CN1234092C - Predictive modelling method application to computer-aided medical diagnosis - Google Patents

Abstract

The invention discloses a predictive modeling method suitable for computer-aided medical diagnosis, which includes obtaining symptom vectors of symptoms of objects to be treated through medical symptom detection equipment, and obtaining predictive results through predictive model processing. The method includes the following steps : (1) If the prediction model is not trained well, then execute step (2), otherwise go to step (6); (2) Use the historical cases to generate the initial training data set; (3) Use the initial training data set to train a neural network Network integration; (4) use neural network integration to process the initial training data set to generate a rule training data set; (5) use rule learning technology to generate a rule model from the rule training data set; (6) use the rule model to predict and give Results and explanations. The invention has the advantage of providing a high precision and high comprehensibility predictive modeling method for a computer aided medical diagnosis device.

Description

A Predictive Modeling Approach for Computer-Aided Medical Diagnosis

1. Technical field

The invention relates to a computer-aided medical diagnosis device, in particular to a high-precision and high-understandability predictive modeling method using neural network integration technology and rule learning technology.

2. Background technology

With the development of computer technology, computer-aided medical diagnosis equipment has become an important auxiliary diagnosis method because it is not affected by factors such as fatigue and emotion. Computer-aided medical diagnosis devices usually use some predictive modeling methods to analyze historical cases to establish a predictive model, and then use the predictive model to diagnose new cases, and the results are submitted to medical experts for further analysis and diagnosis. Reduce the workload of medical experts to a certain extent. Therefore, predictive modeling methods are the key to computer-aided medical diagnosis devices. On the one hand, due to the precision of medical diagnosis, the applicable predictive modeling method must have high precision; It has high comprehensibility, that is, it is necessary to provide an explanation for the diagnosis after the diagnosis conclusion is made. This is not only the need of the patient and his family members, but also the need of medical experts to check the diagnosis process. However, although existing technologies such as neural networks have high precision, they do not have high comprehensibility; and rule learning, etc., have high comprehensibility, but do not have high precision, which affects the performance of computer-aided medical diagnosis devices. had an adverse effect.

3. Contents of the invention

The purpose of the present invention is to provide a high-precision, high-understandability predictive modeling method to assist in improving the Performance of computer-aided medical diagnostic devices.

In order to achieve the stated purpose of the present invention, the present invention provides a method for predictive modeling using neural network integration technology and rule learning technology in machine learning, the method comprising the following steps: (1) if the predictive model is not well trained, then Execute step 2, otherwise go to step 6; (2) use historical cases to generate an initial training data set; (3) use the initial training data set to train a neural network ensemble; (4) use the neural network ensemble to perform an initial training data set Process to generate a rule training data set; (5) use rule learning technology to generate a rule model from the rule training data set; (6) use the rule model to predict and give results and explanations; (7) end.

The advantage of the present invention is that it provides a high-precision and high-understandability predictive modeling method for computer-aided medical diagnosis devices, so as to help improve the performance of computer-aided medical diagnosis devices.

The preferred embodiment will be described in detail below with reference to the accompanying drawings.

4. Description of drawings

Fig. 1 is a working flow chart of the computer-aided medical diagnosis device.

Figure 2 is a flow chart of the method of the present invention.

Fig. 3 is a flow chart of generating regular training data sets with neural network ensemble.

5. Specific implementation

As shown in Figure 1, the computer-aided medical diagnosis device uses medical symptom detection equipment such as body temperature, blood pressure measurement equipment, etc. to obtain the symptoms of the subject to be diagnosed, such as body temperature, blood pressure, etc., and then quantifies the symptoms to obtain a symptom vector, such as [t ₁ , _t2 ,..., _tn ], where _t1 represents the first symptom value, _t2 represents the second symptom value, and so on. The symptom vector is handed over to the prediction model for processing, and the digital representation of the prediction result and explanation can be obtained. After text processing, the final diagnosis conclusion and explanation submitted to the user are produced.

The method of the present invention is shown in FIG. 2 . Step 10 is the initial action. Step 11 judges whether the prediction model has been trained, and if it has been trained, it can handle the diagnosis task, then go to step 16; otherwise, it needs to be trained, go to step 12. Step 12 uses historical cases to generate an initial training data set, which is called L ₀ for the convenience of description. L ₀ contains the symptom vector and its category corresponding to each historical case, that is, the diagnosed specific disease category ("no disease" is also used as a category). Step 13 uses repeatable sampling techniques commonly used in statistics to generate N data sets from L ₀ , and uses each of the N data sets to train a neural network, and these neural networks constitute a neural network ensemble. N is a user-preset integer value such as 9, which determines the number of neural networks included in the neural network ensemble. The neural network used here can be any type of neural network as long as it can perform prediction tasks, for example, the multi-layer feed-forward BP network introduced in the neural network textbook can be used. Step 14 utilizes neural network integration to generate a rule training data set L ₁ for establishing a rule model, and this step will be described in detail in the following part with reference to FIG. 3 .

Step 15 in FIG. 2 uses L ₁ to train a regular model. The rule model is a prediction model composed of many IF-Then or similar rules, which is trained from a training data set (here L ₁ ) by a rule learning method. Any type of rule learning method can be used here, as long as the model it produces can perform prediction tasks, for example, RIPPER, C4.5 Rule, etc. introduced in machine learning textbooks can be used. Step 16 receives a symptom vector to be diagnosed. Step 17 submits the symptom vector to the trained regular model for prediction. Step 18 gives the prediction results generated by the rule model and the rules used in the prediction process, and these rules constitute the interpretation of the prediction results. Step 19 is the end state.

Because the predictive model that the method of the present invention establishes is a regular model, it has high comprehension; and because the method utilizes the neural network integration with high precision to produce the training data set that establishes the regular model, this can be regarded as the initial data The set has undergone benign processing such as denoising and enhancement, so the established rule model is also of high precision.

FIG. 3 illustrates step 14 in FIG. 2 in detail, and its function is to use neural network integration to generate a rule training data set L ₁ for building a rule model. Step 140 of Figure 3 is the initial state. Step 141 sets L ₁ as an empty set. Step 142 obtains a symptom vector and its category from the initial training data set L ₀ generated in step 12 of FIG. 2 . Step 143 sets a counter respectively for each category, and these counters are used for recording how many neural networks give the prediction result is this category, and each category here corresponds to the specific disease category diagnosed ("no disease" is also used as a category). Step 144 clears all counters. Step 145 sets the control parameter k to 1. k is an integer value greater than or equal to 1 but less than or equal to N in step 13 in FIG. Step 146 obtains the prediction result given by the kth neural network in the neural network ensemble for the symptom vector to be diagnosed. For the convenience of description, this result is called F _k . Step 147 increments the counter of the class corresponding to F _k by one. Step 148 increments k by one. Step 149 judges whether k is less than or equal to the number of neural networks in the neural network ensemble, that is, N in step 13 in FIG.

Step 150 of Fig. 3 compares the values in all counters, finds the counter with the largest value, and uses its corresponding category as the new category of the current symptom vector; if the values in multiple counters are all maximum values, then use The disease category with the greatest chance of occurrence in the category corresponding to these counters is used as a new category of the current symptom vector. Step 151 adds the current symptom vector and its new category to L ₁ . Step 152 judges whether there are unexamined symptom vectors in L0, if so, go to step 142; otherwise, go to step 153, which is the end state of FIG. 3 .

Claims (1)

1, a kind of forecast modeling method that is applicable to computer-aided medical diagnosis comprises the symptom of obtaining the follow-up object by the medical symptom checkout equipment, then symptom is quantized to obtain symptom vector [t ₁, t ₂..., t _n], t wherein _nRepresent n symptom value, the symptom vector is given forecast model and is handled, and the digitized representations form that can be predicted the outcome and explain is characterized in that this method may further comprise the steps:

(1) if forecast model does not train, execution in step (2) then, otherwise forward step (6) to;

(2) utilize historical case to produce the initial training data set;

(3) it is integrated to utilize the initial training data set to train a neural network;

(4) but the neural network of utilizing the repeated sampling technology to generate is integrated that the initial training data set is handled with the generation rule training dataset;

(5) utilize the rule learning technology to concentrate the generation rule model from regular training data;

(6) utilize rule model to predict and provide result and explanation;

(7) finish;

In (4), utilize the integrated generation of neural network to be used to set up the regular training dataset L of rule model ₁Step be:

(4.1) with L ₁Be changed to empty set;

(4.2) from initial training data set L ₀In obtain a symptom vector and classification thereof;

(4.3) for each classification a counter is set respectively, is used for writing down the generic number that predicts the outcome that neural network provides;

(4.4) with all counter O resets;

(4.5) controlled variable k is changed to 1, k be one more than or equal to 1 but smaller or equal to neural network integrated in the number N of neural network;

(4.6) obtain neural network integrated in k the F that predicts the outcome that neural network provides follow-up symptom vector _k

(4.7) with F _kThe counter of pairing classification adds 1;

(4.8) k is added 1;

(4.9) judge k whether smaller or equal to neural network integrated in the number N of neural network, if show that then other neural networks are investigated as yet in addition, forward step (4.6) to; Otherwise execution in step (4.10);

(4.10) value in all counters is compared, the counter that the value of finding out is maximum, and with the new classification of its corresponding class as current symptom vector; If there is the value in a plurality of counters to be maximal value, then to occur the new classification of the kinds of Diseases of chance maximum in these counter corresponding class as current symptom vector;

(4.11) current symptom vector and new classification thereof are added L ₁

(4.12) judge L ₀In whether also have the symptom vector of not investigating, if having then forward step (4.2) to; Otherwise enter step (4.13);

(4.13) finish.

Images