total of 1,000 pens, first take 250 pens (1/4) and start training for the model. Since not every database is suitable for neural network analysis, the samples of the database may be excessively extreme and selected. Variables are not stable or important factors affecting the output layer are not selected; when training the model, multiple parameter settings must be included, such as the learning rate, the number of hidden layers, and the training method of the neural network. The above conditions are suitable for neural network training. At first, random sampling is performed from the total database. For the first time, 250 strokes are drawn. 250 strokes will train his Model, and then the remaining 750 strokes (3/4) of the data will be randomly selected. 250 samples (1/4) were taken as the reliability and validity of the validation model. In order to avoid sampling bias, five samples were repeatedly trained to train five models, and each model validation had an average absolute error rate (Mean Absolute). Percentage Error (MAPE), for example, under the first Model, If a piece of data predicts a bone density of 0.9, the actual DXA test result is 1, the absolute value of the error value is 0.1, and the actual DXA test result 1 is divided by 10%. The error rate of all the data is averaged and averaged. After that is MAPE.

The reference real value is the DXA bone density check value. There is an error value when compared with the value calculated by the Model. The error value is divided by the real value as the error rate. The average error of the 250 error values can be calculated after the verification phase. The standard deviation (SD) of the error rate and error rate. The middle of the error rate is 0. The two ends are positive and negative errors. The standard deviation of the positive and negative error rates is the 95% confidence interval of the error rate. The models created by the sub-sampling will have different average error rates and standard deviations of the error rates. If the 95% confidence interval of the five model error rates has five simultaneous overlapping ranges, the model will be deemed to be under relevant conditions. (As described in paragraph 0036) is stable. This method can be used as a method to verify the reliability of the model. At the same time, the 95% confidence interval of the error rate of the five times the model repeats with each other needs to include 0 as the model validity verification.

The relevant parameters of the model can be adjusted through the change of MAPE or (the average of the error value x the standard deviation of the error value) during the setting process, such as adjusting the learning rate, comparing different learning rates, and choosing the minimum MAPE or (the average x The standard deviation of the error value) is the smallest parameter, and the adjustment method can be adjusted with the following data. First use 10 ⁿ , N = + 2, +1,0, -1, -2 as the benchmark for testing. Select The best MAPE or (the average of the error values x the standard deviation of the error values), if the best n value is the largest or smallest (such as +2 or -2), you need to add five more layers up or down Test (such as + 2, + 3, + 4, + 5, + 6), then Ax10 ⁿ , A = 4,3,2,1,0,0.9,0.8,0.7,0.6,0.5 Parameter value.

After that, I began to try to reduce the factors of the Model input layer. When generating a prediction model through neural network training, we can simultaneously give weight to the image power of the output layer results according to the factors, and all input layer factor weights. The sum is 100%, which indicates the importance ratio of each input layer factor to the output layer in the model. After five random sampling training models are generated, the weights of each factor are averaged five times. The next step starts with The highest weighted factors are summed up to the lowest weighted factors. Due to the concepts of 95% confidence interval and α error 5% commonly used in medicine, when adding the factor weights, the total weight is greater than 95% and stops. All remaining factors are deleted. After the deletion, repeat the second step and the third step, and confirm that the 95% confidence interval of the error rate of the model five times after changing the input layer factors still has five simultaneous overlapping ranges to confirm the model after deleting the factors. If the model confirms that it is suitable, this step will be repeated, new weights will be calculated, and the input layer factors will be deleted and reduced.

Repeat the steps of deleting factor items according to the input layer factor weights until it is stopped in one of the following three cases: 1. Add up all input layer factor weights greater than 95% and must include all factors to achieve this requirement; 2. After removing the factor, the 95% confidence interval of the model error rate generated by five random samplings cannot have five simultaneous overlapping ranges; the 95% confidence interval of the error rate repeated by the three or five models does not include 0; after stopping, the last The remaining input layer factors serve as the input layer for the final prediction model.

Finally, the final model is established based on the input layer factor items and all relevant parameter settings. Under this condition, 500 (1/2) of all the total databases are randomly selected for model training, and the remaining 500 Test with pen (1/2).

The 95% confidence interval of the error rate of the resulting model during testing needs to include 0 to indicate that the model has sufficient validity. If it contains 0, it means that there is an absolute error directly between the prediction model and the actual measurement value. If the 95% confidence interval in medicine does not include 0, it means that there is a significant difference, indicating that the prediction model and the actual measurement are two independent events that are not related. Situation prediction models are not suitable for clinical use.

，其中，SD(standard deviation)為 資料庫中所有真實數值(DXA檢測數據)之標準差，mean為資料庫中所有真實數值之平均值，MAPE為預測模型之平均絕對值誤差率，其數值越大越好，亦代表Model的預測效果越良好，此公式表示透過預測模型將可較直接依照資料庫平均值盲猜的效果優秀多少，如某一群體DXA檢測數值之平均值為1，標準差為0.2，我們可得知若個案未做檢查，我們直接猜他的骨密度為1，大部分的個案骨密度也差不多在1±0.2之間，因此以模型之MAPE與此盲猜的數值做比較，可以得知預測模型的效果較盲猜良好的程度，而不同的模型也可透過此公式比較優劣度。 The validity of the model can be expressed as the MAPE value. For example, if the MAPE is 10%, it means that most of the results calculated by the model and the actual measured values will fall within ± 10%. The benefits of the model are used as the difference between different models. The prediction effect comparison can be compared through the model efficacy formula.

Among them, SD (standard deviation) is the standard deviation of all true values (DXA test data) in the database, mean is the average of all true values in the database, and MAPE is the average absolute value error rate of the prediction model. The bigger the better, the better the model's prediction effect. This formula indicates that the prediction model will be better than the direct guess of the average value of the database. For example, the average value of the DXA detection value of a group is 1, and the standard deviation is 0.2, we can know that if the case is not checked, we directly guess that his bone density is 1, and most cases have a bone density of about 1 ± 0.2, so we compare the MAPE of the model with the value of this blind guess It can be known that the prediction model performs better than blind guessing, and different models can compare the pros and cons through this formula.

In summary, the present invention uses a neural network to assist data exploration and big data data identification for disease prediction. The method converts data by inputting disease-related data, extracts features in disease data, and encodes data to obtain The data to be trained is used to calculate the similarity between the features based on the Euclidean distance, and the data is grouped to obtain the model to be trained. Finally, the disease-related data without data conversion is input and the model to be trained uses Euclidean Data are grouped by distance to obtain a training model. Supervised learning is used first, then unsupervised learning is used, and neural network applications such as reinforcement learning are used to predict the disease in order to improve The accuracy and validity of the training model, and the method of parameter adjustment during model training can improve the accuracy and stability of the model, so it can indeed achieve the purpose of cost invention.

However, the above are only the preferred embodiments of the present invention. When the scope of implementation of the present invention cannot be limited by this, that is, the simple equivalent changes and modifications made according to the scope of the patent application and the description of the invention, All are still within the scope of the invention patent.

Claims (6)

A method for predicting disease by using neural network to assist data exploration and data identification of big data, including the following steps: (A) input disease-related data for data conversion, extract features in disease data, and perform data coding to obtain To-be-trained data, in which word vector conversion is used to convert each word in the disease data to a vector. All words converted can represent a vector space, in which the vector distance between words can be calculated, and Give different proportion weights to the features, and then perform data clustering; (B) Calculate the similarity between the features using the Euclidean distance between the data to be trained, and perform data clustering to obtain the model to be trained, and the data clustering is the data to disease Grading; (C) input disease-related data without data conversion and use the Euclidean distance to cluster the data to obtain a training model; and (D) verify the training model, which includes the following sub-steps: (d1 ) Input the training model; (d2) Take 1/4 of the total number of cases of the training model for model training, and then take the training 1/4 of the total number of cases of the model is used for model verification; and (d3) 1/2 of the total number of cases of the trained model is used for the final model training, and the trained model is then taken from the total number of additional cases of the trained model. Test to confirm the accuracy of the model verification. According to the method described in item 1 of the scope of the patent application, a method for predicting disease by using neural network to assist data exploration and big data data identification is performed. In this step (d2), after confirming that curve fitting can be performed, Go to the next step. The method for predicting disease by using neural network to assist data exploration and big data data identification according to item 2 of the scope of patent application, wherein in this step (d2) and step (d3), the training model is randomly taken Case. The method for predicting disease by using neural network to assist data exploration and big data data identification according to item 3 of the scope of patent application, wherein in this step (d3), confirm that the correct rate of model verification is 95% trusted Interval for verification. The method for predicting diseases by using neural network to assist data exploration and big data data identification as described in item 4 of the scope of patent application, wherein in this step (d2), the feature weight of the input layer is repeatedly deleted until it appears. One of the following three situations is stopped: the total weight of all input layer features is greater than 95% and all features must be included; the 95% confidence interval of the model error rate generated by five random samplings after deleting the features cannot have five simultaneous intervals The overlapping range; the 95% confidence interval of the error rate of the five models repeating each other does not contain 0. After stopping, the last input layer feature is used as the input layer for the final model training. A method for predicting disease by using neural network to assist data exploration and big data data identification according to item 5 of the scope of patent application, wherein in this step (d3), the benefits of the model are established and used as accurate comparisons between different comparison models. Degree method meets The relational expression, where SD (standard deviation) is the standard deviation of all true values in the database of the training model, mean is the average of all true values in the database, and MAPE is the average absolute value error rate of the prediction model.