CN113349746A

CN113349746A - Vital sign monitoring alarm system

Info

Publication number: CN113349746A
Application number: CN202110834939.8A
Authority: CN
Inventors: 黄伟红; 胡建中; 周建辉; 岳丽青; 黄凌瑾; 赖娟; 张磊; 刘硕; 高武强; 黄佳
Original assignee: Beijing Hezheng Medical Technology Co ltd; Xiangya Hospital of Central South University
Current assignee: Beijing Hezheng Medical Technology Co ltd; Xiangya Hospital of Central South University
Priority date: 2021-07-21
Filing date: 2021-07-21
Publication date: 2021-09-07

Abstract

A vital sign monitoring alarm system, characterized by: the method comprises the following steps: s1, data acquisition, namely measuring vital sign information of a human body through various measuring equipment, collecting various electronic and numerical sign information, and storing data acquired for at least 8 weeks into a database through the Internet of things; dividing a database into two subsets, wherein one subset is a training set, and the other subset is a testing set; s2, summarizing characteristic values of different sign information, and continuously verifying the availability of the characteristic values; s3, forming a real alarm/invalid alarm classification model through machine learning; and S4, verifying the robustness of the classification model through cross-validation combined with a training algorithm and a verification algorithm. The vital sign monitoring alarm method provided by the invention can reduce the probability of invalid alarm, so that medical care resources can be reasonably distributed, and the medical care working efficiency and the service quality can be improved.

Description

Vital sign monitoring alarm system

Technical Field

The invention relates to the technical field of vital information monitoring, in particular to a vital sign monitoring and alarming system.

Background

When a doctor is used for treating a disease, vital sign information of the patient needs to be acquired, wherein the vital signs comprise body temperature (T), respiration (R), pulse (P), Blood Pressure (BP) and the like, and medical care personnel can observe the change of the disease through the vital sign information, so that an important basis is provided for diagnosis, treatment and nursing of the disease. Therefore, it is important to timely monitor vital sign information of a patient, and there are some instruments capable of measuring vital sign data and methods for automatically sending an alarm according to the measurement result of the vital sign in the prior art, for example, CN110522413A discloses a vital sign monitoring system, which includes a vital sign monitoring module and a terminal, where the terminal includes a vital sign display module, a vital sign comparison and analysis module, a data storage module and an alarm module, and the vital sign monitoring module is used for monitoring the vital sign of the patient; the vital sign display module is used for displaying vital signs of the patient; the vital sign comparison and analysis module is used for comparing and analyzing the monitored vital sign data of the patient with the input standard vital sign data; the data storage module is used for recording and storing the vital sign data of the patient; the alarm module is used for giving an alarm to the abnormal phenomenon of the vital signs of the patient. The invention has diversified monitoring to the vital signs of the patient, so that medical personnel or accompanying personnel can know the real-time vital sign change of the patient in real time, and the condition that the patient is not in danger when the patient is accompanied by the patient can be ensured. However, in the actual use process, the implementation of the alarm function has defects, such as not distinguishing the priority of the urgency; and ineffective alarm caused by lead falling, wearing detection equipment falling and the like is excessive.

The vital sign monitoring alarm method can effectively judge the authenticity of the vital sign monitoring alarm, reduces the probability of invalid alarm, so that medical resources are reasonably distributed, and the medical work efficiency and the service quality are improved.

Disclosure of Invention

The purpose of the invention is: the defects of the prior art are overcome, and the vital sign monitoring and alarming system can reduce the probability of invalid alarm occurrence, so that the medical resources are reasonably distributed, and the medical work efficiency and the service quality are improved.

The technical scheme of the invention is as follows: a vital sign monitoring and alarming system comprises the following working processes:

s1, collecting the data,

measuring vital sign information of a human body through various measuring devices, and collecting various electronic and numerical sign information;

s2, summarizing characteristic values of different sign information, and continuously verifying the availability of the characteristic values;

s3, forming a real alarm/invalid alarm classification model through machine learning;

and S4, verifying the robustness of the classification model through cross-validation combined with a training algorithm and a verification algorithm.

Further, a database is established in step S1, and data collected for at least 8 weeks are stored in the database through the internet of things.

Further, the database is divided into two subsets, wherein one subset is a training set, and the other subset is a testing set.

Further, the data volume of the training set is roughly equivalent to the data volume of the test set.

Further, non-invasive vital sign monitoring data of heart rate, respiratory rate and pulse oximetry as well as systolic and diastolic blood pressure were recorded at a frequency of 1/20 Hz;

further, heart rate <40 or > 140; respiratory rate <8 or > 36; systolic blood pressure <80 or > 200; diastolic pressure > 110; pulse oximetry < 85%, defined as a crisis event that requires alarm notification.

Further, the crisis event is made to last for a tolerance of 40 seconds and, if intermittent, for a minimum of 4 minutes or cumulatively for 4 of 5 minutes. A 40 second tolerance indicates that if the next violation occurs within 40 seconds of the current violation, they are the same event. Since the sampling frequency is 1/20Hz, this also means that only one normal reading is allowed to occur between two abnormal readings.

Further, in step S2, the training set data is used to determine rules for which an alarm event is defined as authentic or invalid. The 1/20Hz time plot in the training set data is provided to a reviewer consisting of at least 3 clinical experts, who classifies the alarm data as true or false according to clinical experience to establish the judgment condition of "ground truth".

Furthermore, an SVM support vector machine is adopted to extract characteristic values on the basis of basic facts. These rules capture the majority of events for each vital sign and evolve into textual rules describing the multi-signal patterns of the most commonly encountered ineffective events. For digital feature extraction, these text rules are converted into a set of digital features derived from multi-signal time series data.

Further, all feature values are refined through one iteration for the selected feature of each vital sign. Events such as tags that do not conform to the characterization rules are reviewed by experts, and then they are relabeled, new features are derived, or text rules are redefined, and their features are redefined. The clinical expert reviewer simultaneously annotates all alarms in the test set as true or false as a test set. And determining the digital characteristics at the end of the characteristic learning step, and judging the usability of the characteristic value by using the AUC and the ROC curve.

Further, to reduce the complexity of the machine learning model and the risk of overfitting, a AUC score of 0.7 was used as an arbitrary cut-off point for blood pressure and pulse oximetry, and a respiratory rate cut-off point of 0.85 to filter the less informative features, leaving 10 features for pulse oximetry, 5 features for respiratory rate, and 5 features for blood pressure.

Further, in step S3, the binary label of the alarm is classified, i.e. the alarm is in both true and invalid states, given the characteristic values above. The machine learning algorithms used are: k nearest neighbors (KNN, at different K), Linear Discriminant Analysis (LDA), naive bayes classifier (NB), Logistic Regression (LR), Support Vector Machine (SVM), and Random Forest (RF). The algorithms are all known mechanical learning algorithms, and the main purpose is to complete the calculation of the characteristic value and form a clear judgment condition.

Further, in step S4, the best classification model applied to different sign information is determined by using ten-fold cross validation in combination with the training algorithm and the validation algorithm AUC score. The selected model is applied to the data in the test set to obtain a result. The AUC scores were from ten-fold cross-validation experiments on an algorithmic array of three classes of vital sign events.

Compared with the prior art, the invention has the beneficial effects that: the vital sign monitoring and alarming system provided by the invention can reduce the probability of invalid alarm, so that medical care resources can be reasonably distributed, and the medical care working efficiency and the service quality can be improved.

Drawings

FIG. 1 is a graph of AUC for various algorithms in example 1 of the present invention;

FIG. 2 is a ROC diagram in example 1 of the present invention, wherein FIGS. a and b are set 1 and set 2, respectively;

in the figure: NN-nearest neighbor, LDA-linear discriminant analysis, NB-naive Bayes classifier, LR-logistic regression, SVM-support vector machine, RF-random forest, RR-respiratory frequency, BP-blood pressure, SpO 2-peripheral pulse blood oxygen saturation.

Detailed Description

The present invention will be described in further detail with reference to specific examples, and methods or processes not specifically described in the examples are all prior art.

Example 1

The embodiment is a specific implementation of a vital sign monitoring and alarming system, and the specific working process comprises the following steps:

1. preparing data:

through the mode of the Internet of things, the physical sign monitoring alarm information is digitalized and electronized, and continuous 8-week data are acquired and then enter a database. Dividing the whole database into two sets, wherein the set 1 is a training set and 308 people are admitted to the hospital; set 2 is the test set, with a total of 326 persons admitted.

2. Algorithm design:

1) noninvasive vital sign monitoring data of heart rate, respiratory rate and pulse oximetry as well as systolic and diastolic pressures were recorded at a frequency of 1/20 Hz.

The cardiopulmonary instability risk threshold is set as follows:

heart rate <40 or > 140;

respiratory rate <8 or > 36;

systolic blood pressure <80 or > 200;

diastolic pressure > 110;

pulse oximetry < 85%;

when the monitored corresponding vital sign deviation exceeds the threshold, a critical event of unstable heart and lung is defined, and the critical event needs to be informed by an alarm. In this example 634,137 total crisis occurred in two foci.

In this embodiment, the event is required to initially last for a tolerance of 40 seconds, and if intermittent, for a minimum of 4 minutes or cumulatively for 4 of 5 minutes (80% duty cycle). A 40 second tolerance indicates that if the next violation occurs within 40 seconds of the current violation, they are the same event.

Since the sampling frequency is 1/20Hz, which also means we only allow one normal reading to occur between two abnormal readings, this embodiment results in 2333 qualifying events, 812 in set 1; there are 1521 in set 2.

2) Rules for determining whether an alarm event is defined as authentic or invalid using the set 1 data. The 1/20Hz time plot in the set 1 data is provided to a 3-bit clinical expert reviewer, and the reviewer classifies the alarm data as true or invalid according to clinical experience, thereby establishing a judgment condition of 'basic fact'. In this embodiment, the reviewer annotates 493 (60%) of the 812 events in set 1 as true and 319 (40%) as invalid. In the event of inefficiency, 43% is respiratory rate, 40% is pulse oximetry, 15% is blood pressure, and 2% is heart rate.

3) And extracting characteristic values by adopting an SVM (support vector machine) on the basis of the basic fact. Some rules in the vector machine capture most events of each vital sign and evolve into textual rules describing the multi-signal patterns of the most commonly encountered ineffective events. For digital feature extraction, these text rules are converted into a set of digital features derived from multi-signal time series data.

4) All feature values are refined by one iteration for the selected feature of each vital sign. Events such as tags that do not conform to the characterization rules are reviewed by experts, and then they are relabeled, new features are derived, or text rules are redefined, and their features are redefined. The clinical expert reviewer simultaneously annotates all alarms in set 2 as true or invalid as a test set. And determining the digital characteristics at the end of the characteristic learning step, and judging the usability of the characteristic value by using the AUC and the ROC curve. To reduce the complexity of the machine learning model and the risk of overfitting, as shown in fig. 1, the AUC score of 0.7 is used as an arbitrary boundary point for blood pressure and pulse oximetry, and a respiratory rate boundary point of 0.85 in this embodiment to filter the less informative features, leaving 10 features for pulse oximetry, 5 features for respiratory rate, and 5 features for blood pressure.

3. The algorithm is realized as follows:

various machine learning algorithms are applied to learn and validate the classification model to classify the binary label (true or invalid) of the alarm given the selected feature values. The machine learning algorithm used in this embodiment includes: k nearest neighbors (KNN, at different K), Linear Discriminant Analysis (LDA), naive bayes classifier (NB), Logistic Regression (LR), Support Vector Machine (SVM), and Random Forest (RF), all of which are well-known mechanical learning algorithms, mainly aiming at completing the calculation of feature values and forming definite judgment conditions.

4. And (3) algorithm verification:

in this embodiment, the best classification model applied to different sign information is determined by using ten-fold cross validation in combination with a training algorithm and a validation algorithm AUC score. The selected model is applied to the set 2 data to yield the result. As shown in fig. 2, the AUC scores were from ten-fold cross-validation experiments on an algorithmic array of three classes of vital sign events. Training and validation (ten fold cross validation) on the set 1 (grey line) data is shown, as well as the average AUC score when the algorithm was tested on the set 2 test set data (black line). The best performing algorithm varies depending on the vital sign type: random forests of Respiratory Rate (RR), logistic regression of Blood Pressure (BP), naive bayes of pulse oximetry (SpO 2).

FIG. 2 shows the performance of the best model from three important classes of set 1 data (random forest of RR, logistic regression of BP, and naive Bayes of SpO2) learning, and testing the True Positive Rate (TPR) versus false positive rate (FPR; 7A) and True Negative Rate (TNR) versus false negative rate (FNR, 7B) on set 2 data. The respiratory rate (RR, dark grey line) is the easiest to classify, and the pulse oximetry (SpO2, black line) is the most difficult to distinguish. The blood pressure manifestations (BP, light gray line) are between the two.

Positive refers to true alarm and negative refers to invalid alarm.

The present invention is not limited to the above embodiments, and various combinations and modifications of the above technical features may be provided for those skilled in the art, and modifications, variations, equivalents, or uses of the structure or method of the present invention in other fields without departing from the spirit and scope of the present invention are included in the protection scope of the present invention.

Claims

1. A vital sign monitoring alarm system, characterized by: the working process comprises the following steps: s1, data acquisition, namely measuring vital sign information of a human body through various measuring equipment, collecting various electronic and numerical sign information, and storing data acquired for at least 8 weeks into a database through the Internet of things; dividing a database into two subsets, wherein one subset is a training set, and the other subset is a testing set; s2, summarizing characteristic values of different sign information, and continuously verifying the availability of the characteristic values; s3, forming a real alarm/invalid alarm classification model through machine learning; and S4, verifying the robustness of the classification model through cross-validation combined with a training algorithm and a verification algorithm.

2. The vital sign monitoring alarm system of claim 1, wherein: in step S1, the data size of the training set is equivalent to the data size of the test set.

3. A vital signs monitoring alarm system as claimed in claim 2, wherein: in step S1, non-invasive vital sign monitoring data of heart rate, respiratory rate and pulse oximetry as well as systolic and diastolic pressures are recorded at a frequency of 1/20 Hz.

4. A vital signs monitoring alarm system according to claim 3, wherein: heart rate <40 or > 140; respiratory rate <8 or > 36; systolic blood pressure <80 or > 200; diastolic pressure > 110; pulse oximetry < 85%, defined as a crisis event that requires alarm notification.

5. The vital sign monitoring alarm system of claim 4, wherein: the crisis event is sustained for a tolerance of 40 seconds and, if intermittent, for a minimum of 4 minutes or cumulatively for 4 of 5 minutes.

6. The vital sign monitoring alarm system of claim 1, wherein: in step S2, rules defining alarm events as true or invalid are determined using the training set data, a 1/20Hz time map in the training set data is provided to a reviewer composed of at least 3 clinical experts, and the alarm data is classified as true or invalid by the reviewer according to clinical experience to establish judgment conditions of ground truth.

7. The vital sign monitoring alarm system of claim 6, wherein: and extracting characteristic values by adopting an SVM (support vector machine) on the basis of basic facts.

8. The vital sign monitoring alarm system of claim 7, wherein: all feature values are refined through an iteration as a selected feature of a certain vital sign.

9. The vital sign monitoring alarm system of claim 8, wherein: the availability of the characteristic value is judged by using the AUC and the ROC curve.

10. A vital signs monitoring alarm system according to claim 9, wherein: in step S4, the best classification model applied to different sign information is determined by using the ten-fold cross validation in combination with the training algorithm and the validation algorithm AUC score, and the selected model is applied to the data in the test set to obtain a result.