Research On A Deep Learning-based Pediatric Clinical Diagnosis Algorithm

In recent years,with the popularization of digital medicine in the global scope,the application of artificial intelligence in the medical industry is also gradually carried out.Studies such as death prediction,cancer diagnosis and medical image recognition have been published one after another.Recently,there are still three challenges in the clinical application of artificial intelligence technology in medicine.First,existing medical artificial intelligence applications mostly focus on the stage of medical picture processing and prediction based on structured data.However,deep mining and auxiliary diagnosis on non-structured electrical medical records is still at its early stage.Second,most of existing clinical decision support systems are built on the basis of logic reasoning on knowledge bases.This pattern leads to the geometric increase in the complexity of the pre-processing works that medical professionals need to do every time when a new disease is expanded.Such complexity hinders the large-scale promotion of medical artificial intelligence algorithms.Third,it is difficult for the clinical decision support systems to accurately extract the patient's features from the medical records written by free-texts.This could prevent the auxiliary diagnosis algorithm from making accurately diagnosis among similar diseases.In summary,existing works generally are lack of an intelligent algorithm that could not only meet the requirement of clinical diagnosis,but also effectively improve the accuracy of diagnosis and prediction.In order to address the aforementioned challenges,this paper proposes a pediatric clinical auxiliary diagnosis algorithm(called NLP-BiRNN)integrating natural language processing technologies with deep learning method.NLP-BiRNN aims at assisting pediatricians with little clinical experience to make diagnosis prediction rapidly and accurately.The main research contents of this paper include:First,a natural language processing solution of massive medical texts is proposed for the unformatted texts freely written in electronic medical records;Second,a bidirectional recurrent neural network based method is used to train the model from massive Electronical Medical Records in Chinese.Third,a large-scale experiment was conducted on 81,476 pediatric electronic medical record datasets from a first-class hospital in China to verify the effectiveness of the proposed algorithm.The prediction accuracy of NLP-BiRNN algorithm is about 3.5%?6.5%higher than that of baseline algorithms,such as CNN,RNN,LSTM,etc.The research significance of this paper are,on the one hand,can help doctors with insufficient clinical experience or community outpatient hospitals without professional pediatricians to make both rapid and accurate diagnosis.Hereafter,patients can receive timely and effective diagnosis,so that medical resources can be made maximum good use of towards community or countryside hospitals.On the other hand,the natural language processing solution on massive Chinese electronic medical records proposed in this paper can assist clinicians and medical students to find the patient's major symptom information quickly and accurately,and effectively promote the clinical decision makings.