WO2022166158A1 - 一种基于卷积生存网络的血透并发症长期风险预测系统 - Google Patents

一种基于卷积生存网络的血透并发症长期风险预测系统 Download PDF

Info

Publication number
WO2022166158A1
WO2022166158A1 PCT/CN2021/113135 CN2021113135W WO2022166158A1 WO 2022166158 A1 WO2022166158 A1 WO 2022166158A1 CN 2021113135 W CN2021113135 W CN 2021113135W WO 2022166158 A1 WO2022166158 A1 WO 2022166158A1
Authority
WO
WIPO (PCT)
Prior art keywords
convolutional
survival
network
risk
hemodialysis
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2021/113135
Other languages
English (en)
French (fr)
Inventor
李劲松
王丰
朱世强
田雨
周天舒
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang Lab
Original Assignee
Zhejiang Lab
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang Lab filed Critical Zhejiang Lab
Publication of WO2022166158A1 publication Critical patent/WO2022166158A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems

Definitions

  • the invention belongs to the technical field of medical treatment and machine learning, and in particular relates to a long-term risk prediction system for hemodialysis complications based on a convolutional survival network.
  • Recurrent neural network is a type of recurrent neural network that takes sequence data as input, performs recursion in the evolution direction of the sequence, and connects all nodes in a chain.
  • Convolutional neural network is a kind of feedforward neural network with convolution calculation and deep structure. It has the ability of representation learning and can classify the input information according to its hierarchical structure. Therefore, it is also called “translation invariant artificial neural network. Neural Networks".
  • the features identified by recurrent neural networks are complex and abstract, and it is difficult to visualize and understand them and provide enlightening results.
  • Convolutional neural networks can often only predict the risk value up to a single time point, and cannot carry out long-term and continuous risk prediction for the risk of complications, and it is difficult to provide accurate and effective decision support for clinicians.
  • traditional recurrent neural networks and convolutional neural networks can only analyze simple data structures, but cannot deal with censored data commonly used in clinical analysis.
  • Deep learning is a popular research direction in the field of artificial intelligence. Many achievements have been made in search, machine translation, natural language processing, multimedia learning, recommendation and personalization technology, and other related fields. However, at present, deep learning is rarely applied to blood. Through the field of prognostic analysis of complications.
  • the purpose of the present invention is to propose a long-term risk prediction system for hemodialysis complications based on convolutional survival network in view of the shortcomings of traditional deep learning methods;
  • the visualization method of the convolution kernel obtains interpretable and inspiring results;
  • the present invention improves the convolutional neural network structure, so that it can make full use of the censored data, predict the relative risk of patients, and then carry out long-term, continuous risk analysis predict.
  • a long-term risk prediction system for hemodialysis complications based on convolutional survival network includes: a data acquisition module for collecting blood pressure information of hemodialysis patients; A data preprocessing module for processing missing values, normalizing and normalizing raw data into a two-dimensional matrix; a learning prediction module for deep learning modeling; a result display module for visual output and long-term risk changes;
  • the processing process of the data preprocessing module is specifically as follows: the blood pressure data of continuous hemodialysis is regularized into a two-dimensional matrix, arranged in the order of hemodialysis treatment, and each row of data corresponds to a hemodialysis treatment process, which is normalized by Min-Max. the two-dimensional matrix to ensure that the blood pressure waveform is preserved;
  • the processing process of the learning prediction module includes two parts:
  • Convolutional survival network is trained using convolutional neural network architecture combined with Cox proportional hazards loss function; the convolutional survival network consists of multiple convolutional layers and fully connected layers The input is a two-dimensional matrix output by the data preprocessing module. Each convolution layer abstracts the feature layer by layer, and finally abstracts it into several pattern features. Each pattern feature is output as a node through the fully connected layer to represent the relative risk of an event. ;Optimize network parameters by Cox proportional hazards loss function;
  • the hazard function h(t,x) represents the probability of an event occurring at a given moment, and the formula is as follows:
  • t represents time
  • p is the number of neurons in the last fully connected layer of the convolutional survival network
  • h 0 (t) is the benchmark risk function
  • ⁇ T x is the relative risk, the output of the convolutional survival network
  • n refers to the number of patients
  • T i and T j represent the individual patient xi and the survival time of x j ;
  • R i is the sample set at risk at time t i ;
  • the risk function h(t, x) describing the long-term risk is calculated according to the formula (1).
  • the data preprocessing module first screened out the data with systolic blood pressure less than 60mmHg or greater than 250mmHg, and then screened the systolic blood pressure data of 36 consecutive hemodialysis sessions into a two-dimensional matrix, arranged in the order of hemodialysis treatment, and each row of data corresponds to In a 5-hour hemodialysis treatment process, the interval between columns is 10 minutes, a total of 36 rows, and each row has 30 points.
  • the convolutional survival network receives the two-dimensional matrix output by the data preprocessing module, firstly identifies the one-dimensional blood pressure pattern, then fills the perceptual field into a square, abstracts the feature layer by layer, and then connects to the global mean pool.
  • the transformation layer obtains several pattern features, and each pattern feature is output as a node through the fully connected layer to represent the relative risk of an event; the activation function of the last layer is Linear, and the activation functions of the remaining layers are Relu.
  • the input of the convolutional survival network is a 1-channel 36*30 size two-dimensional matrix output by the data preprocessing module, and the one-dimensional blood pressure pattern is identified through 16 1*5 convolutions, and then Through 32 vertical 5*1 convolution operations, the perceptual field is filled into a square, and then 2 layers of 16 5*5 convolution kernels and 1 layer of 9 3*3 convolution kernels are connected to abstract the feature layer by layer, and then Then, the global mean pooling layer is connected to obtain 9 pattern eigenvalues, and the 9 pattern eigenvalues are output as a node through the fully connected layer to represent the relative risk of an event.
  • y i _pred is the output of the convolutional survival network, corresponding to ⁇ T x i ; since M is arranged in descending order of T, It is the exponential accumulation sum of the output of the first i rows of the matrix M after passing through the convolutional survival network, that is, the first i rows The cumulative sum of , let the cumulative sum function be recorded as cumsum;
  • ⁇ i is the value of the individual survival outcome event E in the ith row of matrix M, denoted as E i ;
  • the loss function in order to make the loss function corresponding to the training set and the test set take the same magnitude, the summation function sum is replaced with the mean value function mean, and the loss function is expressed as:
  • the learning prediction module when optimizing the network parameters through the Cox proportional hazards loss function, the training set is randomly divided into 10 batches, and the proportion of the survival outcome events E in each batch is equal, and each batch is divided into 10 batches.
  • the survival data are arranged in reverse order of survival time T to calculate the loss function, thus traversing the data set to update the network parameters 10 times.
  • the present invention utilizes the convolutional neural network to process multi-dimensional hemodialysis time series features; the convolutional neural network combines the Cox proportional hazards assumption to propose a convolutional survival network; on the basis of using the convolutional survival network, the Breslow estimated a baseline risk function to calculate long-term changes in risk for patients.
  • the invention can make full use of the censored data common in medical research; the main structure of the convolutional neural network is applied, which is convenient for visual analysis, and makes interpretable and inspiring results; it can predict the long-term risk changes of patients.
  • Fig. 1 is the structural block diagram of the long-term risk prediction system of hemodialysis complications based on the convolutional survival network of the present invention
  • Figure 2 is a flowchart of long-term risk prediction of hemodialysis complications based on convolutional survival network
  • Figure 3 is a schematic diagram of the convolutional survival network model.
  • the convolutional survival network described in the present invention a convolutional neural network applied to survival analysis, which can process time series and image data and perform survival analysis and risk prediction; long-term risk prediction: different from risk prediction at a time point , "Long-term risk prediction" predicts continuous risk changes over a long period of time; Censored data: If there is no resultant event at the specified end time, it is called censored data, and the time from the starting point to the censorship is called the censored data. Censoring time.
  • a long-term risk prediction system for hemodialysis complications based on convolutional survival network includes: a data acquisition module for collecting blood pressure information of hemodialysis patients; A data preprocessing module for missing value processing, normalization and normalization into a two-dimensional matrix; a learning prediction module for deep learning modeling; a result display module for visual output and long-term risk changes.
  • the data preprocessing module preprocesses the hemodialysis blood pressure data, as follows:
  • Blood pressure data recorded during hemodialysis is correlated with adverse clinical events such as cardiovascular and cerebrovascular complications.
  • data with systolic blood pressure less than 60 mmHg or greater than 250 mmHg are screened out, and the systolic blood pressure data of 36 consecutive times (about 3 months) of hemodialysis are screened into the format of a two-dimensional matrix, arranged in the order of hemodialysis treatment, and each row of data corresponds to A hemodialysis treatment process with a duration of 5 hours (less than 5 hours is filled with 0), the interval between columns is 10 minutes (linear interpolation), a total of 36 rows, and each row has 30 points.
  • This two-dimensional matrix was normalized using Min-Max to ensure that the blood pressure waveform was preserved.
  • This two-dimensional matrix is used as the input of the convolutional survival network, and the two-dimensional matrix is shown as the leftmost input matrix in Figure 3.
  • the processing process of the learning prediction module includes two parts: prediction of relative risk of complications based on convolutional survival network; calculation of long-term risk changes based on Breslow method combined with relative risk, as shown in Figure 2, the specific steps are as follows:
  • the present invention uses the architecture of the convolutional neural network and combines the Cox proportional hazards loss function to train the convolutional survival network.
  • the convolutional survival network is composed of multiple convolutional layers and fully connected layers.
  • the input is the above-mentioned 1-channel 36*30 two-dimensional matrix, and the one-dimensional blood pressure pattern is identified through 16 1*5 convolutions, and then 32 A vertical 5*1 convolution operation fills the perceptual field into a square, followed by 2 layers of 16 5*5 convolution kernels, 1 layer of 9 3*3 convolution kernels to abstract feature layers, and then followed by
  • the global mean pooling layer obtains 9 pattern eigenvalues, and the 9 pattern eigenvalues are fully connected and output as a node representing the relative risk of an event (such as the occurrence of cardiovascular complications), that is, ⁇ T x below; the activation function of the last layer is Linear, and the activation function of the remaining layers is Relu.
  • the present invention optimizes network parameters through Cox
  • the key to a convolutional survival network is that it can handle censored data and optimize network parameters by using the Cox proportional hazards loss function.
  • the risk function h(t, x) represents the probability of an event occurring in an individual at a given time.
  • the risk function is shown in formula (1), which represents the probability of an event occurring in patient x at time t:
  • t time
  • x (x 1 , .
  • the 9 pattern eigenvalues obtained after mean pooling, i.e. p takes 9) are used to characterize patient x
  • h 0 (t) is the benchmark risk function
  • ⁇ T x is the relative risk (i.e. the output of the convolutional survival network).
  • the Cox proportional hazards loss function is:
  • n refers to the number of ESRD patients
  • ⁇ i is the indicator of the outcome event E
  • ⁇ i 1 if an event occurs in individual patient x i
  • ⁇ i 0 if no event occurs
  • T i and T j represent individual patient x i , respectively and the survival time of x j .
  • the loss function is calculated by y_pred and y_true.
  • y_pred is the predicted value output by the model
  • y_true represents its true value.
  • the mean squared difference of y_pred and y_true is the mean squared error loss function.
  • equation (2) only ⁇ T x i and ⁇ T x j as the relative risk of model prediction output can be represented by y_pred, and the patient's real risk of complications is unknown.
  • the present invention adopts the following process:
  • y i _pred is the output of the convolutional survival network, representing relative risk, corresponding to ⁇ T x i . Since M is sorted in descending order of T, It is the exponential accumulation sum of the output of the first i rows of the matrix M after passing through the convolutional survival network, that is, the first i rows The cumulative sum of , let the cumulative sum function be recorded as cumsum;
  • ⁇ i is the value of the individual survival outcome event E in the i-th row of matrix M, and can also be recorded as E i ;
  • the Cox proportional hazards loss function uses the model output y_pred and the survival outcome event E to calculate after using the survival time T to reversely arrange the feature data X and the survival outcome event E. , and since the cumulative sum function cumsum needs to be calculated, the loss function can only be calculated in batches.
  • 1 epoch (traversing the data set once) can only update the network parameters once, and the time cost is high.
  • the invention divides the training set into 10 batches randomly and hierarchically, and the ratio of survival outcome events E in each batch is equal, and the survival data of each batch is arranged in reverse order according to the survival time T to calculate the loss function. Therefore, 1 epoch has 10 batches of data used to update network parameters, and the efficiency is significantly improved by traversing the data set to update network parameters 10 times.
  • h(t, x) on the left side of the equal sign is the risk function describing the long-term risk
  • ⁇ T x on the right side of the equal sign the relative risk, which has been calculated by the convolutional survival network; therefore, it is only necessary to estimate h 0 (t ) can be combined with exp( ⁇ T x) to calculate the risk function h(t, x) that describes the long-term risk.
  • the Breslow method proposes the estimated expression of the benchmark cumulative risk function H 0 (t i ) at time t i :
  • R i is the sample set at risk at time t i .
  • the systolic blood pressure recording training model of a patient with end-stage renal disease in a tertiary hospital for 36 consecutive times of hemodialysis is used, and the stratified ten-fold cross-validation is used to evaluate the accuracy of the model (C-Index).
  • C-Index the maximum value, minimum value, mean value, and blood pressure variability of 36 hemodialysis blood pressure records of patients with end-stage renal disease are used for the input of DeepSurv and traditional Cox proportional hazards regression models, and ten-fold cross-validation is performed for comparative evaluation. Model accuracy.
  • the C-Index of the traditional Cox proportional hazards regression model was 0.646 ⁇ 0.065 (0.95CI)
  • the C-Index of the DeepSurv model was 0.658 ⁇ 0.038 (0.95CI)
  • the C-Index of the model of the present invention was 0.839 ⁇ 0.039 (0.95CI). It can be seen that the DeepSurv model is more stable than the traditional Cox proportional hazards regression model and the accuracy C-Index is almost the same, while the model accuracy C-Index of the present invention is obviously better than the former two, and can realize long-term risk prediction of hemodialysis complications.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Physics & Mathematics (AREA)
  • Medical Informatics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Public Health (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Databases & Information Systems (AREA)
  • Pathology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Epidemiology (AREA)
  • Primary Health Care (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • External Artificial Organs (AREA)

Abstract

本发明公开了一种基于卷积生存网络的血透并发症长期风险预测系统，该系统包括数据采集模块、数据预处理模块、学习预测模块和结果展示模块；本发明利用卷积神经网络处理多维度的血透时序特征；卷积神经网络结合Cox比例风险假设，提出卷积生存网络；在利用卷积生存网络的基础上，采用Breslow估计基准风险函数，计算患者的长期风险变化情况。本发明可以充分利用医学研究中常见的截尾数据；应用卷积神经网络的主体架构，便于可视化分析，做出具有可解释性、具有启发性的结果；可以预测患者长期的风险变化情况。

Description

一种基于卷积生存网络的血透并发症长期风险预测系统 技术领域
本发明属于医疗及机器学习技术领域，具体地，涉及一种基于卷积生存网络的血透并发症长期风险预测系统。
背景技术
终末期肾病在世界范围内的发病呈上升趋势，造成巨大疾病负担。大部分患者需要依靠血液透析(血透)维持生命。长期血透过程中可能发生的血管通路感染、高血压、冠心病等并发疾病，严重影响患者生存状况，对患者、患者家庭以及社会都造成了巨大的影响和负担。因此，对血透并发症进行长期的风险预测以及早期的预防治疗，对于提高终末期肾病患者生存质量至关重要。长期血透过程中积累了大量时序数据，为相关研究带来了挑战的同时也带来了大量机会。近年来，随着信息技术的飞速发展，兴起了许多基于机器学习方法的时序分析方法。其中深度学习方法应用广泛且性能强大，包括循环神经网络以及卷积神经网络等。循环神经网络和卷积神经网络已经被广泛应用于医疗领域内的时序分析场景。
循环神经网络是一类以序列数据为输入，在序列的演进方向进行递归且所有节点按链式连接的递归神经网络。卷积神经网络是一类包含卷积计算且具有深度结构的前馈神经网络，具有表征学习能力，能够按其阶层结构对输入信息进行平移不变分类，因此也被称为“平移不变人工神经网络”。然而循环神经网络识别的特征复杂抽象，难以可视化理解、提供启发性结果。卷积神经网络往往仅能预测截止单个时间点的风险值，无法对并发症发生风险进行长期、连续的风险预测，难以为临床医生提供准确有效的决策支持。除此之外，传统的循环神经网络以及卷积神经网络仅能分析简单的数据结构，而无法处理临床分析中常见的截尾数据。
发明内容
深度学习是人工智能领域的热门研究方向，在搜索、机器翻译、自然语言处理、多媒体学习、推荐和个性化技术，以及其他相关领域都取得了很多成果，但目前深度学习还很少应用到血透并发症预后分析领域。
本发明的目的在于针对传统深度学习方法的不足，提出一种基于卷积生存网络的血透并发症长期风险预测系统；本发明利用深度学习领域的卷积神经网络算法，基于其表征学习能 力以及卷积核的可视化方法，得到具有可解释性、具有启发性的结果；本发明改进卷积神经网络结构，使其可以充分利用截尾数据，预测患者的相对风险，进而进行长期、连续的风险预测。
本发明的目的是通过以下技术方案来实现的：一种基于卷积生存网络的血透并发症长期风险预测系统，该系统包括：用于采集血透患者血压信息的数据采集模块；用于对原始数据进行缺失值处理、归一化处理并规整化为二维矩阵的数据预处理模块；用于深度学习建模的学习预测模块；用于可视化输出、呈现长期风险变化情况的结果展示模块；
所述数据预处理模块的处理过程具体为：将连续血透的血压数据规整成二维矩阵，按血透治疗顺序排列，每行数据对应于一次血透治疗过程，使用Min-Max归一化所述二维矩阵以确保血压波形得以保留；
所述学习预测模块的处理过程包括两部分：
(1)基于卷积生存网络的并发症相对风险预测：使用卷积神经网络架构并结合Cox比例风险损失函数训练卷积生存网络；所述卷积生存网络由多个卷积层和全连接层叠加而成，输入为数据预处理模块输出的二维矩阵，各卷积层将特征层层抽象，最终抽象为若干模式特征，各模式特征经全连接层输出为一个节点表示发生事件的相对风险；通过Cox比例风险损失函数优化网络参数；
风险函数h(t，x)表示个体在给定时刻发生事件的概率，公式如下：
h(t，x)＝h 0(t)exp(β Tx)          (1)
其中t表示时间，x＝(x 1，...，x p) T是p维协变量用于表征患者，p为卷积生存网络最后一层全连接层的神经元个数，β＝(β 1，...，β p) T是回归参数，h 0(t)是基准风险函数，β Tx是相对风险，即卷积生存网络的输出；
Cox比例风险损失函数公式如下：
Figure PCTCN2021113135-appb-000001
其中n指患者个数；δ i是结局事件E的指示符，患者个体x i发生事件则δ i＝1，未发生事件则δ i＝0；T i及T j分别表示患者个体x i和x j的生存时间；
(2)基于Breslow法结合相对风险，计算长期风险变化情况，具体地：基于Breslow法得到t i时刻的基准累积风险函数H 0(t i)，根据H 0(t i)得到h 0(t)，H 0(t i)的估计表达式如下：
Figure PCTCN2021113135-appb-000002
其中R i为t i时刻处于风险中的样本集；
根据基准风险函数h 0(t)结合卷积生存网络计算得到的相对风险β Tx：，根据式(1)计算描述 长期风险的风险函数h(t，x)。
进一步地，所述数据预处理模块首先筛除收缩压小于60mmHg或大于250mmHg的数据，之后筛选连续36次血透的收缩压数据规整成二维矩阵，按血透治疗顺序排列，每行数据对应于一次时长5小时的血透治疗过程，列与列之间间隔10分钟，总计36行，每行30个点。
进一步地，所述学习预测模块中，卷积生存网络接收数据预处理模块输出的二维矩阵，首先识别一维血压模式，然后将感知野补成方形，将特征层层抽象后接全局均值池化层获得若干模式特征，各模式特征经全连接层输出为一个节点表示发生事件的相对风险；最后一层激活函数为Linear，其余层激活函数为Relu。
进一步地，所述学习预测模块中，卷积生存网络的输入为数据预处理模块输出的1通道36*30大小的二维矩阵，通过16个1*5的卷积识别一维血压模式，然后通过32个纵向5*1的卷积操作将感知野补成方形，之后接2层16个5*5的卷积核、1层9个3*3的卷积核将特征层层抽象，再之后接全局均值池化层获得9个模式特征值，9个模式特征值经全连接层输出为一个节点表示发生事件的相对风险。
进一步地，所述学习预测模块中，Cox比例风险损失函数的计算流程如下：
(a)将患者的特征数据X和生存结局事件E统一按生存时间T降序排列，形成矩阵M；
(b)y i_pred为卷积生存网络输出，对应于β Tx i；由于M按T降序排列，
Figure PCTCN2021113135-appb-000003
为矩阵M的前i行通过卷积生存网络后输出的指数累加和，即前i行
Figure PCTCN2021113135-appb-000004
的累加和，令累计求和函数记为cumsum；
(c)δ i即矩阵M第i行个体生存结局事件E的取值，记作E i；
(d)令求和函数记为sum，由此损失函数表示为：
loss＝-sum(E i*(y i_pred-log(cumsum(y i_pred))))。
进一步地，所述学习预测模块中，为了使训练集和测试集对应的损失函数取值为同一个量级，将求和函数sum替换为求均值函数mean，损失函数表达为：
loss＝-mean(E i*(y i_pred-log(cumsum(y i_pred))))。
进一步地，所述学习预测模块中，通过Cox比例风险损失函数优化网络参数时，将训练集随机分层划分成10个batch，每个batch中生存结局事件E的比例相等，并将每个batch的生存数据按生存时间T倒序排列用于计算损失函数，由此遍历一遍数据集更新10次网络参数。
本发明的有益效果是：本发明利用卷积神经网络处理多维度的血透时序特征；卷积神经网络结合Cox比例风险假设，提出卷积生存网络；在利用卷积生存网络的基础上，采用Breslow估计基准风险函数，计算患者的长期风险变化情况。本发明可以充分利用医学研究中常见的 截尾数据；应用卷积神经网络的主体架构，便于可视化分析，做出具有可解释性、具有启发性的结果；可以预测患者长期的风险变化情况。
附图说明
图1为本发明基于卷积生存网络的血透并发症长期风险预测系统结构框图；
图2为基于卷积生存网络的血透并发症长期风险预测流程图；
图3为卷积生存网络模型示意图。
具体实施方式
为使本发明的上述目的、特征和优点能够更加明显易懂，下面结合附图对本发明的具体实施方式做详细的说明。
在下面的描述中阐述了很多具体细节以便于充分理解本发明，但是本发明还可以采用其他不同于在此描述的其它方式来实施，本领域技术人员可以在不违背本发明内涵的情况下做类似推广，因此本发明不受下面公开的具体实施例的限制。
本发明中所述的卷积生存网络：应用于生存分析的卷积神经网络，可以处理时序、图像类的数据并进行生存分析与风险预测；长期风险预测：区别于截止一个时间点的风险预测，“长期风险预测”预测较长一段时间内连续的风险变化情况；截尾数据：如果在规定的结束时间没有出现结果事件的数据称之为截尾数据，从起点到截尾的时间称为截尾时间。
如图1所示，本发明提出的一种基于卷积生存网络的血透并发症长期风险预测系统，该系统包括：用于采集血透患者血压信息的数据采集模块；用于对原始数据进行缺失值处理、归一化处理并规整化为二维矩阵的数据预处理模块；用于深度学习建模的学习预测模块；用于可视化输出、呈现长期风险变化情况的结果展示模块。
数据预处理模块对血透血压数据进行预处理，具体如下：
血透过程中记录的血压数据与患者心脑血管并发症等不良临床事件具有相关性。本发明筛除收缩压小于60mmHg或大于250mmHg的数据，筛选连续36次(约3个月)血透的收缩压数据规整成二维矩阵的格式，按血透治疗顺序排列，每行数据对应于一次时长5小时的血液透析治疗过程(不足5小时以0填充)，列与列之间间隔10分钟(线性插值)，总计36行，每行30个点。使用Min-Max归一化该二维矩阵以确保血压波形得以保留。该二维矩阵作为卷积生存网络的输入，二维矩阵见图3中最左侧的输入矩阵。
学习预测模块的处理过程包括两部分：基于卷积生存网络的并发症相对风险预测；基于Breslow法结合相对风险，计算长期风险变化情况，如图2所示，具体步骤如下：
(1)基于卷积生存网络的并发症相对风险预测
本发明使用卷积神经网络的架构并结合Cox比例风险损失函数训练卷积生存网络。卷积生存网络由多个卷积层和全连接层叠加而成，输入为上述1通道36*30大小的二维矩阵，通过16个1*5的卷积识别一维血压模式，然后通过32个纵向5*1的卷积操作将感知野补成方形，之后接2层16个5*5的卷积核、1层9个3*3的卷积核将特征层层抽象，再之后接全局均值池化层获得9个模式特征值，9个模式特征值全连接输出为一个节点表示发生事件(如心血管并发症发生)的相对风险，即下文β Tx；最后一层激活函数为Linear，其余层激活函数为Relu。本发明通过Cox比例风险损失函数优化网络参数，网络输出为相对风险。卷积生存网络模型结构见图3。
卷积生存网络的关键是它可以通过使用Cox比例风险损失函数来处理截尾数据并优化网络参数。风险函数h(t，x)表示个体在给定时刻发生事件的概率，风险函数如式(1)所示，表示患者x在t时刻发生事件的概率：
h(t，x)＝h 0(t)exp(β Tx)          (1)
其中t表示时间，x＝(x 1，...，x p) T是p维协变量，p为卷积生存网络最后一层全连接层的神经元个数(本实施例中对应于全局均值池化后获得的9个模式特征值，即p取9)用于表征患者x，β＝(β 1，...，β p) T是回归参数，h 0(t)是基准风险函数，β Tx是相对风险(即卷积生存网络的输出)。
Cox比例风险损失函数为：
Figure PCTCN2021113135-appb-000005
其中n指ESRD患者个数，δ i是结局事件E的指示符；患者个体x i发生事件则δ i＝1，未发生事件则δ i＝0；T i及T j分别表示患者个体x i和x j的生存时间。式(2)值越小表示参数β(相当于最后一层全连接层的参数)越能拟合患者的相对风险。
通常损失函数由y_pred和y_true计算得到。y_pred为模型输出的预测值，y_true则表示其真值。例如y_pred和y_true的差方和取均值即为均方误差损失函数。但是式(2)中仅β Tx i和β Tx j作为模型预测输出的相对风险能用y_pred表示，而患者真实并发症发生风险未知。为了计算式(2)的损失函数，本发明采用以下流程：
(1.1)将患者的特征数据X(表示所有患者的特征数据)和生存结局事件E(表示所有患者的生存结局)，统一按生存时间T(表示所有患者的生存时间)降序排列，形成矩阵M；
(1.2)y i_pred为卷积生存网络输出，表示相对风险，对应于β Tx i。由于M按T降序排列，
Figure PCTCN2021113135-appb-000006
为矩阵M的前i行通过卷积生存网络后输出的指数累加和，即前i行
Figure PCTCN2021113135-appb-000007
的累加和，令累计求和函数记为cumsum；
(1.3)δ i即矩阵M第i行个体生存结局事件E的取值，又可记作E i；
(1.4)令求和函数记为sum，由此损失函数可表示为：
loss＝-sum(E i*(y i_pred-log(cumsum(y i_pred))))；
(1.5)由于训练集和测试集样本量通常有差异，为了使得二者损失函数取值为同一个量级，本发明将步骤(1.4)的求和改为求均值；令求均值函数记为mean，损失函数可表达为loss＝-mean(E i*(y i_pred-log(cumsum(y i_pred))))。
可以观察到，与普通损失函数使用y_pred和y_true计算不同的是，Cox比例风险损失函数在利用生存时间T倒序排列特征数据X和生存结局事件E后，使用模型输出y_pred和生存结局事件E进行计算，并且由于需要计算累加和函数cumsum，因此只能批量计算损失函数。
使用Cox比例风险损失函数优化网络参数时，通常是使用整个数据集进行计算。但是如此1个epoch(遍历一遍数据集)只能更新一遍网络参数，时间成本较高。本发明将训练集随机分层划分成10个batch，每个batch中生存结局事件E的比例相等，并将每个batch的生存数据按生存时间T倒序排列用于计算损失函数。因此1个epoch有10个batch的数据用于更新网络参数，遍历一遍数据集更新10次网络参数，效率显著提升。
(2)基于Breslow法结合相对风险，计算长期风险变化情况
式(1)中等号左边h(t，x)即描述长期风险的风险函数；等号右边β Tx：即相对风险，已由卷积生存网络计算得到；因此，仅需估算h 0(t)即可结合exp(β Tx)计算描述长期风险的风险函数h(t，x)。
在已知β的情况下估算基准风险函数h 0(t)，最常用的方法即Breslow法。Breslow法提出t i时刻的基准累积风险函数H 0(t i)的估计表达式：
Figure PCTCN2021113135-appb-000008
其中R i为t i时刻处于风险中的样本集。
本实施例使用某三甲医院终末期肾病患者连续36次血透过程中的收缩压记录训练模型，使用分层的十折交叉验证评估模型准确度(C-Index)。本实施例将终末期肾病患者36次血透血压记录求最大值、最小值、均值、血压变异性，用于DeepSurv及传统Cox比例风险回归模型的输入，并且进行十折交叉验证用于对比评估模型准确度。传统Cox比例风险回归模型C-Index为0.646±0.065(0.95CI)，DeepSurv模型C-Index为0.658±0.038(0.95CI)，本发明模型C-Index为0.839±0.039(0.95CI)。可见DeepSurv模型较传统Cox比例风险回 归模型稳定且准确度C-Index相差无几，而本发明模型准确度C-Index明显优于前两者，能够实现血透并发症长期风险预测。
以上所述仅是本发明的优选实施方式，虽然本发明已以较佳实施例披露如上，然而并非用以限定本发明。任何熟悉本领域的技术人员，在不脱离本发明技术方案范围情况下，都可利用上述揭示的方法和技术内容对本发明技术方案做出许多可能的变动和修饰，或修改为等同变化的等效实施例。因此，凡是未脱离本发明技术方案的内容，依据本发明的技术实质对以上实施例所做的任何的简单修改、等同变化及修饰，均仍属于本发明技术方案保护的范围内。

Claims (7)

  1. 一种基于卷积生存网络的血透并发症长期风险预测系统，其特征在于，该系统包括：用于采集血透患者血压信息的数据采集模块；用于对原始数据进行缺失值处理、归一化处理并规整化为二维矩阵的数据预处理模块；用于深度学习建模的学习预测模块；用于可视化输出、呈现长期风险变化情况的结果展示模块；
    所述数据预处理模块的处理过程具体为：将连续血透的血压数据规整成二维矩阵，按血透治疗顺序排列，每行数据对应于一次血透治疗过程，使用Min-Max归一化所述二维矩阵以确保血压波形得以保留；
    所述学习预测模块的处理过程包括两部分：
    (1)基于卷积生存网络的并发症相对风险预测：使用卷积神经网络架构并结合Cox比例风险损失函数训练卷积生存网络；所述卷积生存网络由多个卷积层和全连接层叠加而成，输入为数据预处理模块输出的二维矩阵，各卷积层将特征层层抽象，最终抽象为若干模式特征，各模式特征经全连接层输出为一个节点表示发生事件的相对风险；通过Cox比例风险损失函数优化网络参数；
    风险函数h(t，x)表示个体在给定时刻发生事件的概率，公式如下：
    h(t，x)＝h 0(t)exp(β Tx)  (1)
    其中t表示时间，x＝(x 1，...，x p) T是p维协变量用于表征患者，p为卷积生存网络最后一层全连接层的神经元个数，β＝(β 1，...，β p) T是回归参数，h 0(t)是基准风险函数，β Tx是相对风险，即卷积生存网络的输出；
    Cox比例风险损失函数公式如下：
    Figure PCTCN2021113135-appb-100001
    其中n指患者个数；δ i是结局事件E的指示符，患者个体x i发生事件则δ i＝1，未发生事件则δ i＝0；T i及T j分别表示患者个体x i和x j的生存时间；
    (2)基于Breslow法结合相对风险，计算长期风险变化情况，具体地：基于Breslow法得到t i时刻的基准累积风险函数H 0(t i)，根据H 0(t i)得到h 0(t)，H 0(t i)的估计表达式如下：
    Figure PCTCN2021113135-appb-100002
    其中R i为t i时刻处于风险中的样本集；
    根据基准风险函数h 0(t)结合卷积生存网络计算得到的相对风险β Tx，根据式(1)计算描述 长期风险的风险函数h(t，x)。
  2. 根据权利要求1所述的一种基于卷积生存网络的血透并发症长期风险预测系统，其特征在于，所述数据预处理模块首先筛除收缩压小于60mmHg或大于250mmHg的数据，之后筛选连续36次血透的收缩压数据规整成二维矩阵，按血透治疗顺序排列，每行数据对应于一次时长5小时的血透治疗过程，列与列之间间隔10分钟，总计36行，每行30个点。
  3. 根据权利要求1所述的一种基于卷积生存网络的血透并发症长期风险预测系统，其特征在于，所述学习预测模块中，卷积生存网络接收数据预处理模块输出的二维矩阵，首先识别一维血压模式，然后将感知野补成方形，将特征层层抽象后接全局均值池化层获得若干模式特征，各模式特征经全连接层输出为一个节点表示发生事件的相对风险；最后一层激活函数为Linear，其余层激活函数为Relu。
  4. 根据权利要求2所述的一种基于卷积生存网络的血透并发症长期风险预测系统，其特征在于，所述学习预测模块中，卷积生存网络的输入为数据预处理模块输出的1通道36*30大小的二维矩阵，通过16个1*5的卷积识别一维血压模式，然后通过32个纵向5*1的卷积操作将感知野补成方形，之后接2层16个5*5的卷积核、1层9个3*3的卷积核将特征层层抽象，再之后接全局均值池化层获得9个模式特征值，9个模式特征值经全连接层输出为一个节点表示发生事件的相对风险。
  5. 根据权利要求1所述的一种基于卷积生存网络的血透并发症长期风险预测系统，其特征在于，所述学习预测模块中，Cox比例风险损失函数的计算流程如下：
    (a)将患者的特征数据X和生存结局事件E统一按生存时间T降序排列，形成矩阵M；
    (b)y i_pred为卷积生存网络输出，对应于β Tx i；由于M按T降序排列，
    Figure PCTCN2021113135-appb-100003
    为矩阵M的前i行通过卷积生存网络后输出的指数累加和，即前i行
    Figure PCTCN2021113135-appb-100004
    的累加和，令累计求和函数记为cumsum；
    (c)δ i即矩阵M第i行个体生存结局事件E的取值，记作E i；
    (d)令求和函数记为sum，由此损失函数表示为：
    loss＝-sum(E i*(y i_pred-log(cumsum(y i_pred))))。
  6. 根据权利要求5所述的一种基于卷积生存网络的血透并发症长期风险预测系统，其特征在于，所述学习预测模块中，为了使训练集和测试集对应的损失函数取值为同一个量级，将求和函数sum替换为求均值函数mean，损失函数表达为：
    loss＝-mean(E i*(y i_pred-log(cumsum(y i_pred))))。
  7. 根据权利要求1所述的一种基于卷积生存网络的血透并发症长期风险预测系统，其特 征在于，所述学习预测模块中，通过Cox比例风险损失函数优化网络参数时，将训练集随机分层划分成10个batch，每个batch中生存结局事件E的比例相等，并将每个batch的生存数据按生存时间T倒序排列用于计算损失函数，由此遍历一遍数据集更新10次网络参数。
PCT/CN2021/113135 2021-02-08 2021-08-18 一种基于卷积生存网络的血透并发症长期风险预测系统 Ceased WO2022166158A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110179779.8A CN112530594B (zh) 2021-02-08 2021-02-08 一种基于卷积生存网络的血透并发症长期风险预测系统
CN202110179779.8 2021-02-08

Publications (1)

Publication Number Publication Date
WO2022166158A1 true WO2022166158A1 (zh) 2022-08-11

Family

ID=74975643

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/113135 Ceased WO2022166158A1 (zh) 2021-02-08 2021-08-18 一种基于卷积生存网络的血透并发症长期风险预测系统

Country Status (2)

Country Link
CN (1) CN112530594B (zh)
WO (1) WO2022166158A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117373585A (zh) * 2023-10-31 2024-01-09 华脉汇百通信息技术(北京)有限公司 一种基于人工智能的血液透析模型的构建方法
CN118039183A (zh) * 2024-04-11 2024-05-14 四川互慧软件有限公司 基于循环神经网络预测感染风险的方法、装置及电子设备
CN118711755A (zh) * 2024-08-27 2024-09-27 吉林大学 血液透析中心信息化管理系统及方法

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112530594B (zh) * 2021-02-08 2021-05-11 之江实验室 一种基于卷积生存网络的血透并发症长期风险预测系统
CN114360687A (zh) * 2022-01-19 2022-04-15 滨州医学院烟台附属医院 一种肾内科血透并发症长期风险预测系统及方法
CN114496235B (zh) * 2022-04-18 2022-07-19 浙江大学 一种基于深度强化学习的血透患者干体重辅助调节系统
CN114913982B (zh) 2022-07-18 2022-10-11 之江实验室 基于对比学习的终末期肾病并发症风险预测系统
CN118762796B (zh) * 2024-09-09 2024-12-27 中国人民解放军总医院第二医学中心 基于动态环境感知的智能血液透析辅助护理方法及系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020077352A1 (en) * 2018-10-12 2020-04-16 Human Longevity, Inc. Multi-omic search engine for integrative analysis of cancer genomic and clinical data
CN111582577A (zh) * 2020-05-07 2020-08-25 北京思特奇信息技术股份有限公司 一种电信用户离网预测方法、系统、介质及设备
CN112289455A (zh) * 2020-10-21 2021-01-29 王智 一种人工智能神经网络学习模型构建系统、构建方法
CN112530594A (zh) * 2021-02-08 2021-03-19 之江实验室 一种基于卷积生存网络的血透并发症长期风险预测系统

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11720809B2 (en) * 2019-06-05 2023-08-08 The Ronin Project, Inc. Modeling for complex outcomes using clustering and machine learning algorithms
CN111914516B (zh) * 2020-08-20 2024-03-22 杭州安恒信息技术股份有限公司 一种网络数据预测序列生成方法、装置、设备及存储介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020077352A1 (en) * 2018-10-12 2020-04-16 Human Longevity, Inc. Multi-omic search engine for integrative analysis of cancer genomic and clinical data
CN111582577A (zh) * 2020-05-07 2020-08-25 北京思特奇信息技术股份有限公司 一种电信用户离网预测方法、系统、介质及设备
CN112289455A (zh) * 2020-10-21 2021-01-29 王智 一种人工智能神经网络学习模型构建系统、构建方法
CN112530594A (zh) * 2021-02-08 2021-03-19 之江实验室 一种基于卷积生存网络的血透并发症长期风险预测系统

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
"Chinese Master Thesis Zhejiang University", 15 January 2021, ZHEJIANG UNIVERSITY, CN, article WANG FENG : "Study on Hemodialysis Blood Pressure Patterns Recognition and Prognosis Analysis in Patients with End Stage Renal Disease", pages: 1 - 92, XP055956030 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117373585A (zh) * 2023-10-31 2024-01-09 华脉汇百通信息技术(北京)有限公司 一种基于人工智能的血液透析模型的构建方法
CN118039183A (zh) * 2024-04-11 2024-05-14 四川互慧软件有限公司 基于循环神经网络预测感染风险的方法、装置及电子设备
CN118711755A (zh) * 2024-08-27 2024-09-27 吉林大学 血液透析中心信息化管理系统及方法

Also Published As

Publication number Publication date
CN112530594B (zh) 2021-05-11
CN112530594A (zh) 2021-03-19

Similar Documents

Publication Publication Date Title
CN112530594B (zh) 一种基于卷积生存网络的血透并发症长期风险预测系统
CN111967495B (zh) 一种分类识别模型构建方法
CN111180068A (zh) 一种基于多任务学习模型的慢病预测系统
EP3306500A1 (en) Method for analysing medical treatment data based on deep learning, and intelligent analyser thereof
CN110974214A (zh) 一种基于深度学习的自动心电图分类方法、系统及设备
CN111968741A (zh) 基于深度学习与集成学习的糖尿病并发症高危预警系统
CN113768514B (zh) 基于卷积神经网络与门控循环单元的心律失常分类方法
CN113274031B (zh) 一种基于深度卷积残差网络的心律失常分类方法
Li et al. Fusing deep metric learning with KNN for 12-lead multi-labelled ECG classification
WO2025112186A1 (zh) 状态预测方法、装置、计算机设备和存储介质
CN119722707A (zh) 一种基于Transformer和U-Net的肺部CT图像分割方法
CN113288157A (zh) 基于深度可分离卷积和改进损失函数的心律失常分类方法
Luo et al. The prediction of hypertension based on convolution neural network
CN109119155A (zh) 基于深度学习的icu死亡危险性评估系统
CN119811643A (zh) 基于shap的心血管疾病预测模型的可解释方法
CN114191665A (zh) 机械通气过程中人机异步现象的分类方法和分类装置
CN119014878A (zh) 一种基于注意力机制的cnn-gru的ecg信号分类方法
CN118969269A (zh) 糖尿病风险评估方法、终端设备及存储介质
CN119361141A (zh) 慢性胃炎演化趋势的预测方法及装置
Riyaz et al. Ensemble learning for coronary heart disease prediction
CN114255865B (zh) 一种基于循环神经网络的诊疗项目预测方法
CN114417956A (zh) 一种基于深度学习和可解释机制的预后预测模型建模方法
CN119296789B (zh) 一种可解释的无痛内镜检查低氧风险评估方法及相关装置
Eldem An integrated machine learning framework for the early diagnosis of hypertension disease
Zhang A Highly Accurate Attention-Based Convolutional Neural Network for Classification of Brain Tumors

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21924158

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21924158

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 13/02/2024)

122 Ep: pct application non-entry in european phase

Ref document number: 21924158

Country of ref document: EP

Kind code of ref document: A1