节点文献
多源医疗数据的智能分析与应用研究
Intelligent Analysis and Application of Multi-source Medical Data
【作者】 李振;
【导师】 李学相;
【作者基本信息】 郑州大学 , 计算机技术, 2018, 硕士
【摘要】 医疗健康数据是人们感知健康状况的重要途径,而健康数据主要通过医院的医疗化验单、检测报告和可穿戴的传感器等检查手段获取。从不同医院、不同途径获取的医疗健康数据存在着多源化、多层次以及结构差异较大等问题。如何对这些多源数据进行处理和分析成为了当前亟待解决的问题。本文的目标是通过深度学习技术对多源的健康数据进行分析,并挖掘隐含的内在价值信息,使其可以为医院管理提供更好的决策支持,为用户提供更加全面精准的健康风险评估。本文的研究内容如下:首先,针对纸质医疗单据的多源性问题,本文提出了一种基于深度学习的多源医疗单据的识别的模型。该模型对医疗单据图像进行降噪、抗扭斜、膨胀和腐蚀等形态学操作;然后提出了一种改进的inception卷积神经网络对阈值分割后的图像进行分类,为医疗自动识别医疗单据中的文字奠定基础。其次,针对图像中文字识别难的问题,本文提出了一种适用于识别多源医疗单据文本的深度学习模型:SCRNN。该模型将卷积神经网络与长短周期神经网络相结合以充分提取图像中的空间特征及文本的上下文关系,达到将医疗单据的数据结构化,从而可以完成个人健康数据的采集,逐渐形成历史健康档案。与开源引擎Tesseract相比,SCRNN中英混合词组的识别方面准确率较高,准确率为97.2%。最后,针对心脏病难以通过心电图进行自动诊断的问题,本文提出了一种基于长短周期神经网络的实时ECG诊断模型M-ECG。该模型采用标记后的双导联ECG信号作为训练数据,在模型训练过程中,将三组连续的心拍作为下一个周期的输入数据,同时将多个导联的不同时刻的数据用以训练诊断模型,从而有效地提高了M-ECG预测的准确率。M-ECG模型已在MIT-BIH心律失常公开数据及上进行了验证,实验结果表明:M-ECG具有比1D-CNN更高的精度和更好的性能,与同类的深度学习网络相比,该模型对ECG信号预测的准确率有所提高。
【Abstract】 The health data,mainly obtained from medical examinations such as blood test and ElectroCardioGram(ECG),is an important way for people to perceive their physical conditions.In recent years,the development of wearable smart devices provides new ways and tools for people to get their health data.However,health data obtained from different hospitals and different ways is diversified,multi-layered.How to deal with and analysis these multi-sourced data is becoming a pressing issue.The goal of this report is to analysis health data using deep leaning and to excavatepotential information within the data,which provides better decision support for hospital management and more accurate assessments of health risk for people.The main contributions of this report are as follows:Firstly,for the multi-source problems of paper medical examinations,this paper presents a model of multi-source medical examinationsclassifier based on deep learning.The model performs morphological operations such as noise reduction,anti-skew,swelling,and corrosion on medical examinations images;then,an improved inception convolution neural network is proposed to classify the threshold-segmented images.It lays the foundation of textsrecognition in medical examinations.Secondly,in view of the difficulty of text recognition in images,this paper proposes a deep learning model: SCRNN which is suitable for recognizing the text of multi-source medical documents.This model combines convolution neural networks with long-term short-term neural networks to fully extract the spatial features of the images and the contextual relationships of the texts,so as to structure the data of the medical examination so that the personal health data can be collected,and historical health can be gradually formed.file.Compared with the open source engine Tesseract,the recognition accuracy of the SCRNN Chinese-English mixed phrase group is relatively high,and the accuracy rate is 97.2%.Finally,to achieve the automatic diagnosis of heart diseases through electrocardiogram,a neural network named M-ECG based on Long Short TimeMemory network is proposed to diagnose ECG in real time.Labeled bipolar leads of electrocardiogram is used to train M-ECG,and signals within three continuous heart beat from the two polars are utilized as the input data during the training process,which effectively improves the prediction accuracy of M-ECG.The proposed model has been validated on the public MIT-BIH arrhythmia data set,and experimental results demonstrate that M-ECG achieves better accuracy and performance over1D-CNN and improves the prediction accuracy compared to other deep learning networks.
【Key words】 Multi-source medical examinations; OCR; ECG classification; Deep learning;