Research On Multi-label Classification Algorithm For Obstetric Auxiliary Diagnosis

Data-driven medical health processing has become a new obstetric development direction,especially the auxiliary diagnosis based on obstetric electronic medical records(EMRs),which is of great significance to improve the reproductive health of the population.In this thesis,obstetric auxiliary diagnosis problems are transformed into multi-label classification,to enhance the overall effect by improving the multilabel classification algorithm.The main work is as follows:There are similarities among some diagnostic results of obstetric auxiliary diagnosis.Strengthening the discrimination between them can help to improve the prediction precision.Therefore,in the Backpropagation for Multi-label Learning algorithm(BP-MLL),two error functions,pairwise labels quotient and pairwise labels product,are proposed to minimize the error function by maximizing the distance between relevant labels and irrelevant labels,thus improving the discrimination between pairwise labels.In the experiments of Yeast and Emotions public datasets,the average precision reaches 0.768 and 0.820,respectively.In the experiment of Chinese Obstetrics EMRs dataset,the average precision reaches 0.7667.Supplementary medical knowledge can improve the result of obstetric auxiliary diagnosis.Hence the Hierarchical Attention Network Integrating Label Knowledge(HAN-LK)is proposed.The model compares the similarity between the first disease treatment process record and all the label knowledge texts by cosine similarity.Several label knowledge texts with the highest similarity ranking are input into HAN together with the current first disease treatment process record.In the experiments of DeliciousMIL and Hep-categories public datasets,the average precision reaches 0.6386 and 0.8233 respectively by supplementing relevant labels knowledge in Wikipedia.In the experiment of Chinese Obstetrics EMRs dataset,the average precision reaches 0.8577 by supplementing labels knowledge of Obstetrics and Gynecology and the Baidu baike.