CN114464319B - AMS susceptibility assessment system based on slow feature analysis and deep neural network - Google Patents

AMS susceptibility assessment system based on slow feature analysis and deep neural network Download PDF

Info

Publication number
CN114464319B
CN114464319B CN202210037587.8A CN202210037587A CN114464319B CN 114464319 B CN114464319 B CN 114464319B CN 202210037587 A CN202210037587 A CN 202210037587A CN 114464319 B CN114464319 B CN 114464319B
Authority
CN
China
Prior art keywords
data
ams
susceptibility
sleep quality
stage
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.)
Active
Application number
CN202210037587.8A
Other languages
Chinese (zh)
Other versions
CN114464319A (en
Inventor
史大威
王磊
肖融
陈婧
王军政
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.)
Beijing Institute of Technology BIT
Original Assignee
Beijing Institute of Technology BIT
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 Beijing Institute of Technology BIT filed Critical Beijing Institute of Technology BIT
Priority to CN202210037587.8A priority Critical patent/CN114464319B/en
Publication of CN114464319A publication Critical patent/CN114464319A/en
Application granted granted Critical
Publication of CN114464319B publication Critical patent/CN114464319B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G16H50/50—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for simulation or modelling of medical disorders

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Public Health (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Medical Informatics (AREA)
  • Evolutionary Biology (AREA)
  • Software Systems (AREA)
  • Primary Health Care (AREA)
  • Biophysics (AREA)
  • Databases & Information Systems (AREA)
  • Computing Systems (AREA)
  • Pathology (AREA)
  • Molecular Biology (AREA)
  • Epidemiology (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)

Abstract

The invention discloses a data-driven AMS susceptibility assessment system based on performance monitoring and a deep neural network, which can be applied to the dynamic performance assessment of hypoxia tolerance by combining a monitoring technology with the deep neural network, learn the slow SpO2 characteristics and key information in data acquired in an IHT process, and classify the hypoxia tolerance of a human body; a classifier is constructed by utilizing slow feature analysis and a long-term and short-term memory network, essential features of physiological data are fully mined, and the individual hypoxia tolerance is classified quickly and accurately; the system is characterized in that the idea of performance monitoring and a deep neural network is introduced to carry out individual AMS susceptibility assessment on dynamic physiological data, and meanwhile, a hypoxia tolerance adaptive domain formed by sleep quality data and LLS is fused in the assessment process so as to achieve a multi-angle comprehensive assessment effect; the hypoxia tolerance adaptive domain is constructed by combining sleep quality data and LLS indexes for the first time, so that the subjectivity problem of LLS is solved.

Description

AMS susceptibility assessment system based on slow feature analysis and deep neural network
Technical Field
The invention relates to the field of acute altitude disease susceptibility assessment and prevention, in particular to a data-driven AMS susceptibility assessment system based on performance monitoring and a deep neural network.
Background
Acute altitude disease (AMS), commonly known as altitude sickness, is a syndrome caused by low atmospheric pressure and hypoxia in the plateau area. When people who cannot adapt to high altitude environment suddenly reach altitude, uncomfortable symptoms such as headache, nausea, anorexia, insomnia, fatigue and the like can occur, and serious people can influence normal work and life. The assessment of individual susceptibility by AMS helps to develop preventive strategies in advance, but AMS has difficulty in objective diagnosis. From a physiological perspective, a common diagnostic method for AMS is Lewis Lake Score (LLS) index assessment. This is an index that relies on the individual's self-assessment of the severity of the symptom by evaluating the form of a questionnaire, and is highly subjective. Furthermore, the specificity of the LLS method for diagnosing AMS is low, and the potential reason is that static LLS cannot reflect the whole process information of stress of individuals facing an anoxic environment. From a medical point of view, assessing the susceptibility of AMS by markers or detection kits (patent nos.: 201811637725.6, 201811038384.0), requires sampling of subject blood or DNA and then making a diagnosis, which is medically feasible, but does not have such conditions for most people who live working in plateau areas. With the development of Wearable device technology, a recent review article (s.r. Muza, wearable physiological sensors and real-time algorithms for detection of acid mountain health sites, journal of Applied Physiology 124 (3) (2018) 557-563.) points that continuous measurement of physiological variables by Wearable devices can facilitate AMS susceptibility assessment and diagnosis. Thus, data-driven dynamic process performance assessment based on monitored variables is a very promising approach to predict AMS susceptibility.
In the literature of industrial process control, process performance evaluation methods have proven to be an effective solution for evaluating non-linear systems in operation. Due to the existence of dynamic monitoring data, data-driven performance evaluation methods by dynamic rather than static indicators have been widely studied. Ma et al (y.ma, s.zhao, b.huang, feature extraction of constrained dynamic variables, IEEE Transactions on Industrial information 15 (10) (2019) 567-5645.) developed a novel state transition model to learn constrained dynamic features from observed data with influence probability descriptions. Yin et al (j.yin, x.yan, structural-dynamic space automation for fault detection, industrial and Engineering Chemistry Research 58 (47) (2019) 21614-21624) propose a nonlinear dynamic fault detection model based on Mutual information and stacked sparse self-encoders. The model can not only automatically extract the significant features of the observed data, but also learn the potential information contained in the dynamic process data. In particular, slow Feature Analysis (SFA) is a prominent method to reflect essential features of dynamic process data. Super et al (ang, b. Huang, f. Yang, d. Huang, slow feature analysis for monitoring and diagnosis of Control performance, journal of Process Control 39 (2016) 21-34.) propose SFA-based monitoring method to achieve comprehensive evaluation of overall Process Control performance by extracting Slow features and mining essential features of data deeply. Zhao Chunhui et al (C.ZHao, B.Huang, A full-condition monitoring method for non-reactive dynamic chemical processes with correlation and slow feature analysis, AIChE Journal 64 (5) (2018) 1662-1681.) propose a full-condition monitoring method based on the slow and fast characteristics of SFA synchronous analysis data. Scott et al (D.Scott, C.Shang, B.Huang, D.Huang, A holtistic basic frame for monitoring non-stationary industrial processes, IEEE Transactions on Control Systems technologies (2020) 1-8.) propose a new non-stationary probability SFA algorithm to fully describe the non-stationary and stationary changes of process measurements during normal operation. In addition, the feasibility of the dynamic performance monitoring method for realizing AMS risk assessment is preliminarily verified. Chen Jing et al (patent No. 202010264555.2) proposed dynamic peripheral blood oxygen saturation index (DSI) to evaluate AMS based on continuous measurement data from wearable devices, suggesting that kinetic information of physiological signals in a hypoxic environment helps reflect the AMS susceptibility of individuals. However, how to design a system to quickly extract key information from dynamically monitored physiological data and objectively judge the susceptibility of AMS remains an open question.
Disclosure of Invention
The invention aims to provide a data-driven AMS susceptibility assessment system based on performance monitoring and a deep neural network, which can rapidly and objectively assess AMS susceptibility and is not limited by time and place.
An acute altitude disease AMS susceptibility assessment system comprises a data acquisition and preprocessing module, a hypoxia tolerance adaptation domain construction module and an individual acute altitude disease susceptibility assessment module;
the data acquisition and preprocessing module is used for acquiring the monitoring data of the first stage and the second stage and preprocessing the monitoring data;
wherein, the process of monitoring data acquisition in the first stage is as follows: performing respiratory training of periodic switching of high oxygen and low oxygen environment by multiple subjects, and collecting multiple physiological index data of the subjects on the last day after the training lasts for multiple days, wherein the physiological index comprises SpO2Data;
the process of obtaining the monitoring data in the second stage is as follows: after the respiratory training of the first stage is completed, the subject carries out the sleep test in the high altitude hypoxia environment and obtains the sleep quality data X1:nN represents the number of observation samples of the sleep quality data, each sample comprising a plurality of observation variables;
the hypoxia-tolerant adaptive domain building block is for:
for sleep quality data X1:nPerforming Louis lake scoring test on each observation sample of the subject to obtain a scoring result of the susceptibility of the acute altitude disease AMS;
sleep quality data X based on k-means clustering method1:nCarrying out susceptibility classification on the acute altitude disease AMS to obtain a classification result of the susceptibility of the acute altitude disease AMS;
obtaining mutual information values of a scoring result of a Louis lake scoring test and a classification result of a k-means clustering method, and selecting a group of clustering results with the largest mutual information values as a final hypoxia tolerance adaptive domain of a subject as a label of AMS susceptibility of the subject;
the individual acute altitude disease susceptibility evaluation module is used for constructing and training a long-term memory network and evaluating the hypoxia tolerance of an individual to be evaluated, and specifically comprises the following steps:
the construction process of the training data for training the long-time and short-time memory network comprises the following steps: extraction of SpO of each subject in first-stage monitoring data2The slow characteristic data of the data is spliced with the data of a plurality of physiological indexes of each subject in the last day to obtain training data; the label of the training data is the label of acute altitude disease AMS susceptibility obtained by the hypoxia tolerance adaptive domain building module;
the evaluation process is as follows: acquiring physiological index data of the individual to be evaluated in the last day according to the monitoring data acquisition process in the first stage, inputting the physiological index data into a trained long-time and short-time memory network, and outputting an acute altitude disease AMS susceptibility label of the individual to be evaluated.
Preferably, the physiological index data further includes heart rate and respiratory rate.
Preferably, when the data acquisition and preprocessing module monitors data at the first stage, the periodic switching mode of high oxygen and low oxygen is as follows: as an IHT fragment, lasting 1 hour per day; in each IHT fragment, hyperoxia was introduced for 3 minutes and hypoxia for 5 minutes in each cycle, and the cycles were alternated until 1 hour was reached.
Preferably, in obtaining the monitoring data of the first stage, regarding the data of the low oxygen section, if the average value of the data of the later 200 seconds is lower than the set threshold value by 90% or the data length is less than 250 seconds, the data is regarded as contaminated data, and the data is rejected.
Preferably, the data acquisition and preprocessing module selects the data of the last IHT segment in the physiological index data of the last day, and performs reconstruction processing on the data, specifically: adopting a sliding window with a set length to split and reconstruct the data of each physiological index of the last day, and reconstructing the reconstructed data and the corresponding SpO2The slow feature data of the image are spliced to be used as training data.
Preferably, the hypoxia tolerance adaptive domain construction module further adopts a PCA method to process the sleep quality data X1:nAnd (5) performing dimensionality reduction, and classifying the sleep quality data after dimensionality reduction by adopting a k-means clustering method.
Preferably, the hypoxia tolerance adaptive domain construction module is used for constructing the sleep quality data X based on a k-means clustering method1:nThe process of performing AMS susceptibility classification is:
firstly, randomly selecting k temporary samples from sleep quality data as initial clustering centroids of adaptive domains, wherein each clustering centroid represents an adaptive domain category; calculating the distance from other data samples except the clustering mass center in the sleep quality data to each clustering mass center respectively; for each data sample, determining the adaptive domain class to which the data sample belongs as the class to which the clustering centroid closest to the data sample belongs;
for each adaptive domain category, selecting a new clustering centroid point of the category according to all data samples in the category; then, calculating the distance from other data samples except the clustering centroid to each clustering centroid respectively, and determining the class of the adaptive domain to which the data sample belongs as the class to which the clustering centroid closest to the class belongs; by analogy, after multiple iterations, the adaptive domain type to which each data sample belongs is determined, namely the classification result of the AMS susceptibility.
Preferably, when the iteration of the k-means clustering method is carried out, and the number of the loop iteration reaches the maximum set number or the error reaches the condition of expected minimum error, the iteration is stopped.
Preferably, the sleep quality data includes 10 observation variables, specifically: sleep time, deep sleep time scale, wake up time, number of wakeups, average heart rate, average blood oxygen, morning wake up blood oxygen, body movement time index, and body movement number index.
Preferably, the batch size of the long-time memory network is set to 64, and the number of the units is 100; the optimizer selects Adam.
The invention has the following beneficial effects:
the invention applies the performance monitoring technology and the deep neural network to the dynamic performance evaluation of the hypoxia tolerance, learns the SpO2 slow characteristic and key information in data acquired in the IHT process, and classifies the hypoxia tolerance of a human body.
The system is characterized in that the individual AMS susceptibility assessment is carried out on dynamic physiological data by introducing the ideas of performance monitoring and deep neural network. Meanwhile, the evaluation process integrates a hypoxia tolerance adaptive domain formed by sleep quality data and LLS so as to achieve a multi-angle comprehensive evaluation effect.
The invention combines the sleep quality data and the LLS index to construct the hypoxia tolerance adaptation domain for the first time so as to eliminate the subjectivity problem of LLS.
According to the method, a classifier is constructed by utilizing slow characteristic analysis and a long-term and short-term memory network, essential characteristics of physiological data are fully mined, and the individual hypoxia tolerance is rapidly classified with high accuracy.
The data acquisition of the system only needs the individual to enter the simulated plateau environment, but does not need the individual to actually enter the plateau environment, and the system can be stopped at any time, so that the individual with weak hypoxia tolerance capability is prevented from being unnecessarily injured.
Assessing the individual's hypoxia tolerance in a non-plateau environment can enhance the convenience and flexibility of the actual assessment, can make a prejudgment on the individual's potential risk of entering the plateau and can make corresponding preventive measures.
Drawings
FIG. 1 is an overall architecture diagram of the system of the present invention;
FIG. 2 is a schematic diagram of reconstructed data based on a sliding window in an embodiment of the present invention;
FIG. 3 is a plot of principal component variance versus variance ratio for an embodiment of the present invention;
FIG. 4 is a graph of Gap statistics according to an embodiment of the present invention;
FIG. 5 shows SpO according to an embodiment of the present invention2A slow feature map of;
Detailed Description
The system of the present invention will be described in further detail with reference to the accompanying drawings and the following detailed description, which are provided for implementing the present invention, but the scope of the present invention is not limited to the following embodiments.
As shown in FIG. 1, the AMS susceptibility assessment system of the invention comprises a data acquisition and preprocessing module, a hypoxia tolerance adaptation domain construction module and an individual acute altitude disease susceptibility assessment module.
The device comprises a first data acquisition and preprocessing module.
This example performed 18 subjects with two-stage data acquisition, the first stage being intermittent hypoxia training and the second stage being sleep quality data monitoring simulating a hypoxic environment.
In the first phase, subjects were wearing an oxygen mask and were performing periodic hyperhypoxic inhalation training for a period of 10 days, lasting 1 hour per day, which is an IHT fragment. In each IHT process, high oxygen (35% -38% O) is introduced in each cycle2) Duration of 3 minutes, hypoxia (11% -14%)2) Lasting for 5 minutes, alternately cycling the cycle until 1 hour, each IHT segment having about 7 effective cycles, and continuously monitoring human body dynamic physiological index data including heart rate, respiratory rate and SpO during training2The sampling period is 1 second.
Is provided with
Figure BDA0003468612350000051
Sequence data for the mth fragment in IHT,
Figure BDA0003468612350000052
each represents SpO2Heart rate, respiratory rate, and oxygen concentration.
Figure BDA0003468612350000053
Representing the raw data of the ith index at the kth time point. Physiological data (mainly SpO) collected in the present invention2And heart rate), there will be some outlier points kαSo that
Figure BDA0003468612350000054
The invention adopts a linear interpolation technology to replace outlier point data. The invention focuses more on low oxygen segment data, in case of low oxygen, spO2The value will fall first and then slowly plateau. However, some subjects will remove the mask during the period of hypoxia to relieve the discomfort of hypoxia and inhale the oxygen of normal concentration in the air, resulting in SpO2The phenomenon of rising after a short decline in the hypoxic segment. Typically, one hypoxic segment is 300 seconds, and the invention provides that if the average of the data in the last 200 seconds is below the threshold of 90% or the length of the data is less than 250 seconds, the data is considered contaminated and then rejected.
In the second stage, 18 subjects exposed to a simulated high altitude hypoxic environment (i.e., hypoxic chamber) for overnight sleep after performing stage 1 training, the monitoring device monitors their sleep quality data without interfering with the subjects' sleep, the sleep quality data including 10 observed variables, specifically: sleep time, deep sleep time and proportion, wake time and times, average heart rate, average blood oxygen, morning wake blood oxygen, body movement time index and times index.
The sleep quality data for stage 2 is defined as:
Figure BDA0003468612350000061
wherein x is11…x1pP observation variables representing a first observation sample; there were a total of n observation samples.
After the subjects awaken the next morning, they were subjected to the louisis lake LLS test, which was conducted by assigning an AMS susceptibility score to each observed sample of subjects based on expert experience.
The present invention reconstructs data by shifting a time window in consideration of differences between training time points in one IHT fragment obtained by an individual in the first stage. Since the last day of phase 1 the physiological data monitored by training is closest to the physiological state of phase 2, the data of the last IHT fragment M of phase 1 is selected as the reconstructed object. As shown in fig. 2, the sequence data is split into multiple segments, and assuming h segments, the reconstructed data can be expressed as:
Figure BDA0003468612350000062
wherein
Figure BDA0003468612350000063
Indicating the low-oxygen segment data of the jth index in the ith sliding window, wherein i belongs to 1, 2., h; j ∈ 1,2,3,4. The window width of the time-sliding window was set to 128 seconds, and the interval of each sliding was 64 seconds. Since the window width may be larger than the data segment length when the sliding window slides to the last segment of data, the data segment in which this condition exists is discarded.
Second, a hypoxia tolerance adaptive domain building block. Different people have different physical stress responses to the hypoxic environment in the plateau, and particularly reach the plateau quickly in a short time. The present invention therefore proposes the concept of a hypoxia-tolerant adaptive domain. The data collected in the stage 2 is used for monitoring human physiological sleep indexes under the condition of simulating hypoxia environment in a high altitude area, 10 dimensions are still available except the Louis lake score, the PCA method is firstly used for reducing the dimensions, then the k-means clustering method is used for identifying hypoxia tolerance, a hypoxia tolerance adaptive domain is preliminarily constructed, finally mutual information values between the clustering and the Louis lake score are calculated, and a group of clustering results with the largest mutual information values are taken as a final hypoxia tolerance adaptive domain.
Sleep quality data X for stage 21：nAfter normalization of zero mean unit variance is carried out, key features in data are extracted by using a PCA method, and PCA maximization is achieved through capturing in sequenceThe variance of the principal component. Thus, the first principal component contains the largest variance, and subsequent principal components are successively decremented. For sleep data X1：nSingular value decomposition is carried out to obtain:
X1：n＝UDVT，Z＝UD，
U＝[u1，u2，...，up]is an eigenvector matrix, corresponding to the eigenvalue matrix D = diag (e)1，e2，...，ep) The diagonal elements of (a). V = [ V ]1，v2，...，vp]Is a load matrix, Z = [ Z ]1，z2，...，zp]Is a scoring matrix.
To ensure that the sleep data is as undistorted as possible, the principal component values with the first d larger variances are selected. The variance value of the ith principal component is:
Figure BDA0003468612350000064
the cumulative percentage of explained variance (CPV) is used to determine the number of principal components, d, defined as:
Figure BDA0003468612350000071
where the threshold m is an acceptable percentage of the explained variance. Further, the sleep data is converted into:
Figure BDA0003468612350000072
the dimensionality is effectively reduced to d principal component dimensions.
After the characteristics are selected, X is clustered by using a k-means clustering methodPCClustering into k categories, respectively representing k different adaptation domains of individual hypoxia tolerance, specifically:
firstly, the main component data X of sleep qualityPCRandomly selecting k temporary samples as initial clustering centroid of adaptation domain
Figure BDA0003468612350000073
Each cluster centroid represents an initial adaptation domain class. Calculating XPCThe distances from the other n-k data samples except the clustering centroid to each clustering centroid are respectively expressed as follows:
Figure BDA0003468612350000074
wherein x isi∈XPCData sample xiWill fall into the adapted domain class to which the cluster centroid closest to it belongs. x is a radical of a fluorine atomiClass of belonging
Figure BDA0003468612350000075
Is defined as:
Figure BDA0003468612350000076
for each hypoxia tolerant domain class
Figure BDA0003468612350000077
Centroid points for all samples in a class
Figure BDA0003468612350000078
And selecting the new cluster centroid point of the category. So the cluster centroid point after e-round iterative update
Figure BDA0003468612350000079
Is defined as follows:
Figure BDA00034686123500000710
x after e-round iterative updateiCluster category to which it belongs
Figure BDA00034686123500000711
Is defined as:
Figure BDA00034686123500000712
and repeating the operation of updating the classification category and the clustering centroid point until the number of loop iterations reaches the maximum set number or the error reaches the condition of expected minimum error.
Among them, the selection of the value of k is a critical step, which has a great influence on the results of the k-means method. Commonly used methods include the elbow method, which requires the location of the elbow point to be determined, but is sometimes unclear, and gap statistics. Therefore, we use gap statistics to determine the k value.
Since both the hypoxia tolerance adaptive domain and the lewis lake score represent to some extent the risk of susceptibility to AMS, we assumed that there was a correlation between them. Therefore, mutual information values between the clustering results and the LLS are calculated, which can strictly quantify the degree of uncertainty reduction of the LLS due to the known values of the hypoxia tolerance adaptive domain. And selecting a group of clustering results with the maximum mutual information value as a final hypoxia tolerance adaptive domain of the subject, and taking the clustering results as a label of the individual hypoxia tolerance capacity, and recording the label as Y.
Thirdly, evaluating the susceptibility of the individual acute altitude diseases. The evaluation method based on the dynamic physiological index data is an effective method for avoiding subjectivity of AMS evaluation. The invention is based on a slow feature analysis and long-and-short time memory network method, and aims to evaluate the AMS susceptibility of individuals by using dynamic monitoring data.
There is an invariant factor due to the changing signal to reflect its inherent properties. The invention adopts SFA method to extract SpO in the stage 12The slow characteristic of the data is specifically as follows:
the goal of the SFA is at the input SpO2Sequence data
Figure BDA0003468612350000081
And outputting the attribute signal
Figure BDA0003468612350000082
Find effective mapping between
Figure BDA0003468612350000083
Make the characteristic
Figure BDA0003468612350000084
The change is the slowest (i.e., slow characteristic). SpO2Is defined as the time average of the squares of the first order differences, i.e.
Figure BDA0003468612350000085
Wherein
Figure BDA0003468612350000086
Is that
Figure BDA0003468612350000087
First order differential with time point k, defined as
Figure BDA0003468612350000088
<·>tRepresenting the time average of the signal data. SpO2The slow feature problem can be described as an optimization problem with the goal of minimizing SpO2Is characterized by
Figure BDA0003468612350000089
The changes, namely:
Figure BDA00034686123500000810
is constrained by:
zero mean value:
Figure BDA00034686123500000811
unit variance:
Figure BDA00034686123500000812
and (3) decorrelation:
Figure BDA00034686123500000813
slow feature analysis mainly extracting SpO2Medium slow varying characteristics. Linear models of slowly varying features can be viewed as linear combinations of the original variables, i.e.
Figure BDA00034686123500000814
Where W is the weight matrix. The slow eigen-analysis solution problem can be viewed as a generalized eigenvalue solution problem. In analyzing SpO2The slow features of the data are preceded by a centering process to satisfy a zero mean constraint. The optimization problem is rewritten as:
Figure BDA00034686123500000815
meanwhile, the unit variance can also be rewritten as follows:
Figure BDA00034686123500000816
wherein, CoAnd CderiEach represents SpO2Data of
Figure BDA00034686123500000817
And their differential values
Figure BDA00034686123500000818
Time average of the covariance matrix of (a). Furthermore, the above optimization problem is solved by solving the generalized eigenvalues of the following equations:
CderiW＝CoWD
where D is a diagonal matrix of generalized eigenvalues. And after the weight transformation matrix is solved by the generalized eigenvalue, the decorrelation constraint is automatically satisfied. Eigenvalues λ in an eigenvalue matrixiAnd sorting the materials in the order from small to large. Slower features with smaller indices vary more slowly over time, i.e. slow features
Figure BDA0003468612350000091
According to the timeThe slowest variation between the phases, followed by the slow feature
Figure BDA0003468612350000092
And so on. The function with the slowest change speed is finally selected as the slow characteristic and redefined as
Figure BDA0003468612350000093
Data collected by the human body monitoring system are sequence data, the problem of long-term dependence on characteristics exists, and not only is a relationship existing between short-term spacing data, but also a potential relationship may exist between data points with long-term spacing. As the obvious advantage of the long-time and short-time memory network is the practicability of the long sequence, the invention introduces the long-time and short-time memory network to extract the characteristics from the physiological data. The long-short time memory network consists of a plurality of long-short time memory modules in the chain, and the modules can selectively remember and forget the hidden characteristics of the physiological signals in a plurality of time intervals by designing a plurality of gate signals with different functions. Input data H of long-time and short-time memory networkM+Monitoring data H obtained for phase 1MSpO re-splicing2Slow characteristic of
Figure BDA0003468612350000094
Obtained, defined as:
Figure BDA0003468612350000095
hypoxia tolerant adaptive domain Y as input data HM+The label of (a), wherein the label corresponding to each sample of input data of each subject is the label of the subject; will input data HM+And the corresponding labels are input into the long-time and short-time memory network to train the long-time and short-time memory network.
The core of the long-time memory network is a door module comprising a forgetting door ftAnd an input gate itAnd an output gate ot. The forgetting gate is an implicit state input that selectively forgets the physiological signal that was converted from the previous time node. It doesCell state c at a given timet-1C reserved to the current timet. Forget door ftIs defined as:
ft＝σ(Wf·[ht-1，HM+]+bf)，
where σ is sigmoid function, WfAnd bfAre the weight parameter and the deviation parameter of the forgetting gate. HtAnd ht-1Respectively, refers to the hidden state of the physiological data at the current time t and the previous time t-1. HtIs defined as follows:
Figure BDA0003468612350000096
the input gate selectively memorizes the physiological signals input at the current stage, i.e. determines the input H at the current momentM+In cell state ctHow much can be retained. The input gate is defined as:
it＝σ(Wi·[ht-1，HM+]+bi)，
wherein WiAnd biAre the weight parameter and the bias parameter of the input gate.
The output gate determines which current cell state is to be considered the current state output. The formula is defined as:
ot＝σ(Wo.[ht-1，HM+]+bo)，
wherein WoAnd boAre the weight parameter and the bias parameter of the output gate. Cell state ctAnd ct-1Respectively representing the state of the cell at the current time t and the previous time t-1. CtThe formula of (c) is defined as:
Figure BDA0003468612350000103
wherein
Figure BDA0003468612350000104
Hadamard product defined as a matrix。WcAnd bcAre the weight parameter and the bias parameter of the cell state.
In the invention, the batch size of the long-time memory network is set to be 64, and the number of units is 100. The model selects the Adam optimizer. In order to avoid overfitting in the data training process, a regularization operation is performed by utilizing dropout, and some neurons are discarded in a neural network. The extracted features are sent to a softmax layer for individual hypoxia tolerance classification.
And for the individual to be evaluated, acquiring reconstructed data according to a first-stage mode, inputting the reconstructed data into the trained long-time memory network, and outputting a susceptibility label of the individual to be evaluated.
The embodiment is as follows:
the invention also carries out performance verification analysis and evaluation on the method. The PCA method is used to highlight key features of the sleep data. As shown in fig. 3, the histogram shows the variances corresponding to each principal component, arranged in order of magnitude. The first few principal components have large variances indicating that they contain the main information of the data. In particular, the variance of the first principal component is about half of the sum of all variances. The CPV is used as the selection basis of the PCA main component number. The CPV threshold is set to 90%. Thus, d =6 was chosen as the number of principal components, and the data with stage 2 was mapped to data with 6 key feature dimensions. After the data dimensionality is reduced, clustering is performed using k-means to determine a preliminary hypoxia tolerance adaptive domain. The k value of the cluster was determined using Gap statistics, as shown in fig. 4, when k =2, the Gap value takes the maximum value. Thus, hypoxia-adaptive domains are divided into two categories, strong hypoxia tolerance and weak hypoxia tolerance. And finally, obtaining mutual information values between the clustering results and the Louis lake scores, and selecting a group of clustering results with the largest mutual information values between the clustering results and the Louis lake scores as the hypoxia tolerance adaptive domain as shown in the table 1.
TABLE 1 mutual information of clustering results and Louis lake scores
Figure BDA0003468612350000101
For stage 1 monitoring data, spO is first extracted2As shown in fig. 5, the solid line represents SpO2Raw data of (3), dotted line represents SpO2The slow signature of (2). The method provided by the invention is used for classifying the strength of the individual hypoxia tolerance based on the combination of the long-time memory network and the short-time memory network and the slow characteristic analysis, and the confusion matrix of the classification is shown in table 2.
TABLE 2 confusion matrix results for hypoxia tolerance classification
Figure BDA0003468612350000102
Figure BDA0003468612350000111
The present invention adopts six quantitative indicators to quantitatively evaluate the performance of classification, including Accuracy (ACC), precision (PRE), sensitivity (SEN), specificity (SPE), F1 score and Markov Correlation Coefficient (MCC), and the results are shown in table 3.
TABLE 3 quantification of hypoxia tolerance classification
Figure BDA0003468612350000112
In addition, whether the slow characteristic is effective to the classification result is analyzed by using the Area (AUC) under the Receiver Operating Characteristic (ROC) curve. When the input value of LSTM comprises SpO2The AUC of (a) was 0.925, which is higher than the AUC =0.866 obtained for the model without the slow feature. The results reflect that the slowness characteristic of the combined physiological signals in the model can well reflect the most basic characteristics of the signals, thereby improving the classification performance of hypoxia tolerance.

Claims (10)

1. The AMS susceptibility assessment system is characterized by comprising a data acquisition and pretreatment module, a hypoxia tolerance adaptation domain construction module and an individual AMS susceptibility assessment module;
the data acquisition and preprocessing module is used for acquiring the monitoring data of the first stage and the second stage and preprocessing the monitoring data;
wherein, the process of monitoring data acquisition in the first stage is as follows: performing respiratory training of periodic switching of high oxygen and low oxygen environment by multiple subjects, and collecting multiple physiological index data of the subjects on the last day after the training lasts for multiple days, wherein the physiological index comprises SpO2Data;
the process of obtaining the monitoring data in the second stage is as follows: after the respiratory training of the first stage is completed, the subject carries out the sleep test in the high altitude hypoxia environment and obtains the sleep quality data X1:nN represents the number of observation samples of the sleep quality data, and each sample comprises a plurality of observation variables;
the hypoxia-tolerant adaptive domain building block is for:
for sleep quality data X1:nPerforming Louis lake scoring test on each observation sample of the subject to obtain a scoring result of the susceptibility of the acute altitude disease AMS;
sleep quality data X based on k-means clustering method1:nCarrying out susceptibility classification on the acute altitude disease AMS to obtain a classification result of the susceptibility of the acute altitude disease AMS;
obtaining mutual information values of a scoring result of the Louis lake scoring test and a classification result of a k-means clustering method, and selecting a group of clustering results with the largest mutual information values as a final hypoxia tolerance adaptive domain of the subject as a label for AMS susceptibility of the subject;
the individual acute altitude disease susceptibility evaluation module is used for constructing and training a long-term memory network and evaluating the hypoxia tolerance of an individual to be evaluated, and specifically comprises the following steps:
the construction process of the training data for training the long-time memory network comprises the following steps: extraction of SpO of each subject in first-stage monitoring data2Slow characterization of the data, which was compared to multiple physiology of each subject on the last daySplicing the data of the indexes to obtain training data; the label of the training data is the label of acute altitude disease AMS susceptibility obtained by the hypoxia tolerance adaptive domain building module;
the evaluation process comprises the following steps: acquiring physiological index data of the individual to be evaluated in the last day according to the monitoring data acquisition process in the first stage, inputting the physiological index data into a trained long-time memory network, and outputting an acute altitude disease AMS susceptibility label of the individual to be evaluated.
2. The system of claim 1, wherein the physiological index data further comprises heart rate and respiratory rate.
3. The AMS susceptibility assessment system for acute altitude sickness of claim 1, wherein the data collection and preprocessing module is used for periodically switching hyperoxia and hypooxia during the monitoring data of the first stage: as an IHT fragment, lasting 1 hour per day; in each IHT fragment, hyperoxia was introduced for 3 min and hypooxia for 5 min at each cycle, and the cycles were alternated until 1 hour was reached.
4. The AMS susceptibility assessment system for acute altitude diseases of claim 3, wherein in obtaining the first stage monitoring data, regarding the low oxygen stage data, if the average value of the data in the later 200 seconds is lower than the set threshold value 90% or the data length is less than 250 seconds, the data is considered as contaminated data, and the data is removed.
5. The AMS susceptibility assessment system for acute altitude diseases of claim 3, wherein the data collection and pre-processing module selects the data of the last IHT fragment in the physiological index data of the last day and reconstructs the data, specifically: adopting a sliding window with a set length, splitting and reconstructing the data of each physiological index of the last day, and reconstructing the reconstructed data and the corresponding SpO2And splicing the slow characteristic data to be used as training data.
6. The AMS susceptibility assessment system for acute altitude diseases of claim 1, wherein said hypoxia tolerance adaptive domain construction module further adopts PCA method to evaluate sleep quality data X1:nAnd (5) performing dimensionality reduction, and classifying the sleep quality data after dimensionality reduction by adopting a k-means clustering method.
7. The AMS susceptibility assessment system for acute altitude diseases of claim 1, wherein said hypoxia tolerance adaptive domain construction module is based on k-means clustering method to sleep quality data X1:nThe process of performing AMS susceptibility classification is:
firstly, randomly selecting k temporary samples from sleep quality data as initial clustering centroids of adaptive domains, wherein each clustering centroid represents an adaptive domain category; calculating the distances from other data samples except the clustering mass center in the sleep quality data to all the clustering mass centers respectively; for each data sample, determining the adaptive domain class to which the data sample belongs as the class to which the clustering centroid closest to the data sample belongs;
for each adaptive domain category, selecting a new clustering centroid point of the category according to all data samples in the category; then, calculating the distance from other data samples except the clustering centroid to each clustering centroid respectively, and determining the class of the adaptive domain to which the data sample belongs as the class to which the clustering centroid closest to the class belongs; by analogy, after multiple iterations, the adaptive domain type to which each data sample belongs is determined, namely the classification result of the AMS susceptibility.
8. The AMS susceptibility assessment system for acute altitude diseases of claim 7, wherein when the iteration of k-means clustering method is performed, the iteration is stopped when the number of loop iteration reaches the maximum setting number or the error reaches the condition of expected minimum error.
9. The system for assessing AMS susceptibility to acute altitude diseases as claimed in claim 1, wherein the sleep quality data comprises 10 observed variables, in particular: sleep time, deep sleep time scale, wake up time, number of wakeups, average heart rate, average blood oxygen, morning wake up blood oxygen, body movement time index, and body movement number index.
10. The AMS susceptibility assessment system for acute altitude diseases of claim 1, wherein the batch size of said long and short term memory network is set to 64, the number of cells is 100; the optimizer selects Adam.
CN202210037587.8A 2022-01-13 2022-01-13 AMS susceptibility assessment system based on slow feature analysis and deep neural network Active CN114464319B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210037587.8A CN114464319B (en) 2022-01-13 2022-01-13 AMS susceptibility assessment system based on slow feature analysis and deep neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210037587.8A CN114464319B (en) 2022-01-13 2022-01-13 AMS susceptibility assessment system based on slow feature analysis and deep neural network

Publications (2)

Publication Number Publication Date
CN114464319A CN114464319A (en) 2022-05-10
CN114464319B true CN114464319B (en) 2022-11-01

Family

ID=81409446

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210037587.8A Active CN114464319B (en) 2022-01-13 2022-01-13 AMS susceptibility assessment system based on slow feature analysis and deep neural network

Country Status (1)

Country Link
CN (1) CN114464319B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117059270A (en) * 2023-08-14 2023-11-14 北京理工大学 Acute altitude disease risk assessment system combining medical priori knowledge pseudo tags
CN117079813A (en) * 2023-08-17 2023-11-17 北京理工大学 Priori knowledge-based small sample acute altitude disease risk assessment system

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104997516A (en) * 2015-06-18 2015-10-28 中国人民解放军第三军医大学第一附属医院 Plateau work capability assessment and early-warning index system
CN105212945A (en) * 2015-10-15 2016-01-06 成都艾克尔特医疗科技有限公司 A kind of vital signs bracelet that uses is for the health analysis method of altitude sickness
US10561863B1 (en) * 2012-04-06 2020-02-18 Orbital Research Inc. Biometric and environmental monitoring and control system
CN111093476A (en) * 2017-09-21 2020-05-01 京瓷株式会社 Monitoring device, monitoring method, and monitoring program
CN111513726A (en) * 2020-04-07 2020-08-11 北京理工大学 System for evaluating AMS risk based on IHT dynamic performance
CN113361665A (en) * 2021-08-11 2021-09-07 成都理工大学 Highland mountain tourism safety risk early warning method based on reinforcement learning
CN113611414A (en) * 2021-06-22 2021-11-05 北京理工大学 Dynamic monitoring and performance evaluation method and system for acute altitude disease
CN113851224A (en) * 2021-06-22 2021-12-28 北京理工大学 Digital simulator for AMS risk assessment

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2019113562A1 (en) * 2017-12-08 2019-06-13 The United States Of America As Represented By The Secretary Of The Army Systems and methods for monitoring subjects in potential hypoxic distress

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10561863B1 (en) * 2012-04-06 2020-02-18 Orbital Research Inc. Biometric and environmental monitoring and control system
CN104997516A (en) * 2015-06-18 2015-10-28 中国人民解放军第三军医大学第一附属医院 Plateau work capability assessment and early-warning index system
CN105212945A (en) * 2015-10-15 2016-01-06 成都艾克尔特医疗科技有限公司 A kind of vital signs bracelet that uses is for the health analysis method of altitude sickness
CN111093476A (en) * 2017-09-21 2020-05-01 京瓷株式会社 Monitoring device, monitoring method, and monitoring program
CN111513726A (en) * 2020-04-07 2020-08-11 北京理工大学 System for evaluating AMS risk based on IHT dynamic performance
CN113611414A (en) * 2021-06-22 2021-11-05 北京理工大学 Dynamic monitoring and performance evaluation method and system for acute altitude disease
CN113851224A (en) * 2021-06-22 2021-12-28 北京理工大学 Digital simulator for AMS risk assessment
CN113361665A (en) * 2021-08-11 2021-09-07 成都理工大学 Highland mountain tourism safety risk early warning method based on reinforcement learning

Non-Patent Citations (6)

* Cited by examiner, † Cited by third party
Title
IoT-enabled Intelligent Dynamic Risk Assessment of Acute Mountain Sickness Based on Data from Wearable Devices;Jing Chen 等;《2021 4th IEEE International Conference on Industrial Cyber-Physical Systems》;20210705;第132-137页 *
IoT-Enabled Intelligent Dynamic Risk Assessment of Acute Mountain Sickness：The Role of Event-Triggered Signal Processing;Jing Chen 等;《IEEE Transactions on Industrial Informatics》;20211213;第1-9页 *
Prediction of the development and susceptibility to acute mountain sickness (AMS) by monitoring oxygen saturation (SpO2) - literature review;Karol Mazur 等;《Journal of Education，Health and Sport[online]》;20200212;第10卷(第2期);第79-84页 *
The Use of Pulse Oximetry in the Assessment of Acclimatization to High Altitude;Tobias Dünnwald等;《Sensors》;20210122;第21卷(第4期);第1-27页 *
Unveiling Interpretable Key Performance Indicators in Hypoxic Response：A System Identification Approach;Jing Chen 等;《IEEE Transactions on Industrial Electronics》;20211229;第69卷(第12期);第13676-13685页 *
急进高原人群急性高原病与高原头痛的个体易感性研究;申阳;《中国优秀硕士学位论文全文数据库 医药卫生科技辑》;20210115;第2021年卷(第01期);第E065-420页 *

Also Published As

Publication number Publication date
CN114464319A (en) 2022-05-10

Similar Documents

Publication Publication Date Title
CN109009017B (en) Intelligent health monitoring system and data processing method thereof
CN111261282A (en) Sepsis early prediction method based on machine learning
CN114464319B (en) AMS susceptibility assessment system based on slow feature analysis and deep neural network
EP3614301A1 (en) Artificial intelligence-based interference recognition method for electrocardiogram
CN108492877B (en) Cardiovascular disease auxiliary prediction method based on DS evidence theory
CN113040711A (en) Cerebral stroke attack risk prediction system, equipment and storage medium
CN111081379B (en) Disease probability decision method and system thereof
Chen et al. A new deep learning framework based on blood pressure range constraint for continuous cuffless BP estimation
CN114512239B (en) Cerebral apoplexy risk prediction method and system based on transfer learning
CN111951975A (en) Sepsis early warning method based on deep learning model GPT-2
CN110491506A (en) Auricular fibrillation prediction model and its forecasting system
CN111261278A (en) Heart disease detection method based on deep learning model of three-dimensional image
CN113593708A (en) Sepsis prognosis prediction method based on integrated learning algorithm
Zhang et al. Novel covid-19 screening using cough recordings of a mobile patient monitoring system
Herath et al. Autism spectrum disorder diagnosis support model using Inception V3
CN113796873A (en) Wearable dynamic electrocardiosignal classification method and system
Hao et al. An automated digital biomarker of mobility
CN117116475A (en) Method, system, terminal and storage medium for predicting risk of ischemic cerebral apoplexy
CN114521900B (en) Arrhythmia classification and identification method based on transfer learning
Prabha et al. A Novel Analysis and Detection of Autism Spectrum Disorder in Artificial Intelligence Using Hybrid Machine Learning
CN114464317A (en) Sleep posture classification method based on pseudo item recognition fusion and HMM algorithm
REKHA et al. Brain Stroke Prediction Using Random Forest And Adaboost Algorithm
CN111513702A (en) Method for detecting electrocardiographic data and storage medium
Song et al. [Retracted] Health Promotion Effects of Sports Training Based on HMM Theory and Big Data
CN116369941B (en) Sleep quality judging method based on EEG physiological information

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant