The Study On Multi-disease Risk Prediction Model Based On Deep Learning

With the development of medical digital technology,artificial intelligence and big data technology,the medical model has gradually changed from treatment-based to prevention-based.Combining artificial intelligence and big data technology to handle disease risk prediction is a key study of the intelligent medical.Disease risk prediction refers to the discovery of the latent risks and trends of the disease,which plays an important role in the prevention,intervention and management of the disease.In real life,it is often found that people have latent risks and trends to get multiple diseases at the same time.This problem belongs to multi-disease risk prediction problem.In order to handle multi-disease risk prediction problem efficiently,researchers have proposed many preferable algorithms.Deep learning is very popular in recent years.This paper uses deep learning methods to handle multi-disease risk prediction problem,and focuses on deep learning algorithm design and improvement in the disease risk prediction model design.Firstly,this paper uses the problem transformation method to convert the multi-disease risk prediction problem into a multi-label learning problem.Because the problem transformation method makes the algorithm independent,and only need to perform multi-disease label conversion work.Binary Relevance(BR)and Label Powerset(LP)are chosen to convert multiple disease labels into multi-labels separately.Based on the problem transformation methods,this paper designs a novel convolutional neural network framework and named it GroupNet and then combine the GroupNet with BR and LP correspondingly.The core component of the GroupNet network framework is the Group Block proposed in this paper.The group block consists of two parts: group convolution and cluster convolution,which has the effect in alleviating convolution redundancy and clustering.The experimental results demonstrate that the GroupNet achieves better performance than several classical convolutional neural network frameworks.Secondly,since BR method does not take correlation between labels into consideration,and this paper proposes a loss named correlated loss to alleviate this shortcoming.In this paper,the CL is compared with the focal loss(FL)and the cross entropy loss(CE).The experimental results show that the CL gets better performance than other two classic loss functions.In order to further improve the performance of multi-disease risk prediction model,this paper organically combines the GroupNet and ensemble algorithms(such as random forest algorithm and LightGBM)to get an integrated model,which can take use of advantages of different algorithms.The experimental results show that the integrated model based on two powerful methods increases at least 1% better than the single GroupNet and the ensemble algorithm.The experimental results show that the integration model based on two powerful methods increases at least 1% better than single GroupNet and ensemble algorithm.Finally,in order to verify the effectiveness of the proposed methods,this paper also uses several classical machine learning algorithms to compare with the proposed methods using label-based evaluation measures.The proposed GroupNet and the integrated model proposed achieve the accuracy with 81.13 and 82.68% respectively,and they achieve better performance than compared machine algorithms obviously.Furthermore,they also get similar improvements in precision,recall and F1 when comparing with other methods.