CN115798708A - First-aid injury classification method based on long-time sequence - Google Patents
First-aid injury classification method based on long-time sequence Download PDFInfo
- Publication number
- CN115798708A CN115798708A CN202211496521.1A CN202211496521A CN115798708A CN 115798708 A CN115798708 A CN 115798708A CN 202211496521 A CN202211496521 A CN 202211496521A CN 115798708 A CN115798708 A CN 115798708A
- Authority
- CN
- China
- Prior art keywords
- model
- injury
- sample
- rate
- feature
- 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.)
- Pending
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Medical Treatment And Welfare Office Work (AREA)
Abstract
The invention discloses an injury classification method for judging first-aid injuries based on a long-time sequence, which comprises the following steps: constructing a data set containing four injury categories according to predicted mortality indexes of APACHE IV in an eICU emergency database; carrying out serial pretreatment of data cleaning, standardization and feature screening on the data set; performing intelligent modeling on the time sequence characteristics by adopting deep learning, and then sparsely inputting characteristics according to requirements, so that the model realizes higher performance through the minimum key characteristics; and selecting a plurality of deep learning models aiming at the long-time sequence, and obtaining a final classification model through model fusion. The accuracy rate of classification of the injury finally related to the method under the characteristic combination and fusion model can reach 92%, and an effective reference method is provided for the battlefield treatment strategy.
Description
Technical Field
The invention belongs to the technical field of medical signal processing, and relates to a method for classifying emergency injuries based on a long-time sequence.
Background
With the advancement of technology, the modern war is just like a digital, intelligent and high-tech war mode. Under a new war mode, the concept of the traditional soldier combat unit generates a qualitative leap, and the combat capability of an individual soldier system directly determines the strength of the whole battle force on a battlefield. In modern battlefield rescue, rapid and timely treatment plays an extremely important role in reducing the death rate of soldiers, and the treatment sequence needs to be determined according to the injury of the wounded so as to achieve the optimal treatment efficiency, so that the rapid and accurate assessment of the severity of the injury of the soldiers is particularly critical.
At present, most of the injury evaluation systems adopt a simple combat injury scoring method by carrying out standard question-answer type communication on the injured and scoring by combining physiological parameters acquired by a sensor, and combine the obtained comprehensive score with an injury ticket to obtain the injury category.
The simple war injury scoring method comprises the following physiological indexes: heart rate, systolic pressure, respiratory rate, body temperature, consciousness, etc. The normal range of heart rate is 51-100 (times/min), the normal range of systolic pressure is 101-199 (mmHg), the normal value of respiratory rate is 9-14 (times/min), and the normal value of body temperature is 35-38.4 (DEG C). The more deviation from normal, the higher the score. The higher the score, the more severe the injury.
Moreover, the simple war injury scoring method is not easy to master by non-professional medical care personnel, and the problems of inaccurate information, quickness and the like exist.
Disclosure of Invention
In order to solve the defects or shortcomings in the existing injury evaluation technology, the invention aims to provide a method for classifying the injury of first aid based on a long-time sequence, which can be used for rapidly and accurately judging the injury type of the injured soldier by combining various physiological parameters of the injured soldier, assisting a doctor to rapidly and accurately rescue the injured soldier and further reducing the injury and death conditions of the soldier.
In order to realize the task, the invention adopts the following technical solution:
a first-aid injury classification method based on a long-time sequence is characterized by comprising the following steps:
step one, preprocessing an eICI (enhanced Research Database, eICI) to construct a data set for model training and testing;
and step two, intelligent modeling is carried out by adopting deep learning, and injury classification based on static and dynamic time sequence characteristics is realized.
And step three, sparsifying the feature input, so that the model can realize higher performance under the condition of the least feature quantity.
And step four, selecting a plurality of mainstream deep learning models (LSTM, TCN, transformer and Informer), and obtaining a final classification model through model fusion.
According to the invention, the specific implementation steps of the first step are as follows:
step 101: dividing the injury data set according to the predicted ICU mortality index of APACHE IV in the APACHE PatientResult list in the cooperative research database; wherein:
the stroke with the death rate within 0-10% is classified as light injury, the stroke with the death rate within 10-30% is classified as medium injury, the stroke with the death rate within 40-60% is classified as heavy injury, and the stroke with the death rate above 70% is classified as dangerous and heavy injury.
Step 102: according to the id of the patient in the injury data set, two static characteristics of the sex and the age of the patient are extracted from the patient list. The physiological characteristics of heart rate, respiration, blood oxygen, systolic pressure, diastolic pressure, mean pressure, body temperature, glasgow coma index (GCS), etc. are extracted from the nurseCharting list. A total of 26485 samples, 11 features, 100 time points were extracted.
Step 103: processing for the case that the sample has feature missing:
1) If a certain characteristic of the sample has a few values missing, if the characteristic of the sample has missing at the time t, the value at the time t-1 is filled.
2) If there is a feature missing from the sample, the average of the class in which the sample is located is filled.
3) And for the processing of abnormal values, sample filtering is carried out by referring to the upper and lower bounds of a reasonable value given by an expert according to different physiological characteristics, and a sample is discarded when any characteristic value of the sample is not in a reasonable range.
Step 104: normalizing features
Calculating the mean value and the standard deviation of each feature in the whole data set dimension, and standardizing the feature values by using the calculated mean value and standard deviation to eliminate the influence of different feature dimensions; the formula for a dataset with a total number of samples m is as follows:
in the formula, x i Is a characteristic of the ith sample in the data set.
Step 105: according to the following steps of 6:2:2, dividing the training set, the verification set and the test set.
Specifically, in the second step, in order to fully utilize the multi-dimensional characteristic information of the sample on the long-time node, the method firstly adopts a long-time sequence prediction model Informer based on an attention-free mechanism to intelligently model the time sequence characteristics on the constructed data set, and quantitatively evaluates the performance of the model through various evaluation indexes; the method comprises the following specific steps:
step 201: model training and validation
The model adopts a two-layer Informer encoder and a one-layer decoder structure, and is optimized by using an Adam optimizer, wherein the hyperparameter beta 1 And beta 2 Set to 0.99 and 0.9, respectively. The loss function employs multi-class cross-entropy loss, which is mathematically represented as follows:
wherein M is the number of classes, y ic As a function of the sign, p ic Is the predicted probability that the observation sample i belongs to class c.
During training, static features are converted into one-dimensional feature vectors with the same length as the time sequence feature vectors through copying, and then the one-dimensional feature vectors and k groups of time sequence feature vectors with the length of 100 are spliced along the channel dimension to form a two-dimensional feature map
As an input. The training batch size is 64, the learning rate is 0.00005, the learning rate is reduced by 10 times in 20 training rounds, and the maximum number of training rounds is 30. And saving the model of the last round as the optimal model.
Step 202: model testing
Loading the optimal model weight saved in step 201, and calculating the areas under the Accuracy (ACC), F1 score (F1-score, F1) and receiver operating characteristic curve (ROC) of each category on the test set, wherein the calculation formula of ACC and F1 score is as follows:
wherein TP, TN, FP, FN represent true positive rate, true negative rate, false positive rate, and false negative rate, respectively, and R and P represent recall rate and precision rate, respectively.
Further, in step three, considering that the physiological characteristics of the battlefield environment are limited to be easily monitored, it is necessary to achieve high performance using as few physiological characteristics as possible. Sex (Gender, G) and Age (Age, a) of each sample were taken as fixed static characteristics, and indices such as Heart Rate (HR), respiratory Rate (RR), body Temperature (BT), blood oxygen Saturation (O2 Saturation, O2), systolic Pressure (SP), diastolic Pressure (DP), and Glasgow Coma Score (GCS) were taken as time series characteristics. According to performance of different feature combinations on an Informer model, aiming at requirement sparse input features, selecting a feature combination with the highest performance as the input of a final model, and specifically comprising the following steps:
step 301: setting a characteristic set theta, wherein the specific combination is as follows:
step 302: and (5) inputting the models by taking different characteristic combinations, and repeating the step two to obtain the performance expression of each model on the test set.
Further, in the fourth step, in order to further improve the accuracy and robustness of the injury classification model, a neural network model for evaluating a mainstream aiming at a long-time sequence is selected and a classification result of a final model is obtained through model fusion, and the specific implementation steps are as follows:
step 401: and (5) network structure fusion. Selecting a time domain Convolutional Network (TCN) and an Informer as basic models, deleting original output layers of the two models, flattening the eigenvectors output by neurons in the last layer of hidden layer, and splicing along the channel dimension to form new eigenvectors. And fusing through a full connection layer and outputting a final class feature vector.
Step 402: and step two is a model training configuration reference step, two groups of same time sequence characteristic sequences are input simultaneously during training, characteristics are extracted through two paths of a TCN (train control network) and an Informer respectively, and finally, a multi-layer perceptron is utilized to perform characteristic fusion among different models and output a prediction result.
Step 403: and testing the model with the good training model on the test set to obtain the final model.
Step 404: finally, through a network model after TCN and Informer are fused, the classification accuracy rate is 92% under the combination of (A, G, HR, R, BT and CS), and the area under the average ROC curve is 99%. The ROC curve is shown in fig. 2.
The invention discloses a long-time sequence-based emergency injury classification method, which brings technical innovation that:
firstly, data of static dimension and dynamic dimension are fused as characteristics, and a data set of injury severity is divided according to the predicted mortality value of Apache IV;
secondly, in order to fully utilize multi-dimensional characteristic information of a sample on a long-time node, the method comprises the steps of firstly adopting a long-time sequence prediction model inform based on a self-attention mechanism, carrying out intelligent modeling on time sequence characteristics on a constructed data set, and carrying out quantitative evaluation on the performance of the model through various evaluation indexes;
thirdly, various feature combinations are tried, and as few feature data as possible are selected to achieve high performance.
Fourthly, in order to further improve the accuracy and the robustness of the injury classification model, a neural network model with better long-time sequence performance is selected, and the classification result of the final model is obtained through model fusion.
Experiments of the applicant show that the accuracy rate of injury classification under the characteristic combination and fusion model finally related by the method can reach 92%, and an effective reference method is provided for battlefield treatment strategies.
Drawings
Fig. 1 is a flow chart of the first-aid injury classification method based on long time sequence according to the invention.
Fig. 2 shows test set classification ROC curves after model fusion, in which (a) is a severe injury ROC curve (grade area 98.17%), (b) is a severe injury ROC curve (grade area 98.90%), (c) is a medium injury ROC curve (grade area 99.92%), (d) is a light injury ROC curve (grade area 99.92%).
The invention will be described in further detail below with reference to the following figures and examples.
Detailed Description
Referring to fig. 1, the present embodiment provides a method for classifying emergency injuries based on a long time sequence, which includes the following steps:
1) Dividing a cooperation Research Database (eICU) into data sets according to Apache IV predicted mortality and preprocessing, wherein the steps are as follows:
step 101: dividing the estrus data set according to the predicted ICU mortality index of APACHE IV in an apachePatientResult list in the eICU; wherein, the ones with the mortality rate within 0-10% are classified as mild injuries, the ones with the mortality rate within 10-30% are classified as medium injuries, the ones with the mortality rate within 40-60% are classified as heavy injuries, and the ones with the mortality rate above 70% are classified as severe injuries.
Step 102: according to the id of the patient in the injury data set, two static characteristics of the sex and the age of the patient are extracted from the patient list. The physiological characteristics of heart rate, respiration, blood oxygen, systolic pressure, diastolic pressure, mean pressure, body temperature, glasgow coma index (GCS) and the like are extracted from the nurseCharting list. A total of 26485 samples, 11 features, 100 time points were extracted.
Step 103: and (3) processing the sample with the characteristic missing condition:
(1) If a certain characteristic of the sample has a few values missing, if the characteristic of the sample has missing at the time t, the value at the time t-1 is filled.
(2) If there is a feature missing from the sample, the average of the class in which the sample is located is filled.
(3) And for the processing of abnormal values, sample filtering is carried out by referring to the upper and lower bounds of a reasonable value given by an expert according to different physiological characteristics, and a sample is discarded when any characteristic value of the sample is not in a reasonable range.
Step 104: the features are normalized. And calculating the mean value and the standard deviation of each feature on the dimension of the whole data set, and normalizing the feature values by using the calculated mean value and standard deviation so as to eliminate the influence of different feature dimensions. The formula for a dataset with a total number of samples m is as follows:
in the formula, x i Is characteristic of the ith sample in the data set.
Step 105: according to the following steps of 6:2:2, dividing the training set, the verification set and the test set.
2) In order to fully utilize the multi-dimensional characteristic information of the sample on the long-time node, the intelligent modeling is carried out on the time sequence characteristic on the constructed data set, and the steps of evaluating the model performance are as follows:
step 201: model training and validation
The model adopts a two-layer Informer encoder and one-layer decoder structure, and is optimized by using an Adam optimizer, wherein the hyperparameter beta 1 And beta 2 Set to 0.99 and 0.9, respectively. The loss function employs multi-class cross-entropy loss, which is mathematically represented as follows:
where M is the number of classes, y ic As a function of the sign, p ic Is the predicted probability that the observation sample i belongs to class c.
During training, static features are converted into one-dimensional feature vectors with the same length as the time sequence feature vectors through copying, and then the one-dimensional feature vectors and k groups of time sequence feature vectors with the length of 100 are spliced along the channel dimension to form a two-dimensional feature map
As an input. The training batch size is 64, the learning rate is 0.00005, the learning rate is reduced by 10 times in 20 training rounds, and the maximum number of training rounds is 30. And saving the model of the last round as the optimal model.
Step 202: model testing
Loading the optimal model weight saved in step 201, and calculating the area under the Accuracy (ACC), F1 score (F1-score, F1) and the receiver operating characteristic curve (ROC) of each category on the test set, wherein the calculation formula of ACC and F1 score is as follows:
wherein TP, TN, FP, FN represent true positive rate, true negative rate, false positive rate, and false negative rate, respectively, and R and P represent recall rate and precision rate, respectively.
3) For feature input sparseness, the steps for enabling the model to achieve higher performance with as few feature quantities as possible are:
step 301: setting a characteristic set theta, wherein the specific combination is as follows:
step 302: and (5) inputting the models by taking different characteristic combinations, and repeating the step two to obtain the performance expression of each model on the test set.
4) In order to further improve the accuracy and robustness of the injury classification model, a neural network model for evaluating a mainstream aiming at a long-time sequence is selected and a classification result of a final model is obtained through model fusion:
step 401: network architecture convergence
Selecting a time domain Convolutional Network (TCN) and an Informer as basic models, deleting original output layers of the two models, flattening the eigenvectors output by neurons in the last layer of hidden layer, and splicing along the channel dimension to form new eigenvectors. And fusing through a full connection layer and outputting a final class feature vector.
Step 402: and step two is referred to for model training configuration, two groups of same time sequence feature sequences are simultaneously input during training, features are extracted through two paths of TCN and inform, and finally a multilayer perceptron is used for feature fusion among different models and outputting a prediction result.
Step 403: and testing the model with the good training model on the test set to obtain the final model.
Step 404: finally, through a network model after TCN and Informer are fused, the classification accuracy rate is 92% under the combination of (A, G, HR, R, BT and CS), and the area under the average ROC curve is 99%. The ROC curve is shown in fig. 2.
The method for classifying the emergency injury based on the long-time sequence provided by the embodiment relates to three technical levels:
(1) In order to fully utilize multi-dimensional characteristic information of a sample on a long-time node, firstly, a long-time sequence prediction model Informer based on an attention mechanism is adopted to intelligently model a time sequence characteristic on a constructed data set, and the performance of the model is quantitatively evaluated through various evaluation indexes;
(2) Various feature combinations are tried, and as few feature data as possible are selected to achieve high performance.
(3) In order to further improve the accuracy and robustness of the injury classification model, a neural network model with better long-time sequence performance is selected and the classification result of the final model is obtained through model fusion.
The result shows that under the new characteristic combination, the fused model has higher performance, can accurately and quickly judge the injury severity of the patient and provides reference for subsequently adopting corresponding agile treatment strategies.
Claims (5)
1. A first-aid injury classification method based on a long-time sequence is characterized by comprising the following steps:
firstly, preprocessing a cooperative research database to construct a data set for model training and testing;
secondly, intelligent modeling is carried out by adopting deep learning, and injury classification based on static and dynamic time sequence characteristics is realized;
and step three, the feature input is thinned, so that the model can realize higher performance under the condition of the least feature quantity.
And step four, selecting a plurality of mainstream deep learning models, and obtaining a final classification model through model fusion.
2. The method of claim 1, wherein the step one is implemented by the following steps:
step 101: the impairment data set is partitioned according to predicted ICU mortality indicators for APACHEIV in the apache PatientResult list in the collaborative research database, wherein:
the scratch with the mortality rate within 0-10% is classified as a light injury, the scratch with the mortality rate within 10-30% is classified as a medium injury, the scratch with the mortality rate within 40-60% is classified as a heavy injury, and the scratch with the mortality rate above 70% is classified as a critical heavy injury;
step 102: extracting two static characteristics of the sex and the age of the patient from a patient list according to the id of the patient in the injury data set; the following physiological characteristics were extracted in the nurseCharting list: heart rate, respiration, blood oxygen, systolic pressure, diastolic pressure, mean pressure, body temperature, glasgow coma index; 26485 samples, 11 features and 100 time points are extracted;
step 103: processing for sample presence feature missing condition
1) If a certain characteristic of the sample has a few value loss, namely, if the characteristic of the sample has loss at the time t, filling the value at the time t-1;
2) Filling the average value of the category of the sample if the sample has certain characteristic missing;
3) For the processing of abnormal values, sample filtering is carried out by referring to the upper and lower bounds of a reasonable value given by an expert according to different physiological characteristics, and a sample is discarded if any characteristic value of the sample is not in a reasonable range;
step 104: normalizing features
Calculating the mean value and the standard deviation of each feature in the whole data set dimension, and standardizing the feature values by using the calculated mean value and standard deviation to eliminate the influence of different feature dimensions; the formula for a data set with a total number of samples m is as follows:
in the formula, x i Features of the ith sample in the dataset;
step 105: according to the proportion of 6:2:2, dividing the training set, the verification set and the test set.
3. The method of claim 1, wherein in the second step, in order to fully utilize multi-dimensional feature information of the sample on the long-time node, the time-series feature is intelligently modeled on the constructed data set, and the performance of the model is quantitatively evaluated through various evaluation indexes, and the specific steps are as follows:
step 201: model training and validation
The model adopts a two-layer Informer encoder and one-layer decoder structure, and is optimized by using an Adam optimizer, wherein the hyperparameter beta 1 And beta 2 Set to 0.99 and 0.9, respectively; the loss function employs multi-class cross-entropy loss, which is mathematically represented as follows:
wherein M is the number of classes, y ic As a function of the sign, p ic A predicted probability of belonging to class c for the observation sample i;
during training, static features are converted into one-dimensional feature vectors with the length same as that of the time sequence feature vectors through copying, and then the one-dimensional feature vectors and k groups of time sequence feature vectors with the length of 100 are spliced along channel dimensions to form a two-dimensional feature map
As an input; the batch size of training is 64, the learning rate is 0.00005, the learning rate is reduced by 10 times in each 20 training rounds, and the maximum number of the training rounds is 30; saving the model of the last round as an optimal model;
step 202: model testing
Loading the optimal model weight saved in step 201, and calculating the accuracy, F1 score and area under the working characteristic curve of the subject of each category on the test set, wherein the calculation formula of ACC and F1 score is as follows:
wherein TP, TN, FP, FN represent true positive rate, true negative rate, false positive rate, and false negative rate, respectively, and R and P represent recall rate and precision rate, respectively.
4. The method of claim 1, wherein the third step is implemented by the following steps:
step 301: setting a characteristic set theta, wherein the specific combination is as follows:
step 302: and (5) inputting the models by taking different characteristic combinations, and repeating the step two to obtain the performance expression of each model on the test set.
5. The method of claim 1, wherein the fourth step is implemented by the following steps:
step 401: network architecture convergence
Selecting a time domain Convolutional Network (TCN) and an Informer as basic models, deleting original output layers of the two models, flattening a feature vector output by a neuron of a last layer of a hidden layer, splicing the feature vectors along the channel dimension to form a new feature vector, fusing through a full connection layer and outputting a final class feature vector;
step 402: the model training configuration reference step II is that two groups of same time sequence characteristic sequences are input simultaneously during training, characteristics are extracted through two paths of TCN and inform, and finally a multilayer perceptron is used for carrying out characteristic fusion among different models and outputting a prediction result;
step 403: testing the model with the good training model on the test set to obtain a final model;
step 404: finally, through a network model after TCN and Informer are fused, the classification accuracy rate is 92% under the combination of (A, G, HR, R, BT and CS), and the area under the average ROC curve is 99%.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211496521.1A CN115798708A (en) | 2022-11-25 | 2022-11-25 | First-aid injury classification method based on long-time sequence |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211496521.1A CN115798708A (en) | 2022-11-25 | 2022-11-25 | First-aid injury classification method based on long-time sequence |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115798708A true CN115798708A (en) | 2023-03-14 |
Family
ID=85441905
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211496521.1A Pending CN115798708A (en) | 2022-11-25 | 2022-11-25 | First-aid injury classification method based on long-time sequence |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115798708A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116313127A (en) * | 2023-03-23 | 2023-06-23 | 珠海市安克电子技术有限公司 | Decision support system based on pre-hospital first-aid big data |
| CN117598700A (en) * | 2024-01-23 | 2024-02-27 | 吉林大学 | Intelligent blood oxygen saturation detection system and method |
-
2022
- 2022-11-25 CN CN202211496521.1A patent/CN115798708A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116313127A (en) * | 2023-03-23 | 2023-06-23 | 珠海市安克电子技术有限公司 | Decision support system based on pre-hospital first-aid big data |
| CN117598700A (en) * | 2024-01-23 | 2024-02-27 | 吉林大学 | Intelligent blood oxygen saturation detection system and method |
| CN117598700B (en) * | 2024-01-23 | 2024-03-29 | 吉林大学 | Intelligent blood oxygen saturation detection system and method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115798708A (en) | First-aid injury classification method based on long-time sequence | |
| Nikam et al. | Cardiovascular disease prediction using machine learning models | |
| Azrar et al. | Data mining models comparison for diabetes prediction | |
| CN111951975B (en) | Sepsis early warning method based on deep learning model GPT-2 | |
| Alsheref et al. | Blood diseases detection using classical machine learning algorithms | |
| CN111798954A (en) | Drug combination recommendation method based on time attention mechanism and graph convolution network | |
| CN117077786A (en) | Knowledge graph-based data knowledge dual-drive intelligent medical dialogue system and method | |
| CN112992368A (en) | Prediction model system and recording medium for prognosis of severe spinal cord injury | |
| CN115293161A (en) | Reasonable medicine taking system and method based on natural language processing and medicine knowledge graph | |
| CN113223656A (en) | Medicine combination prediction method based on deep learning | |
| CN114724710A (en) | Emergency scheme recommendation method and device for emergency events and storage medium | |
| Pillai | Cardiac disease prediction with tabular neural network | |
| CN116246773A (en) | Knowledge and data combined auxiliary diagnosis and treatment model | |
| CN113903466A (en) | Data processing device and system for auxiliary evaluation of risk degree of cardiovascular diseases of population and application of data processing device and system | |
| Wisaeng | Predict the diagnosis of heart disease using feature selection and k-nearest neighbor algorithm | |
| Bhandari et al. | Comparative analysis of fuzzy expert systems for diabetic diagnosis | |
| Pourahmad et al. | Identifying important attributes for prognostic prediction in traumatic brain injury patients | |
| CN114550896A (en) | Dizziness patient emergency pre-examination triage decision-making method, device and model based on artificial neural network | |
| CN114724701A (en) | Noninvasive ventilation curative effect prediction system based on superposition integration algorithm and automatic encoder | |
| CN114613497A (en) | Intelligent medical auxiliary diagnosis method of patient sample based on GBDT sample level | |
| Umut et al. | Prediction of sepsis disease by Artificial Neural Networks | |
| JP2536287B2 (en) | A classifier using disjunctive normal form stochastic logic | |
| Alagarsamy et al. | A Novel Technique for Prophecy of Brain Strokes | |
| Saleh | Artificial intelligence and statistical techniques to predict probability of injury survival | |
| CN114678132B (en) | Self-learning medical wind control system and method based on clinical behavior feedback |
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 |