Research And Implementation Of Intelligent Disease Consultation Technology Based On Machine Learning

According to the Ministry of Health’s report[1],as many as 70%of urban residents are in sub-health status,only 4.8%go to the hospital and nearly 80%seek information online.The traditional search engine is difficult to provide the correct 1edical information.Public data[2],2020,at least 50%of human-computer interaction devices will have voice capabilities.Based on the above background,this thesis intends to study the intelligent interrogation system based on the voice dialogue mode.The diagnostic procedure of the intelligent interrogation system is as follows:the user states the condition,and the system assigns the initial disease hypothesis according to the statement text.When the condition information is too small to be confirmed,the system enters the interactive intelligent interrogation and matches with the "current symptom" And "initial disease hypothesis" to ask questions about the user and further classify the disease based on user feedback.The research of intelligent interrogation system has the following three difficulties:1)The amount of text of the statement of illness gradually accumulates as the interview progresses,and different amounts of text require different methods of disease classification.There is a problem of sparse text features in the pre-diagnosis period[3],and the sparsity of features in the latter part of the interview period is weakened.The main problem is that there is a higher requirement on the accuracy of the disease classification.2)Identify the symptoms described in the statement of illness.Interactive intelligent interrogation needs to ask users according to the "current symptom" and "initial disease hypothesis." Therefore,studying symptom identification is an important task in this thesis.Traditional dictionary matching fails to account for more colloquial Chinese,similar and similar words.3)Matches the symptoms associated with "Current Symptoms" and "Initial Illness Assumptions." The usual symptom selection strategy is to build a knowledge base of disease symptoms,but the establishment of a disease knowledge base requires medical expert knowledge and extensive rule inference.In view of the existing problems of the intelligent interrogation system,this thesis proposes the corresponding solutions based on the previous studies.Specifically,the contribution and innovation of this article include the following points:1)In order to solve the problem of sparseness of text features in pre-interview state of illness presentation,this thesis proposes a feature expansion method based on similarity measure of keywords.Firstly,the keywords in the text are extracted,and then the keywords are semantically extended.The semantic expansion is to use similar measures to find words with similar or similar semantic meaning to the keywords in the interrogation corpus.The final experiment shows that F1 value is increased by about 2%respectively in the interview corpus and Fudan corpus used by predecessors.2)For the latter part of the interview,the complexity of the corpus is increased,the learning depth is not enough,and the diagnosis accuracy of the disease needs to be higher.In this paper,a convolution neural network is used to classify the long texts of the later period of interrogation.Experiments show that,for long texts,convolution neural network diagnostic accuracy than traditional machine learning 3.9%.3)Aiming at the situation that there are many colloquial,similar and similar words in Chinese,based on the traditional dictionary matching,"Synonyms Lin"[4]and word vector similarity calculation are respectively used,and different symptom matching strategies are adopted.Finally,the three methods of weighting the recognition results,select the score to reach the threshold of the symptoms.The final experiment showed that the recall rate of dictionary matching rose from 50.89%before the improvement to 83.51%,In response to the selection strategy of symptoms,this thesis establishes a naive Bayes-based correlation symptom matching model.The input to the model is the extracted symptom and the initially diagnosed disease,and the output is the conditional probability of the relevant symptom.This article calculates the frequency of co-occurrence of each disease and symptom as a baseline for symptom selection.The final experiment shows that the symptom selection space for the relevant symptom matching model proposed in this thesis is only 12.95%of the baseline.4)Realization of intelligent interrogation system.Based on the research of algorithms,this thesis has initially realized the intelligent inquiry system,which includes Android client and server.The client mainly includes the voice interaction module and the HTTP communication module,and the server mainly includes a text preprocessing module,a symptom identification module,a disease diagnosis module and a related symptom matching module.The research results of this thesis mainly include:improving the accuracy of symptom recognition,effectively reducing the impact of feature sparseness on disease diagnosis,improving the accuracy of disease diagnosis at the later stage of diagnosis,and proposing a new symptom selection strategy In interactive conversation.These achievements will provide a comprehensive theoretical reference and support for the practical application of intelligent interrogation.