Research On Medical And Health Decision Making Support For Multimodal Data

The application of medical big data has a huge impact on the development of national health,and how to mine useful knowledge from multimodal medical data and provide decision support for clinical medical diagnosis and patient selection is a major challenge for medical big data applications.This thesis uses machine learning and deep learning methods to solve problems including loss of decision making efficiency caused by imbalanced data miscalculation or missing judgment,difficulty of obtain core features caused by high-dimensional attribute redundancy,contradictory issue caused by knowledge granularity and knowledge reasoning ability of unstructured text,and loss of deep learning network inference efficiency caused by small samples,few or none labeled medical images due to medical imaging principles and acquisition costs.The main contributions are as follows:To solve imbalanced classification problems of intelligent medical diagnosis,Ada C-TANBN algorithm has been proposed,which uses a variable misclassification cost determined by sample distribution probability to represent the misclassification cost of positive and negative samples.Its performance has been verified by using four medical imbalanced datasets and the results show that it is superior to other excellent algorithms in terms of accuracy,specificity,sensitivity,AUC value and ROC curve.To solve multi-attribute classification problems of intelligent medical diagnosis,hybrid heuristic GA-MLP algorithm has been proposed,which uses GA algorithm to optimize attribute weights and combines with the corresponding attributes as new inputs,then finds the MLP optimal parameter combination to complete classification.Its performance has been verified by using multi-attribute medical datasets and the results show that it is superior to other comparison algorithms in terms of accuracy,specificity,sensitivity,AUC value,ROC curve and epoch value.To solve sentiment polarity classification problems of medical text comments,semi-supervised learning algorithm based on mutual information feature weights has been proposed,which uses NTUSD simplified Chinese emotion dictionary of Taiwan university segment the crawled medical text comments,then builds semi-supervised learning model based on mutual information feature weights from attribute granularity.Its performance has been verified by using the corpus and the results show that it has excellent performance on medical text comments emotional polarity classification.To slove medical image intelligent diagnosis decision,convolutional neural network model based on image enhancement and transfer learning has been proposed.Firstly,it adopts brightness dimming and Gaussian noise to enhance medical image datasets,then adopts Le Net,Alex Net and Res Net network to implement transfer learning on the enhanced datasets.Its performance has been verified by using four breast medical image datasets and the results show that in addition to MRI images in.dcm format,Alex Net network has excellent performance in terms of classification accuracy,loss value and the computation time.