Research On Prediction Of Diabetes With Multimodal Data And Knowledge

Diabetes is a chronic disease with abnormal blood sugar as the main manifestation.Long-term abnormal blood sugar can lead to a series of complications.At present,artificial intelligence methods such as machine learning are widely used in the prediction of diabetes risk.Most of the current diabetes prediction algorithms use the numeric data of patients,but relying solely on physical examination data for diabetes risk prediction is not reliable.Therefore,this thesis analyzes and studies the risk prediction of diabetes,and proposes a diabetes prediction method with multimodal data and knowledge.This thesis mainly does the following three aspects of work:(1)This thesis proposes a diabetes prediction method based on multimodal data.The existing diabetes prediction algorithm mainly relies on numerical physical examination data.Since it is difficult to make full use of other types of data generated by patients,the prediction effectiveness is not as good as expected.This thesis uses the numerical physical examination data of patients and their own symptom descriptions,and combines the information contained in the two for the prediction of diabetes.In the selection of diabetes prediction model,this thesis compares different types of deep learning models,and selects a prediction model based on convolutional neural network according to the experimental results.The characteristics of multimodal data and convolutional neural network architecture,which can make full use of the patient’s physical examination data and symptom description to improve the accuracy of diabetes prediction.The experimental results show that our method based on multimodal data is superior to most diabetes prediction algorithms based on single modal data in accuracy,and is more flexible in the selection of data sources.(2)In this thesis,a diabetes prediction method with multimodal data and knowledge is proposed.The multimodal data used in this thesis contains patient symptom descriptions.Symptom descriptions belong to short medical texts,and there are problems such as insufficient contextual information and rare words.Based on multimodal data,this thesis introduces external knowledge for semantic enhancement of symptom descriptions.Firstly,this thesis uses the entity link method to find relevant entities in the patient’s symptom description,and then uses these entities to find entity related conceptual knowledge in the knowledge base.In the process of encoding conceptual knowledge,this thesis introduces some attention mechanisms to measure the importance of external knowledge.Finally,the conceptual knowledge with attention weight and the patient’s symptom description are fused to achieve the purpose of semantic enhancement of symptom description.The experimental results show that the combination of external knowledge and attention mechanism can fully mine the effective information in symptom description,especially when dealing with rare medical vocabulary.In addition,the proposed method can further improve the accuracy of diabetes prediction with certain interpretability.(3)Based on the proposed algorithm,a diabetes prediction system is designed and implemented.The system adopts a modular design,so that users can freely combine their own data to obtain the probability of diabetes according to the type of data they hold.The system test results show that the system has good practicability and accuracy,which also verifies the effectiveness and practicability of the proposed method from the engineering point of view.