摘要
疾病预测模型通过利用收集到的医疗数据,能够在患者疾病发作前准确地进行疾病预测.目前在疾病预测方面深受欢迎的深度神经网络,它依靠增加网络层数来提升模型的准确率,利用梯度下降来进行权重的更新,而这导致了模型梯度爆炸、训练速度慢等问题.一旦数据更新,深度神经网络需要重新训练,进而导致模型更新困难.宽度学习(Broad Learning System,BLS)无须梯度下降的特性与其可通过增量学习快速重构的优势为有效解决上述问题提供了技术方案,但是BLS无法提取到隐藏在医疗数据中深层次的特征,其在复杂的医疗环境下仍然表现不佳.针对该问题,本文提出一种基于降噪自动编码器(Denoising AutoEncoder,DAE)与宽度学习的增量式疾病预测模型——DAE-BLS.所提模型将DAE引入BLS的架构设计中,结合了DAE在混乱环境下的降噪能力与BLS的简洁快速的特点,既保证了高效的运算能力又增强了特征提取能力,因而更适用于复杂医疗环境.将DAE-BLS在包含不同格式以及不同数据量的糖尿病、心力衰竭、心电异常和乳腺癌数据集上进行模拟预测实验,实验结果表明,DAE-BLS能够在保留宽度结构的神经网络快速高效特点的同时,在不同格式的数据上表现出很好的性能,分别达到96.62%,94.53%,98.50%与83.64%的准确率,并能在需要更改模型结构时通过增量学习技术快速重构以适应用户不断变化的疾病数据.
Disease prediction models use collected medical data to accurately predict a patient's disease before its on⁃set.At present,deep neural network,which is popular in disease prediction,relies on increasing the number of network lay⁃ers to improve the accuracy of the model,and uses gradient descent to update the weight,which leads to problems such as model gradient explosion and slow training speed.At the same time,once the data is updated,the deep neural network needs to be retrained,which makes it difficult to update the model.Broad learning system(BLS),which does not need gra⁃dient descent and has the advantage of rapid reconstruction through incremental Learning,provides a technical solution to solve the above problems effectively.However,BLS cannot extract the deep features hidden in medical data,and still per⁃forms poor in complex medical environment.To solve this problem,we propose an incremental disease prediction model based on denoising autoencoder(DAE)and BLS,called DAE-BLS.In the proposed model,DAE is introduced into the ar⁃chitecture design of BLS.The model combines the denoising ability of DAE in chaotic environment with the simplicity and speed of BLS,which not only ensures the efficient computing capability of the model but also enhances the feature extrac⁃tion capability of the model,making it more suitable for complex medical environment.Applying DAE-BLS to prediction experiments on diabetes,heart failure,ECG abnormalities and breast cancer datasets with different formats and data vol⁃umes,experimental results show that DAE-BLS can retain the fast and efficient characteristics of BLS,and show good per⁃formance in different data formats,reaching 96.62%,94.53%,98.50%and 83.64%accuracy respectively,and can be rapidly reconstructed to adapt to users'changing disease data through incremental learning techniques when the model structure needs to be changed.
作者
漆华妹
胡宇轩
袁正一
QI Hua-mei;HU Yu-xuan;YUAN Zheng-yi(School of Computer Science and Engineering,Central South University,Changsha,Hunan 410075,China)
出处
《电子学报》
EI
CAS
CSCD
北大核心
2023年第6期1474-1485,共12页
Acta Electronica Sinica
关键词
神经网络
自动编码器
疾病预测
宽度学习
增量学习
智慧医疗
neural network
auto-encoder
disease predictions
broad learning system
incremental learning
smart healthcare
作者简介
漆华妹,女,1978年出生,江西南昌人.中南大学计算机学院副教授,硕士生导师.主要研究方向为机器学习、边缘计算、云计算、物联网、5G通信等.E-mail:qhm@csu.edu.cn;胡宇轩,男,1999年出生,湖南娄底人.中南大学计算机学院硕士研究生.主要研究方向为机器学习与智慧医疗.E-mail:204712249@csu.edu.cn;袁正一,男,2001年出生,江西南昌人.中南大学计算机学院本科生.主要研究方向为机器学习与智能算法.E-mail:2250100280@qq.com。
