JPH0784981A - Diagnostic processor - Google Patents

Abstract

(57) [Summary] [Object] With regard to diagnostic processing, it is an object of the present invention to provide a diagnostic processing device capable of relatively easily configuring a diagnostic system using a neural network. [Structure] The weight generation unit 1 is composed of a first neural network having a phenomenon input terminal and a weight output terminal corresponding to each phenomenon item of a plurality of phenomenon items, and a state of the phenomenon corresponding to each phenomenon item. By inputting a phenomenon value representing a predetermined value into the phenomenon input terminal, each required weight value is output to each weight output terminal. The weighting unit 2 calculates a weighted phenomenon value for each phenomenon item from the phenomenon value and the weight value output from the weight generation unit 1, and the diagnosis unit 3
Is a weighted phenomenon input terminal corresponding to each phenomenon item,
A second neural network having a required cause output terminal, and by inputting the weighted phenomenon value of each of the phenomenon items to the weighted phenomenon input terminal, it is possible to estimate the required cause at the cause output terminal. Configure to output a value.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention easily configures a device for processing a diagnosis for estimating a cause of a disease or a failure from a phenomenon by a neural network, and particularly a neural network for obtaining a highly accurate diagnostic result. The present invention relates to a diagnostic processing device that can be performed.

[0002]

It is well known that a diagnostic system such as a medical diagnostic system that automatically estimates the cause from a symptom is constructed using a so-called expert system. Is.

However, the diagnostic system constructed by the expert system requires knowledge and rule coding, which requires a great number of man-hours. If the knowledge is increased, the diagnostic speed becomes slow and the system becomes large. There are problems such as being expensive.

Therefore, a neural network has come to be used as an alternative to the expert system. As is well known, a neural network is composed of, for example, three layers of an input layer, an intermediate layer, and an output layer as illustrated in FIG. Are configured to be input to all nodes of the node, and the sum of the inputs multiplied by these weights is configured as the input of this node, and the weights can be different for each node pair. Each node outputs a predetermined function value determined by the input value.

In order to make a diagnosis using such a neural network, for example, for a phenomenon item of each symptom (for example, fever, cough, headache, etc.), the presence or absence of the symptom is represented by a value of 0 or 1 Is input to the node corresponding to each phenomenon item in the input layer of the neural network, and each node in the output layer is assigned to each probable cause (eg, disease name) that should be the recognition result, for example, when a "flu" is estimated. The neural network is trained so that the first node outputs a value as close to 1 as possible, and the nodes assigned to other causes output a value close to 0.

This learning is well known, for example, "cold".
When a number of patterns with various variations are sequentially input according to the symptom pattern to be recognized and the required output signal as in the above example is designated as the teacher signal, each node is connected between adjacent layers in the neural network. This is done for all necessary causal items because it is done by reducing the difference between the two signals by changing the path weights according to a predetermined algorithm based on the difference between the actual output signal and the teacher signal. This completes the learning.

In this way, the neural network can automatically learn the desired relationship between the input and output, but in order to improve the accuracy of cause estimation, various patterns for each estimated cause should be used as much as possible. It is necessary to collect and learn, and generally, it is easy to require preparation of a huge amount of data and learning for a long time in order to obtain high cause estimation accuracy.

In the case of diagnosing a disease, etc., in the method of determining the disease name based only on the presence / absence of the symptom, the weights of all the symptoms are uniformly evaluated. In particular, it is necessary to use a large number of cases for learning because they have the same effect.

However, it is difficult at present to collect a sufficient number of cases depending on the situation of the dairy farm and the situation of the veterinarian when trying to diagnose a disease of a dairy cow in dairy farming.

On the other hand, in diagnosing a disease, a doctor accurately specifies a disease name by considering the combination of symptoms and the relative weight of each symptom that he has empirically based on the combination. Therefore, if the symptom is weighted as such as the preprocessing for determining the disease name, the influence on the diagnosis result can be reduced even if the symptom is not important.

In order to obtain information on the relationship between the combination of symptoms and the weighting for each symptom in each case, the empirical knowledge of the veterinarian in the above example should be sought out, arranged and digitized appropriately. Good, so easier than collecting a huge number of cases from a dairy farm.

It is an object of the present invention to provide a diagnostic processing device capable of relatively easily constructing a diagnostic system capable of increasing the accuracy of estimating a cause with a relatively small number of cases by considering the weight of a phenomenon.

[0013]

FIG. 1 is a block diagram showing the configuration of the present invention. The diagram shows the configuration of the diagnostic processing apparatus, which includes a weight generation unit 1, a weighting unit 2, and a diagnostic unit 3.

The weight generation unit 1 is composed of a first neural network having a phenomenon input terminal corresponding to each phenomenon item of a plurality of phenomenon items and a weight output terminal corresponding to each phenomenon item. By inputting a phenomenon value representing a state of a phenomenon corresponding to a phenomenon item with a predetermined value to the phenomenon input terminal, each required weight value is output to each of the weight output terminals.

The weighting unit 2 determines, for each phenomenon item,
A weighted phenomenon value is calculated from the phenomenon value and the weight value output from the weight generation unit 1. The diagnosis unit 3 is composed of a second neural network having a weighted phenomenon input terminal corresponding to each phenomenon item and a required cause output terminal,
By inputting the weighted phenomenon value of each of the phenomenon items to the weighted phenomenon input terminal, a required cause estimation value is output to the cause output terminal.

[0016]

With the diagnosis processing apparatus of the present invention, when a phenomenon value group representing the presence or absence of each phenomenon as 1 or 0 is input as
The apparatus first outputs the weight value of each phenomenon item from the input by the first stage neural network.

Next, from the weight value and the phenomenon value, for example, the weighted phenomenon value obtained as the product of the two is obtained, and the estimated value of the cause is output as the input of the second stage neural network.

As a result, with respect to each stage of the neural network, it is possible to easily arrange the input and output data for learning, and it is possible to perform learning so as to obtain a required output with high accuracy even in a comparatively small number of cases, so that highly accurate diagnosis can be performed. The processing device can be constructed relatively easily.

[0019]

FIG. 2 is a diagram showing an embodiment of the present invention. As a system for diagnosing a disease, the weight generation unit 1 has a fever,
The values corresponding to the presence / absence of each symptom are input as 1/0 to the input layer of the first neural network via the phenomenon input terminal corresponding to each symptom of illness such as cough and headache, and the output layer of each symptom is input. Output weight.

Therefore, the first neural network has an input layer and an output layer each having nodes equal to the number of phenomenon items, and the input / output data set as shown in FIG. Patterns of various symptoms are collected, data of them are given, and learning is performed as described above.

In FIG. 2, the weighting unit 2 outputs, for each phenomenon item, a product of the product of the input phenomenon value and the output weight value, or passes the weight value as it is when the phenomenon value is 1, When the phenomenon value is 0, each gate 4 is composed of a circuit that outputs 0, and the output is input to the diagnosis unit 3.

The diagnosis unit 3 has an input layer having nodes equal to the number of phenomenon items, colds, gastric ulcers, pneumonia, and heat rashes shown in the figure.
It is composed of a second neural network in the output layer having nodes equal in number to the required probable causes such as gastritis.

In the second neural network, the cause output is set to 1, and the phenomenon value group data in which the symptom caused by the cause is weighted is collected for each cause output,
Give these data and carry out learning.

As described above, when a phenomenon value indicating the presence / absence of a symptom is input to all the phenomenon input terminals of this device, an estimated value in the range of 0 to 1 is output to each cause item of the cause output terminal. The size can be used as the leading data for diagnosis.

[0025]

As is apparent from the above description, according to the present invention, in the configuration of the diagnostic system using the neural network, the input / output data for learning of the neural network can be easily aligned, and the accuracy is high even in a relatively small number of cases. Learning is possible to obtain the required output, and there is a remarkable industrial effect that a highly accurate diagnostic processing device can be constructed relatively easily.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of the present invention.

FIG. 2 is a block diagram showing an embodiment of the present invention.

FIG. 3 is a diagram for explaining learning data for weight generation.

FIG. 4 is a diagram illustrating a neural network.

[Explanation of symbols]

 1 Weight generation unit 2 Weighting unit 3 Diagnostic unit 4 Gate

Claims (1)

[Claims] 1. A weight generation unit (1), a weighting unit (2), and a diagnosis unit (3), wherein the weight generation unit (1) has a plurality of phenomenon items for each phenomenon item. A first neural network having a corresponding phenomenon input terminal and a weight output terminal corresponding to each phenomenon item, and a phenomenon value representing a state of a phenomenon corresponding to each phenomenon item with a predetermined value, By inputting to the phenomenon input terminal, each required weight value is output to each weight output terminal, and the weighting unit (2), for each phenomenon item, the phenomenon value and the weight generation unit (1). From the weight value output by
The weighted phenomenon value is calculated, and the diagnosis unit (3) is composed of a second neural network having a weighted phenomenon input terminal corresponding to each phenomenon item and a required cause output terminal. A diagnostic processing apparatus configured to output a required cause estimation value to the cause output terminal by inputting the weighted phenomenon value of an item to the weighted phenomenon input terminal.