Multimodal Breast Cancer Data Analysis And Predictive Modeling

Breast cancer(BC)is the most common cancer and the second greatest cause of cancer deaths in women.The main cause of death in BC is distant metastasis.When BC patients have distant metastasis,it means that BC is in an advanced stage and has lost the opportunity for radical surgery.Therefore,it is important to improve the diagnostic accuracy of BC distant metastasis and find new biomarkers for early prediction of BC distant metastasis.Intuitively,multimodal models that aggregate more information are usually better than single-mode models.The combination of deep learning and multimodal BC data provides a method for more comprehensive analysis of the influencing factors of BC distant metastasis.In this study,we propose a deep fusion auto encoder prediction model based on multimodal feature fusion by integrating multimodal data,deepening the network depth of the encoding and decoding layers,adding batch normalization,linearly summing the uniformly encoded features in the fusion layer,and performing activation mapping.In the image modal data,to enhance the feature representation ability,prediction performance,and interpretability of the model,we propose a hybrid attention mechanism depth neural network model before feature fusion,and the improved hybrid attention mechanism is introduced into the depth neural network.The model introduces an improved hybrid attention mechanism into the deep neural network.The skip connection structure is used to improve the convolutional block attention module(CBAM),and the network depth of the attention module is deepened so that the attention module can effectively learn the weight information between feature channels and the weight information on spatial regions.In the missing processing of clinical structured data,we propose a multiple interpolation method based on threshold discrimination,which integrates four common interpolation methods,including random forest interpolation,support vector machine interpolation,maximum expectation interpolation,and baseline multiple interpolation.The interpolation process consists of intra group optimization and inter group optimization,in which the interpolation results are selected for each column of features,and inter-group optimization,in which the different interpolation methods are selected.This interpolation method uses a threshold constraint on the features and an error measure to effectively avoid interpolation of erroneous data.The results of interpolation experiments show that under different missing rates,multiple interpolation based on threshold discrimination is better than other four single interpolation methods and is more stable on BC dataset.Compared with other baseline models,the deep neural network model with hybrid attention mechanism not only has a significant improvement in prediction performance,but also has a significant improvement in convergence speed.This shows that the improved hybrid attention mechanism module effectively strengthens the model’s abilities for feature representation and feature extraction.The prediction performance of a deep auto encoder prediction model based on multimodal feature fusion is better than that of single-modal data,which indicates that the fused multimodal data provides more effective information for model prediction.The above study shows that the proposed model is a BC prediction model with better prediction performance,higher stability,and stronger interpretability and can provide a meaningful biomarker.This study highlights the value of this model in BC distant metastasis and provides an applicable method for the subsequent study of this model.