CN112382388A - Early warning method for adverse pressure sore event - Google Patents

Abstract

The invention discloses a method for early warning pressure sore adverse events, which comprises the steps of firstly obtaining early warning indexes of the pressure sore adverse events, and constructing an early warning index system of the pressure sore adverse events (particularly, carrying out characteristic information mining on clinical real pressure sore adverse events by using a mixed research method combining half-structure interview, association rules and Delphi expert function inquiry, and constructing the early warning index system of the pressure sore adverse events); secondly, performing feature extraction on the constructed early warning index system of the adverse pressure sore event (particularly performing feature extraction on an ultra-short text vector feature extraction method based on a random up-down sampling text convolution neural network RS-CNN); and finally, establishing a prediction model by using an algorithm of a support vector machine, and outputting an early warning result. The invention provides an early warning method for adverse pressure sore events, which has the advantages of objectivity, effectiveness and practicability.

Description

Early warning method for adverse pressure sore event

Technical Field

The invention relates to the technical field of medical treatment, in particular to a method for early warning adverse pressure sore events.

Background

The comprehensive report of the pressure sore adverse events is an effective means for promoting the management of the pressure sore adverse events, and the most critical content for effectively managing the pressure injuries is that the risk factors of the pressure injuries existing in a patient are fundamentally removed by collecting the occurred pressure injury adverse events for analysis and rule discovery. However, the phenomenon of reporting under and missing of adverse pressure sore events is serious at present, and active reporting is obstructed. The occurrence rule of adverse pressure sore events is not easy to understand comprehensively, the system structure is changed, and the benign outcome of adverse pressure sore events is promoted.

Under the background that artificial intelligence is greatly leap, medical information construction enters a big data era. In recent years, the construction of medical big data platforms of single centers in China is gradually developed, ideas, methods and technologies for big data analysis are vigorously developed and practiced, the big data analysis is mature day by day, and methods and data guarantees are provided for the development and research of medical big data in early warning of pressure injury. A large amount of real data of pressure sore patients are reserved in a hospital information system, and the data comprise treatment, nursing, examination, medication, outcome and the like, and have big data characteristics and mining values. By means of multiple data mining technologies such as artificial intelligence, machine learning and neural networks, pressure sore cases which have already occurred in clinic or related risk factors of pressure sore occurrence can be identified from real, comprehensive and dynamic data of patients existing in a medical information system, and a feasible model is provided for predicting the occurrence of pressure injury.

With the cross fusion of multiple disciplines, researchers combine information technology with pressure injury early warning to improve the risk identification rate to a certain extent, but the researches integrate the pressure injury risk assessment table into the information system construction and assist quality monitoring and dynamic monitoring to realize the collection and enhanced management of pressure injury occurrence conditions, so that the problems of document recording and data archiving are reduced, and the phenomenon of re-reporting and missing reporting caused by bed moving or transferring is reduced.

In addition, most of data collected by the research of the scales and systems in the prior art are numerical data, only the numerical data are processed, and in the big data era, the data volume is large, the processing is inconvenient, and the characteristic processing of the text data is not considered in the data processing.

Therefore, it is of great significance to develop an early warning method for adverse pressure sore events which is objective, effective and practical.

Disclosure of Invention

The invention provides an early warning method for adverse pressure sore events, which comprises the following specific technical scheme:

a method for early warning of adverse pressure sore events comprises the following steps:

step one, acquiring early warning indexes of pressure sore adverse events, and constructing an early warning index system of the pressure sore adverse events;

step two, extracting the characteristics of the early warning index system of the pressure sore adverse events constructed in the step one;

and step three, establishing a prediction model by using an algorithm of a support vector machine, and outputting an early warning result.

Preferably, in the first step, a mixed research method combining a semi-structure interview, an association rule and a Delphi expert function inquiry is used for mining the characteristic information of clinical real pressure sore adverse events, and an early warning index system of the pressure sore adverse events is constructed.

Preferably, the step one specifically includes the steps of:

step 1.1, performing semi-structural interview under the guidance of a crisis management annular model, and extracting early warning indexes of adverse pressure sore events from professional staff relevant to the clinical adverse pressure sore events by collecting the processing experiences and feelings before, during and after the occurrence of the pressure sore; professionals including doctors, nurses and dieticians;

step 1.2, preliminarily screening the pressure sore adverse event early warning indexes extracted in the step 1.1 through discussion of an expert group conference, and reserving indexes which can be recorded and reflected in an information system to form an initial frame of a pressure sore adverse event early warning index system;

step 1.3, screening indexes in the initial frame of the pressure sore adverse event early warning index system obtained in the step 1.2 by adopting a quantitative and qualitative combination mode, and screening out important indexes; the quantitative characteristic is that indexes are screened according to the correlation rule result by performing correlation rule analysis on the real clinical pressure sore adverse events; the quality is that indexes are screened by Delphi expert function inquiry;

and step 1.4, the inviting expert compares the importance of each two of the important indexes screened out in the step 1.3 by adopting a Satty scoring method, and obtains the index weight of the adverse event of pressure sore by utilizing hierarchical analysis software to form an early warning index system of the adverse event of pressure sore.

Preferably, in step 1.4, a comprehensive weighting method is adopted to obtain the weight of the pressure sore adverse event index.

Preferably, the second step is specifically to perform feature extraction by using an ultra-short text vector feature extraction method based on a random up-down sampling text convolutional neural network RS-CNN.

Preferably, the second step specifically comprises the following steps:

step 2.1, performing word segmentation on the early warning index system of the pressure sore adverse event constructed in the step one by using a jieba word segmentation method; pre-training the document data set after Word segmentation through a Word2Vec Word vector extraction model to obtain a feature Word tensor set of each text;

2.2, using a traditional Text-CNN model to combine random up-down sampling to construct a classification model RS-CNN, and converting words in the characteristic word vector set of each Text into a word vector set;

and 2.3, enhancing TF-IDF by integrating collocation as the vocabulary characteristics, and constructing a total characteristic word set.

Preferably, the third step comprises the following steps:

step 3.1, data preprocessing, specifically comprising: converting multiple indexes into a few comprehensive indexes by using a Principal Component Analysis (PCA) technology and utilizing a dimensionality reduction thought for the total feature word set obtained in the step 2.3, and then cleaning the data to obtain preprocessed data;

step 3.2, performing polynomial data feature mapping on the preprocessed data, and converting the data to obtain data required by the model, namely a data sample;

3.3, randomly distributing the data samples obtained in the step 3.2 into a training set and a testing set;

3.4, training the data of the test set and the training set obtained in the step 3.3 by using an SVM model to respectively obtain a training result on the test set and a training result on the training set;

and 3.5, performing ten-fold cross validation to verify the result, and obtaining and outputting an early warning result.

By applying the method, firstly, early warning indexes of the pressure sore adverse events are obtained, and an early warning index system of the pressure sore adverse events is constructed (particularly, a mixed research method combining semi-structural interview, association rules and Delphi expert function inquiry is used for mining the characteristic information of the clinical real pressure sore adverse events, and an early warning index system of the pressure sore adverse events is constructed); secondly, performing feature extraction on the early warning index system of the pressure sore adverse event constructed in the step one (particularly performing feature extraction on an ultra-short text vector feature extraction method based on a random up-down sampling text convolution neural network RS-CNN); and finally, establishing a prediction model by using an algorithm of a support vector machine, and outputting an early warning result. The method has the following beneficial effects:

objectivity and effectiveness of an early warning index: the early warning indexes of the adverse events of pressure sores obtained by the invention not only comprise the content of the original pressure sore risk assessment tool, but also discover clinical objective indexes related to the occurrence of pressure sores in a qualitative interview mode, and the indexes can be reflected by biochemistry, inspection results, medical advice content, definite clinical diagnosis and record and pricing information of patients; the index early warning is real and objective and dynamically changed, so that the real-time performance and accuracy of the early warning can be ensured, and the influence of subjective judgment is reduced.

Scientificity and practicability of the data processing method: in the pressure sore prediction research, because a large amount of text data exists in medical advice data, and problems still exist at home and abroad in the research of carrying out data processing on feature extraction so as to train a prediction model, the method provided by the invention utilizes a convolutional neural network to carry out feature extraction, then utilizes an algorithm of a support vector machine to establish the prediction model, and improves the prediction result to 84.8 percent, so that in the pressure sore prediction, the accuracy of prediction can be effectively improved by utilizing the convolutional neural network to carry out feature extraction and utilizing the algorithm of the support vector machine to establish the prediction model.

Thirdly, economic benefits and social benefits of the early warning model are as follows: compared with the traditional method depending on clinical nurse evaluation, the prediction model obtained by the invention can timely find the clinically-occurred pressure sore or the case with pressure sore risk, and reduce the situations of reporting and missing report, thereby collecting the adverse events of pressure sore as comprehensively as possible, summarizing and analyzing the intervention scheme, reducing the occurrence of pressure sore, saving a large amount of manpower, material resources and financial resources, realizing the early-finding early intervention and early treatment on the risk early warning of pressure injury in real sense, and having important significance for relieving the pain of patients, reducing the medical expense and promoting the development of medical services.

Drawings

FIG. 1 is a schematic flow chart of a method for warning of adverse pressure sore events according to the present invention;

FIG. 2 is a comparison chart of training results of LR, LDA, KNN, NB and RS-CNN models.

Detailed Description

The embodiments of the present invention will be described in detail with reference to the accompanying drawings so that the advantages and features of the invention can be more easily understood by those skilled in the art, and the scope of the invention will be clearly and clearly defined.

Example (b):

an early warning method for adverse pressure sore events is detailed in fig. 1, and specifically comprises the following aspects:

firstly, acquiring early warning indexes of pressure sore adverse events, and constructing an early warning index system of the pressure sore adverse events, which specifically comprises the following steps: the method is characterized by mining the characteristic information of clinical real pressure sore adverse events by using a mixed research method combining a semi-structure interview, an association rule and a Delphi expert function inquiry, and constructing an early warning index system of the pressure sore adverse events, and comprises the following specific steps of:

step 1.1, in order to comprehensively and truly acquire the pressure sore adverse event early warning index, the embodiment is based on literature review, a semi-structure interview (interview to data saturation) is performed under the guidance of a crisis management annular model, and by collecting the processing experience and feeling before, during and after the occurrence of pressure sore of professionals (including doctors, nurses and nurses) related to clinical pressure sore adverse events, the pressure sore adverse event early warning index is extracted from the interview synopsis constructed by a pressure sore adverse event early warning index system for nurses, doctors and nurses:

(1) interview outline (nurse) constructed by pressure sore adverse event early warning index system

Introduction to

Thank you to agree with our interview, and with this study we want to know the signs before, during and after the adverse event of pressure sore in the patient, our conversation will probably take about 1 hour, and the experience of each person will be different, because of the study needs, we may ask questions that do not look relevant to the study, or that look very simple and detailed, so that we can better understand the psychological and behavioral changes of people when they need to solve the same problem

This is a completely voluntary interview, and if some questions you don't want to answer or do not want to mention, please mention, we can change one topic. If you want to turn off the recording, please also show it, we can turn off.

The content we talk to is kept secret and no other irrelevant people will know that this is what you say. We will share common information with clinical staff, but no one other than the study group staff can get our talk information

Do you have any concern about what you have for our recorded conversation;

[ CANCELING-OVER OR CANCELLING ]

Warming-up device

Thank you for our study.

How long you have engaged in the care; in addition to care, do you have other medical practices (e.g., dieticians); what are the skin problems of the patient who is touched after your work; what are the most impressive skin problem events you or your colleagues have experienced;

assessment of prevention

How you find and treat potential skin problem patients in your work;

please confirm that the interviewee talks about the following questions, with additional questions if necessary:

you think that which hospitalized patients are prone to skin problems (potential skin problems exist);

how your and your colleagues evaluate them;

how the evaluation result is recorded is generally described below.

You think which preventive measures should be taken by such patients;

how the documents (doctor's advice, nursing notes, ward logs) are recorded;

do some of the complex skin conditions to be assessed require consultation; please refer to which departments;

what effect the outcome of the consultation has on the patient's skin management;

how your leadership requires with respect to prevention and assessment of patients with potential skin problems;

(details of the unit or department, the method of treatment of the system;)

For preventive assessment of potential skin problems, what you still want to complement;

in the pressure injury nursing process

Please describe the way and idea of your colleague after the patient suffers from stress injury, or do it;

please confirm that the interviewee talks about the following questions, with additional questions if necessary:

how your or your colleagues assess stress-impaired patients;

how the evaluation result is recorded is generally described below.

How your or your colleagues are observing the skin condition of the patient of this type of hand-over;

how your or your colleagues are caring for the skin of such patients;

do some treatments of complex skin conditions require consultation; what is generally the consultation;

what effect the outcome of the consultation has on the patient's skin management;

for skin care of such patients, how your leadership is required;

(details of the unit or department, the method of treatment of the system;)

What you want to complement with respect to skin observation and treatment of a pressure-injured patient;

reporting of adverse events of pressure sores

For the report of adverse events of pressure sores, what experience or idea you or your colleagues have;

please confirm that the interviewee talks about the following questions, with additional questions if necessary:

does you hear or participate in the report of the adverse event of the overpressure sore by colleagues at your sides; what the specific situation is; (who is involved in what role; what information you think they need to provide; what information)

Reporting what influence is caused to you or a reporter by the adverse pressure sore event;

reporting what influence the adverse pressure sore event has on the skin management of the patient;

reporting about adverse events of pressure sores, what you still want to supplement;

what is suggested about our research;

(2) interview outline (doctor) constructed by pressure sore adverse event early warning index system

Introduction to

Thank you for our interview, and through this study we want to know the signs before, during and after the adverse event of pressure sore in the patient, we can talk about 1 hour, and the experience of each person will be different, because of the study needs, we may ask some questions that do not look relevant to the study, or that look simple and detailed, so that we can better understand the psychological and behavioral changes of people when they need to solve the same questions;

this is a completely voluntary interview, and if some questions you don't want to answer or do not want to mention, please mention, we can change one topic. If you want to turn off the recording, please also show it, we can turn off.

The content we talk to is kept secret and no other irrelevant people will know that this is what you say. We will share common information with clinical staff, but no one other than the study group staff can get our talk information

Do you have any concern about what you have for our recorded conversation;

[ CANCELING-OVER OR CANCELLING ]

You have worked for several years; weighing; what are the skin problems of the patient that you pay attention to after work; what are the most impressive skin problem events you or your colleagues have experienced; do you know the pressure injury (pressure sore) of the skin;

assessment of prevention

You think how to more accurately find the patients with potential skin pressure injury problems; how you are dealing with such patients;

please confirm that the interviewee talks about the following questions, with additional questions if necessary:

you think that which hospitalized patients are prone to pressure sores (potential pressure sores exist);

the information just spoken by you is recorded in the patient's admission record, ward round record or hand-over book; how it is described;

if not, you feel that how the care record describes is more convenient for you to know the condition;

in the face of potential pressure sore patients, from which aspects you or your colleagues will perform patient management;

how the presence of potential pressure sores affects the treatment regimen for this type of patient; what is different from the same patient species; please describe below.

Do some of the complex skin conditions to be assessed require consultation; please refer to which departments;

what is common to the results of the consultation;

what effect the outcome of the consultation has on the patient's skin management;

for preventive assessment of potential skin problems, what you still want to complement;

after the patient has pressure sore

How you treat facing a patient who has developed a pressure sore;

please confirm that the interviewee talks about the following questions, with additional questions if necessary:

how your or your colleagues assess pressure sore patients;

does the assessment result have a medical record; how it is described, if any; if not, you feel that how the care record describes is more convenient for you to know the condition;

in medicine, how you or colleagues around you observe and take care of the skin condition of patients, especially patients needing to change medicine;

how you see nurses care for the skin of such patients;

how much pressure sores occur affects the treatment regimen for such patients; please describe below.

Do some treatments of complex skin conditions require consultation; please refer to which departments;

what effect the outcome of the consultation has on the patient's skin management;

what you want to complement with respect to skin observation and care of a pressure-injured patient;

reporting of adverse events of pressure sores

For the report of adverse events of pressure sores, what experience, idea or suggestion you have;

please confirm that the interviewee talks about the following questions, with additional questions if necessary:

does you hear or participate in the report of the adverse event of the overpressure sore by colleagues at your sides; what the specific situation is; (who is involved in what role; what information you think they need to provide; what information)

What influence the reporting of the adverse event has on you or the reporting person;

reporting what content is increased for pressure sore management of a patient by adverse events; prescribing the doctor; please describe below.

Reporting about adverse events of pressure sores, what you still want to supplement;

what is suggested about our research;

(3) interview outline (nutriologist) constructed by pressure sore adverse event early warning index system

Introduction to

Thank you for our interview, and we want to know the traces before, during and after the adverse event of pressure sore through this study, our conversation will probably take about 1 hour, and the experience of each person will be different, because we may ask questions that look irrelevant to the study, or that look simple and detailed, so that we can better understand the different details of the development of the same question.

This is a completely voluntary interview, and if some questions you don't want to answer or do not want to mention, please mention, we can change one topic. If you want to turn off the recording, please also show it, we can turn off.

The content we talk to is kept secret and no other irrelevant people will know that this is what you say. We will share common information with clinical staff, but no one except the research team can get our talk information;

do you have any concern about what you have for our recorded conversation;

[ CANCELING-OVER OR CANCELLING ]

Please confirm that the interviewee talks about the following questions, with additional questions if necessary:

you are currently several grades of dieticians;

what are the skin problems of the patient who is touched after your work;

you think that which hospitalized patients are prone to skin problems (or have potential skin problems);

how such patients are in contact with the nutrition (how you are in contact with such patients);

how your or your colleagues assess their nutritional status; how to record;

the nutritional specialties of such patients examine what is;

how well the common indicators reflect their nutritional status;

what is generally included in the nutritional management of such patients; what is special;

what are the most impressive skin problem events you or your colleagues have experienced; how you did at that time; please describe in detail.

What you want to say about the nutritional management of skin problem patients;

what is suggested about our research;

the list of semi-structured interview topics used in this embodiment is shown in table 1:

TABLE 1 semi-structured interview subject List

Step 1.2, preliminarily screening the pressure sore adverse event early warning indexes extracted in the step 1.1 through discussion of an expert group conference (such as a brainstorm), and reserving indexes which can be recorded and reflected in an information system to form an initial frame of a pressure sore adverse event early warning index system;

step 1.3, screening indexes in the initial frame of the pressure sore adverse event early warning index system obtained in the step 1.2 by adopting a quantitative and qualitative combination mode (namely screening items by association rule and Delphi expert consultation), and screening out important indexes; the quantitative method is characterized in that a subject group of a plurality of daily system researchers (preferably 2 bits) adopts an artificial marking mode to utilize an initial frame of a pressure sore adverse event early warning index system discussed and passed by the subject group, and an Excel table is used for referring to a variable value-giving table to search and record related information in a medical information system in a hospital intranet; the data information comprises general information of the patient and each item in the early warning index item pool of the adverse pressure sore event determined by the head storm, and the general information comprises information such as sex, age, length of stay in hospital and the like; filling missing items, eliminating noise data and carrying out consistency processing, clarifying the relation among indexes by means of an association rule and probably contributing to pressure sore adverse events by each index, and providing reference for weight setting and verification of an early warning model. Deleting the items according to the expert opinions and deleting standards of all indexes, wherein the deleting standards of all indexes are as follows: the importance assignment mean is less than 4; the approval rate is less than 80 percent; the coefficient of variation is greater than 20%. Here, the quality is to screen the index by means of Delphi expert function. In order to ensure that important information is not omitted, a subject group conference is held to inquire the result of an expert letter, analyze the association rule and analyze the expert experience in a combined manner, and finally a content framework of the pressure sore adverse event early warning index system is established.

The specific frame of the index system formed by sorting and screening in this example is detailed in table 2:

TABLE 2 early warning index system content framework for pressure sore adverse events

In the early warning theory, the early warning indexes comprise three types of indexes of a warning source, a warning sign and a warning situation, wherein: the source of police refers to various source factors causing crisis; the warning sign is a sign when abnormal change occurs; the alert refers to the external manifestation of the crisis, a risk that has already been manifested. Since the formation process of the pressure sore is a causal relation, no warning sign index exists in the warning sign system (the warning sign report of the occurrence of the pressure sore is not found in the literature). The search for the source of police is to analyze various risk factors. According to the definition of risk, the risk of any project comes from uncertainty about the various factors of the project. By referring to the "ergonomic" approach, the sources of uncertainty include three areas: human activity (human), physical factors (machine), environmental factors (environment) of the project. Entries in the early warning System of table 2 and dictionary base sources are required to be able to find traces in the Hospital Information System (HIS). The pauses in the HIS that cannot be queried are excluded.

And step 1.4, the inviting expert compares the importance of every two selected important indexes in the step 1.3 by adopting a Satty scoring method, and obtains the index weight of the adverse event of pressure sore by utilizing hierarchical analysis software (namely, a comprehensive weighting method comprising a Delphi method and hierarchical analysis) to form an early warning index system of the adverse event of pressure sore.

The index weight coefficient table of the pressure sore adverse event early warning index system obtained in the embodiment is detailed in table 3:

TABLE 3 PRESS-CREAM INDICATOR WEIGHT COEFFICIENT TABLE

In the embodiment, a pressure sore adverse event early warning index system framework is constructed, and the framework comprises 2 primary themes, 13 secondary themes and 35 tertiary themes. Unlike the case of other researches in which "the sum of the index weights is 1", the weight average of the two first-level indexes, namely the source alarm index and the alert condition index weight, is set to 1 in this embodiment. Because the police source index covers objective indexes directly or indirectly related to pressure sore occurrence, and the contribution degrees of different indexes to the pressure sore occurrence are slightly different, the total weight of the police source index is set to be 1, the weight of each level of indexes under the police source is discussed according to the comparison result of the importance of every two of inquiry experts, and reference can be provided for the pressure sore occurrence risk degree calculated by a later-stage pressure sore adverse event early warning model. And the warning condition index indicates the relevant record after the adverse pressure sore event occurs. These recordings are not needed if no pressure sores occur. And once the written records of the pressure sore description and treatment appear, the pressure sore is indicated to occur. Therefore, each weight of the second-level index and the third-level index under the alarm condition index is 1. This accords with the clinical pressure sore occurrence development and record reality.

Secondly, extracting the characteristics of the constructed early warning index system of the pressure sore adverse events, namely extracting the characteristics of the ultrashort text vector characteristic extraction method based on the random up-down sampling text convolution neural network RS-CNN, wherein the method comprises the following steps:

step 2.1, performing word segmentation on the constructed early warning index system of the adverse pressure sore event by using a jieba word segmentation method, performing word segmentation by using a self-defined dictionary word segmentation tool, collecting common nonsense words, constructing a stop word dictionary, filtering stop words with little meaning, reducing the interference of the model, improving the word segmentation effect of the model, and finally obtaining a de-noised document data set; pre-training the document data set after Word segmentation through a Word2Vec Word vector extraction model to obtain a feature Word tensor set of each text, and laying a foundation for subsequent feature extraction;

step 2.2, considering that most texts such as medical orders and medication records in the pressure sore diagnosis and treatment process are less than 300 characters and belong to ultra-short texts, the method uses a traditional Text-CNN model in combination with random up-down sampling to construct a classification model RS-CNN to convert words in a characteristic word vector set of each Text into a word vector set, and specifically comprises the following steps:

a set of random up-down sampling mechanism is designed to solve the problem of sparse key information features in short text documents, and the method specifically comprises the following steps: the weight of the key features can be amplified through random up-sampling, and the random down-sampling can further reduce the over-fitting phenomenon caused by insufficient sample number on the basis of pooling operation after convolution operation; in addition, the technology of random up-down sampling can effectively fuse the feature information about the word sequence of the text in the shallow network and the semantic features of the deep text, as follows:

unlike conventional convolutional networks, this embodiment introduces a random upsampling operation, assuming the current layer of the network

Is of a size of

Is expressed as an eigentensor

Then the operation of upsampling is defined as:

wherein:

defining the fusion characteristics of the current network layer characteristic tensor and the up-sampling characteristic tensor through characteristic fusion operationACT _CCT () implementation.ACT _US (x) represents an upsampling operation which searches shallow nets in the current net layer for a size matching the current net layer

Feature tensor

After finding out the feature tensor with matched size, the tensor is amplified by one time in each dimension, and feature fusion operation is carried out

Feature fusion is an operation at the tensor level, generating a magnitude of

And the characteristic tensor is the sum of the two tensor channels.

The down-sampling is to reduce the random tensor of the current layer, the method reduces the size of the current tensor and the number of channels by half by adopting a random function, and increases the universality of the method to reduce the overfitting effect by reducing the original tensor, and the operation is defined as follows:

wherein:

for down-sampling operations in which the upper layer tensor size is

Then, the feature tensor obtained by down-sampling

The size of the upper layer is reduced by half, that is, the size of the upper layer is 0.5n*0.5nAnd the characteristic tensor is a half of the upper layer tensor in the number of the channels of the characteristics in the tensor.

The characteristic change conditions after up-down sampling can be obtained through monitoring the characteristic tensors of each layer of the neural network:

1) the initial feature tensor shape is 16 × 2, namely, the dimension of the matrix is 16 × 16, and the number of channels is 2.

16*16*2

2) The shape change of the tensor after a certain convolution operation is 64 x 2:

3) by upsampling, the tensor shape is expanded to 128 × 4:

4) by downsampling, feature subtraction is performed on the basis of the feature tensor of the upper layer 128 × 4, generating a feature tensor whose shape is 64 × 2:

on the basis of word segmentation, a classification model RS-CNN for ultra-short Text vector feature extraction is constructed by combining a traditional Text-CNN model with random up-down sampling, pressure sore type prediction based on data such as medical advice and medication records is gradually realized, the application effect of the Text-CNN model and the word2vec model in an actual scene is deeply analyzed through data preprocessing, word vector training, modeling details, classification functions and data visualization, and words are converted into a word vector form;

and 2.3, enhancing TF-IDF by integrating collocation as lexical characteristics, extracting collocation words according to the part of speech determination, forming specific part of speech forms such as adjective + noun, noun + verb and the like, thereby obtaining relevant statistical information of the document, including the word frequency of each word, and taking the words with the word frequency larger than K and converting the words into word vectors to construct a total characteristic word set.

Because the pressure sore data and the non-pressure sore data are seriously unbalanced, the data are processed by respectively adopting an up-sampling method and a down-sampling method during model training, and the results after two data training are compared are shown in figure 2 in detail, and the results are shown in figure 2: the accuracy of the training results of the KNN and the RS-CNN is very close to each other, but in the actual results, the training result of the RS-CNN is slightly higher than that of the KNN model training, which shows that the detection effect on the unbalanced and ultrashort text pressure sore data set can be effectively improved by the up-and-down sampling based text convolution network RS-CNN.

Thirdly, establishing a prediction model by using an algorithm of a support vector machine, and outputting an early warning result, wherein the specific technical route is as follows:

step 3.1, data preprocessing, specifically comprising: converting multiple indexes into a few comprehensive indexes by using a Principal Component Analysis (PCA) technology and a dimensionality reduction thought for the total feature word set obtained in the step 2.3, and then cleaning the data (checking data consistency, processing invalid values, missing values and the like, see the prior art) to obtain preprocessed data;

in this embodiment, the data is provided by the three hunan-yai hospital, and mainly includes a nursing record sheet, basic information of a patient, a pricing item, an evaluation scoring table, medical advice, diagnosis and other information tables, data in the table is extracted based on an ultra-short text vector feature of a random up-down sampling text convolutional neural network, and extracted data is subjected to correlation analysis using Apriori algorithm, and partial results are as follows:

the PCA principal component analysis technology is used for original data, the idea of reducing dimensions is utilized, multiple indexes are converted into a few comprehensive indexes, then data cleaning is carried out on the data (data consistency is checked, invalid values and missing values are processed, and the like) to obtain preprocessed data, and data needed by a model are obtained after data preprocessing, as shown in a table 4:

TABLE 4 preprocessed data required to obtain model

Step 3.2, performing polynomial data feature mapping on the preprocessed data, and converting the data to obtain data required by the model, namely a data sample;

3.3, randomly dividing the data samples obtained in the step 3.2 into a training set and a test set, wherein the samples are generally randomly divided into 80% serving as the training set and 20% serving as the test set;

3.4, training the data of the test set and the training set obtained in the step 3.3 by using an SVM model, and respectively obtaining a training result on the test set and a training result on the training set;

in the embodiment, the support vector machine uses the loss function to find the optimal parameter; minimization using L1 regularized linear SVM;

and 3.5, performing ten-fold cross validation to verify the result, and obtaining and outputting an early warning result.

The training of the early warning model of this embodiment is to divide the loaded data set into a training set and a verification set, wherein 70% of the training set is used for training, evaluating and selecting in the model, and 30% of the training set is reserved as the verification data set, and the current verification results are shown in table 5:

TABLE 5 verification results table

Wherein: the first row of values is the accuracy of the prediction; the second and third rows are the resulting outputs of a confusion matrix; the lowest of these is an indicator of the model performance metric. The characteristic dimensions of model training are sex, age, immobilization, insufficient nutrient intake, hormonal therapy, use of vasoactive drugs, stool stimulation, traumatic therapy, consultation, skin grafting, posture placement, dressing change, preventive use of auxiliary materials, compression fixation repositioning device, Braden score, stress point skin condition, hypo-perfusion, sweat stimulation, urine stimulation, pressure reduction mattress, skin description, bed rest, unconsciousness, edema, mechanical ventilation, skin type, etc. and the model is in a process of continuous optimization. A user of the model only needs to install relevant software, such as anaconda, python environment configuration and the like, and input relevant parameters to obtain relevant prediction results.

The above description is only an embodiment of the present invention, and not intended to limit the scope of the present invention, and all modifications of equivalent structures and equivalent processes performed by the present specification and drawings, or directly or indirectly applied to other related technical fields, are included in the scope of the present invention.

Claims (7)

1. A pressure sore adverse event early warning method is characterized by comprising the following steps: the method comprises the following steps:

step one, acquiring early warning indexes of pressure sore adverse events, and constructing an early warning index system of the pressure sore adverse events;

step two, extracting the characteristics of the early warning index system of the pressure sore adverse events constructed in the step one;

and step three, establishing a prediction model by using an algorithm of a support vector machine, and outputting an early warning result.

2. The method for early warning of adverse events of pressure sores as claimed in claim 1, wherein in the first step, a mixed research method combining a semi-structured interview, an association rule and a delphire expert function is used for mining the characteristic information of the clinically real adverse events of pressure sores to construct an early warning index system of the adverse events of pressure sores.

3. The method for warning of adverse pressure sore events as claimed in claim 2, wherein the first step specifically comprises the steps of:

step 1.1, performing semi-structural interview under the guidance of a crisis management annular model, and extracting early warning indexes of adverse pressure sore events from professional staff relevant to the clinical adverse pressure sore events by collecting the processing experiences and feelings before, during and after the occurrence of the pressure sore; professionals including doctors, nurses and dieticians;

step 1.2, preliminarily screening the pressure sore adverse event early warning indexes extracted in the step 1.1 through discussion of an expert group conference, and reserving indexes which can be recorded and reflected in an information system to form an initial frame of a pressure sore adverse event early warning index system;

step 1.3, screening indexes in the initial frame of the pressure sore adverse event early warning index system obtained in the step 1.2 by adopting a quantitative and qualitative combination mode, and screening out important indexes; the quantitative characteristic is that indexes are screened according to the correlation rule result by performing correlation rule analysis on the real clinical pressure sore adverse events; the quality is that indexes are screened by Delphi expert function inquiry;

and step 1.4, the inviting expert compares the importance of each two of the important indexes screened out in the step 1.3 by adopting a Satty scoring method, and obtains the index weight of the adverse event of pressure sore by utilizing hierarchical analysis software to form an early warning index system of the adverse event of pressure sore.

4. The method for warning of pressure sore adverse events according to claim 3, characterized in that in step 1.4, a comprehensive weighting method is used to obtain the weight of the pressure sore adverse event index.

5. The early warning method for adverse events of pressure sores according to claim 1, wherein the second step is specifically to perform feature extraction based on an ultra-short text vector feature extraction method of a random up-down sampling text convolution neural network (RS-CNN).

6. The early warning method for adverse events of pressure sores according to claim 5, wherein the second step specifically comprises the steps of:

step 2.1, performing word segmentation on the early warning index system of the pressure sore adverse event constructed in the step one by using a jieba word segmentation method; pre-training the document data set after Word segmentation through a Word2Vec Word vector extraction model to obtain a feature Word tensor set of each text;

2.2, using a traditional Text-CNN model to combine random up-down sampling to construct a classification model RS-CNN, and converting words in the characteristic word vector set of each Text into a word vector set;

and 2.3, enhancing TF-IDF by integrating collocation as the vocabulary characteristics, and constructing a total characteristic word set.

7. The method for warning of adverse pressure sore events as claimed in claim 6, wherein said step three comprises the steps of:

step 3.1, data preprocessing, specifically comprising: converting multiple indexes into a few comprehensive indexes by using a Principal Component Analysis (PCA) technology and utilizing a dimensionality reduction thought for the total feature word set obtained in the step 2.3, and then cleaning the data to obtain preprocessed data;

step 3.2, performing polynomial data feature mapping on the preprocessed data, and converting the data to obtain data required by the model, namely a data sample;

3.3, randomly distributing the data samples obtained in the step 3.2 into a training set and a testing set;

3.4, training the data of the test set and the training set obtained in the step 3.3 by using an SVM model to respectively obtain a training result on the test set and a training result on the training set;

and 3.5, performing ten-fold cross validation to verify the result, and obtaining and outputting an early warning result.

Images