Alarm identification method and device based on machine learning model, electronic equipment and storage medium

Abstract

The invention provides an alarm identification method and device based on a machine learning model, electronic equipment and a storage medium, and relates to the technical field of medical equipment networking. The method comprises the following steps: receiving monitoring data of medical equipment collected by a collector every time, wherein the monitoring data is collected by the collector for the medical equipment when an alarm is triggered or is collected by the collector for the medical equipment at regular time; if it is determined that the received current monitoring data is collected when the medical equipment triggers the alarm, obtaining a triggering moment when the medical equipment triggers the alarm; and judging whether the medical equipment triggers a clinical alarm or not based on all the monitoring data in the identification time period of the triggering moment. According to the invention, the monitoring data of the medical equipment is collected in real time, and whether the medical equipment triggers the clinical alarm or not is judged based on all the monitoring data in the identification time period of the alarm moment when the medical equipment triggers the alarm, so that the timeliness identification of the clinical alarm of the medical equipment is realized; therefore, the pressure of medical staff on a large number of non-clinical alarms is relieved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical device networking. Specifically, it relates to an alarm recognition method, device, electronic device, and storage medium based on a machine learning model. Background Art

[0002] Medical devices such as ventilators and monitors, as the core devices for maintaining the lives of critically ill patients, their alarm systems are important barriers to ensuring patient safety. Currently, the alarms of such medical devices mainly rely on preset physiological parameter thresholds (such as airway pressure, tidal volume, blood oxygen saturation, etc.) and device operation status detection (such as power supply, gas source, sensor signal integrity), and the alarm categories can be divided into clinical alarms and non-clinical alarms. Among them, clinical alarms may indicate that the patient has some emergency situations, and medical staff need to take corresponding medical measures in a timely manner; non-clinical alarms may be caused by technical failures of the device, and non-clinical alarms may also be false alarms that do not require the attention of medical staff. Relevant investigation results show that the proportion of non-clinical alarms is as high as 85%-99%, which means that medical staff need to distinguish real clinical alarms from a large number of device alarms every day.

[0003] Therefore, how to effectively identify clinical alarms of medical devices is the key to ensuring patient safety and reducing waste of medical resources. Summary of the Invention

[0004] The purpose of the present invention is to provide an alarm recognition method, device, electronic device, and storage medium based on a machine learning model to improve the problems existing in the prior art.

[0005] The embodiments of the present invention can be implemented as follows: In a first aspect, the present invention provides an alarm recognition method based on a machine learning model, including: Receiving the monitoring data of the medical device collected by the collector each time, where the monitoring data is collected by the collector for the medical device when an alarm is triggered or the collector periodically collects the medical device; If it is determined that the currently received monitoring data is collected when the medical device triggers an alarm, obtaining the trigger moment when the medical device triggers the alarm; Based on all the monitoring data within the recognition period where the trigger moment is located, determining whether the medical device triggers a clinical alarm.

[0006] In an optional implementation manner, the step of determining whether the alarm triggered by the medical device is a clinical alarm based on all the monitoring data within the recognition period where the trigger moment is located includes: Take each piece of monitoring data within the recognition period where the trigger moment is located as reference monitoring data; the reference monitoring data includes sampling data of multiple indicators; Normalize the sampling data of each of the multiple specific indicators in each reference monitoring data respectively to obtain a target feature matrix; the multiple specific indicators are selected from all the indicators; Input the target feature matrix into a pre-trained target recognition model to obtain a recognition result; wherein the recognition result reflects whether the medical device triggers a clinical alarm.

[0007] In an alternative embodiment, the target recognition model is obtained through the following steps: Construct an original data set, the original data set includes a number of original samples; the original samples include change feature vectors and alarm label vectors of the multiple indicators within a preset monitoring duration; Based on the original data set, select the multiple specific indicators from all the indicators; Delete from each original sample other content except the change feature vectors and alarm label vectors of each specific indicator to obtain a target data set; Use the target data set to train a number of pre-constructed classification models to obtain a number of candidate classification models; Evaluate the performance of each candidate classification model to determine the target recognition model from the multiple candidate classification models.

[0008] In an alternative embodiment, the step of constructing the original data set includes: Obtain a number of monitoring sets; the monitoring sets include multiple historical sampling data and multiple trigger states of each indicator within the preset monitoring duration; For any one of the monitoring sets, normalize all the historical sampling data of each indicator in the monitoring set respectively to obtain change feature vectors corresponding to each indicator; Label all the trigger states of each indicator in the monitoring set respectively to obtain alarm label vectors corresponding to each indicator; Combine the change feature vectors and alarm label vectors corresponding to each indicator to obtain an original sample corresponding to the monitoring set; Traverse each monitoring set to obtain the original data set.

[0009] In an alternative embodiment, the alarm label vector includes S label values; The step of respectively performing tagging processing on all trigger states of each of the indicators in the monitoring set to obtain an alarm tag vector corresponding to each of the indicators includes: For each of the indicators, obtain the s-th trigger state of the indicator in the monitoring set; If the s-th trigger state is to trigger non-clinical treatment, set the s-th tag value in the alarm tag vector corresponding to the indicator to 1; If the s-th trigger state is to trigger technical treatment or no treatment is triggered, set the s-th tag value in the alarm tag vector corresponding to the indicator to 0, where .

[0010] In an alternative embodiment, the step of screening out the multiple specific indicators from all the indicators based on the original data set includes: Based on the original data set, train multiple algorithm models respectively to obtain multiple trained algorithm models; Obtain the weight parameters of each of the indicators in each of the trained algorithm models respectively to obtain an importance list of each of the indicators; Based on the importance list of each of the indicators, screen out the multiple specific indicators from all the indicators.

[0011] In an alternative embodiment, when the number of algorithm models is M, the importance list includes M kinds of importance coefficients; The step of screening out the multiple specific indicators from all the indicators based on the importance list of each of the indicators includes: Sort all the indicators respectively in descending order according to the M kinds of importance coefficients to obtain M sorting results; After deleting the indicators after the N-th position in the M sorting results, determine the intersection of the M sorting results; Take each indicator in the intersection as the specific indicator.

[0012] In a second aspect, the present invention provides an alarm recognition device based on a machine learning model, including: A receiving module, configured to receive the monitoring data of the medical device collected by the collector each time, where the monitoring data is collected by the collector for the medical device when an alarm is triggered or the collector periodically collects the medical device; The receiving module is further configured to, if it is determined that the currently received monitoring data is collected when the medical device triggers an alarm, obtain the trigger moment when the medical device triggers the alarm; A judgment module, configured to judge whether the medical device triggers a clinical alarm based on all the monitoring data within the recognition period where the trigger moment is located.

[0013] In a third aspect, the present invention further provides an electronic device, including: a memory and a processor, where the memory stores a software program, and when the electronic device runs, the processor executes the software program to implement the method as described in the first aspect.

[0014] In a fourth aspect, the present invention further provides a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method as described in the first aspect is implemented.

[0015] Compared with the prior art, the embodiments of the present invention provide a method, apparatus, electronic device, and storage medium for alarm recognition based on a machine learning model. The method is as follows: receiving the monitoring data of a medical device collected by a collector each time, where the monitoring data is collected by the collector for the medical device when an alarm is triggered or the collector periodically collects the medical device; if it is determined that the currently received monitoring data is the data collected when the medical device triggers an alarm, obtaining the trigger moment when the medical device triggers the alarm; based on all the monitoring data within the recognition period where the trigger moment is located, determining whether the medical device triggers a clinical alarm. The present invention collects the monitoring data of the medical device in real time, and when the medical device triggers an alarm, determines whether the medical device triggers a clinical alarm based on all the monitoring data within the recognition period where the alarm moment is located, realizing the timely recognition of the clinical alarm of the medical device, thereby reducing the pressure on medical staff faced with a large number of non-clinical alarms. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0017] Figure 1 It is a schematic flowchart of a method for alarm recognition based on a machine learning model provided by an embodiment of the present invention.

[0018] Figure 2 It is a schematic flowchart of the training process of a target recognition model provided by an embodiment of the present invention.

[0019] Figure 3 It is a schematic structural diagram of an alarm recognition device based on a machine learning model provided by an embodiment of the present invention.

[0020] Figure 4 It is a schematic structural diagram of an electronic device provided by an embodiment of the present invention.

[0021] Icons: 200 - Alarm recognition device based on machine learning model; 210 - Receiving module; 220 - Judgment module; 300 - Electronic device; 310 - Processor; 320 - Memory; 330 - Bus. Detailed implementation manners

[0022] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some but not all of the embodiments of the present invention. The components of the embodiments of the present invention described and illustrated herein generally may be arranged and designed in a variety of different configurations.

[0023] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0024] It should be noted that like reference numerals and letters denote like items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0025] In addition, terms such as "first" and "second" are used for descriptive purposes only and cannot be construed as indicating or implying relative importance.

[0026] It should be noted that the features in the embodiments of the present invention may be combined with each other without conflict.

[0027] Please refer to Figure 1 , Figure 1 which is a schematic flowchart of a model training method provided by an embodiment of the present invention. The execution subject of this method may be, but is not limited to: electronic devices such as smart phones, personal notebooks, personal computers, servers, etc. As Figure 1 , this method includes the following steps S6 to S8: S6. Receive the monitoring data of the medical device collected by the collector each time.

[0028] In this embodiment, the monitoring data is collected by the collector for the medical device when an alarm is triggered or the collector periodically collects the medical device.

[0029] Taking a medical device such as a monitor as an example, the interval period T for the monitor to collect data regularly can be 10s, 20s, 30s, etc. That is, when the interval period T = 10s, in addition to collecting and reporting data every 10s, the collector will also collect and report data each time the monitor triggers an alarm.

[0030] Among them, the collector can be a device specifically designed for monitoring and collecting medical devices. The collector and the medical device can be connected through a USB interface, an Ethernet interface, a type-C interface, etc.

[0031] Exemplarily, the collector can be a modular device chip developed based on STM32F407VET6, equipped with an ESP8266 WiFi module and a LAN8720A Ethernet core, which can connect to the Ethernet and wireless networks simultaneously. The Ethernet connects to the medical device local area network for data collection, and the wireless network uses WiFi to connect to the Internet for uploading data. The collection terminal can achieve WiFi positioning; inside the collector, WiFi positioning is realized in cooperation with the server side, and the main controller directly obtains the wireless AP information from the WiFi network card module. The server side determines the wireless AP device connected by the device by identifying the AP signal mark, and all wireless APs within the networking range are marked with specific positions, thereby determining the device position. When the collector is connected to a medical device without an Ethernet interface, the MAX3232 chip is used to convert the main controller TTL serial port into a 232 serial port to adapt to the additional general 9-core serial port of the medical device. The collector integrates a Winbond W25Q32 SPI Flash storage chip with a capacity of 4MB, which can cache 4000 offline data, that is, it can cache about 10 hours of offline data.

[0032] This example is only for illustration, and the embodiments of the present invention do not limit the integrated structure of the collector.

[0033] S7. If it is determined that the currently received monitoring data is collected when the medical device triggers an alarm, obtain the trigger moment when the medical device triggers the alarm.

[0034] Optionally, each time the collector uploads the monitoring data, it can also upload the trigger flag and the collection moment at the same time. The trigger flag being 0 indicates that the monitoring data is collected regularly by the collector, and in this embodiment, being 1 indicates that the monitoring data is collected when the medical device is sensed to trigger an alarm; In this embodiment, when the currently received monitoring data is received, if the simultaneously received trigger flag is 1, it means that the currently received monitoring data is collected when the medical device triggers an alarm, and the simultaneously received collection moment can be used as the trigger moment when the medical device triggers the alarm.

[0035] S8. Based on all the monitoring data within the recognition period where the trigger moment is located, determine whether the medical device triggers a clinical alarm.

[0036] In this embodiment, the recognition period where the trigger moment is located includes: the period of the first set duration before the trigger moment and the period of the second set duration after the trigger moment. All the monitoring data within the recognition period where the trigger moment is located reflect the operation of the medical device during a period of time before and after the trigger of the alarm. Therefore, it is possible to determine whether the medical device triggers a clinical alarm based on this.

[0037] Taking a monitor as an example, assume that the interval period T for the monitor to collect data regularly is 30s. The first set duration and the second set duration can be the same (for example, both are 2 minutes or 5 minutes), or they can be different (for example, one is 3 minutes and the other is 2 minutes). Therefore, within the recognition period where the trigger moment is located, the number of monitoring data is not fixed and is related to the size of the interval period T, the size of the first set duration, and the size of the second set duration. This example is only for illustration and is not limited here.

[0038] The alarm recognition method based on a machine learning model provided by the embodiments of the present invention realizes the timely recognition of clinical alarms for medical devices by collecting the monitoring data of medical devices in real time and determining whether the medical device triggers a clinical alarm based on all the monitoring data within the recognition period where the alarm moment is located, thereby reducing the pressure on medical staff faced with a large number of non-clinical alarms.

[0039] In the present invention, the collected monitoring data includes the sampling data of multiple indicators of the medical device. Among all the indicators, some are device-related indicators and some are patient-related indicators.

[0040] Exemplarily, taking a monitor as an example, the device-related indicators include: the operating state of the device (working, standby, shutdown, etc.), the set parameters of the device, the operating state (normal, device failure, etc.), the machine ID, the device type, the device model, the device manufacturer, etc.; while the patient-related indicators include: heart rate, peripheral oxygen saturation (SpO2), non-invasive blood pressure (NIBP), respiratory rate (RR), body temperature (T), electrocardiogram (ECG), photoplethysmography (PPG), invasive blood pressure (IBP), central venous pressure (CVP), as well as the alarm type, alarm level, and alarm duration, etc.

[0041] This example is only for illustration and is not limited here.

[0042] In an optional implementation, "judging whether the alarm triggered by the medical device is a clinical alarm" can be implemented in the following way: Set one or more clinically relevant threshold ranges for each specific indicator. These threshold ranges can be determined in advance based on historical experience data or the knowledge of domain experts. The specific indicators can be selected depending on the experience of medical staff; Regard each piece of monitoring data within the recognition period where the triggering moment is located as reference monitoring data; subsequently, check one by one whether the sampling data of the specific indicator in each reference monitoring data exceeds its corresponding threshold range, and then count the number of times the specific indicator exceeds the standard in all reference monitoring data; If the number of times a specific indicator exceeds the standard exceeds a preset number, it can be considered that this indicator may be related to a clinical alarm, that is, the alarm triggered by the medical device belongs to a clinical alarm.

[0043] This rule-based threshold comparison method relies too much on accurate threshold setting and the screening of specific indicators. Although it is simple to implement, both the threshold setting and the selection of specific indicators rely on human experience, and the solution has limitations.

[0044] Therefore, in another optional implementation, the screening of specific indicators can be realized by using machine learning in advance, and a target recognition model can be trained in advance, which is specifically used to judge whether the alarm triggered by the above-mentioned medical device is a clinical alarm. In this way, accurate judgment can be realized based on machine learning without relying on human experience for threshold setting and specific indicator selection. Therefore, for the above step S8, its sub-steps can include S81 to S83.

[0045] S81. Regard each piece of monitoring data within the recognition period where the triggering moment is located as reference monitoring data.

[0046] In this embodiment, each reference monitoring data includes sampling data of multiple indicators.

[0047] S82. Normalize the sampling data of each of the multiple specific indicators in each reference monitoring data respectively to obtain a target feature matrix.

[0048] In this embodiment, the multiple specific indicators can be screened from all indicators in advance based on machine learning algorithms. The specific indicators are indicators that have a strong correlation with the occurrence of clinical alarms. Since the dimensions of each specific indicator are different, normalization is used to eliminate the influence of different dimensions or numerical ranges on the subsequent model input, thereby improving the accuracy and stability of model prediction. Through normalization, the sampling data of each specific indicator in all reference monitoring data can be converted into feature vectors under a unified scale, and finally a target feature matrix is formed.

[0049] S83. Input the target feature matrix into the pre-trained target recognition model to obtain the recognition result.

[0050] In this embodiment, the recognition result reflects whether the medical device triggers a clinical alarm.

[0051] The training process of the target recognition model is introduced below. Please refer to Figure 2 , the training process of the target recognition model may include the following steps S1 to S5: S1. Construct the original data set.

[0052] Among them, the original data set includes a number of original samples, and each original sample may include the change feature vectors and alarm label vectors of multiple indicators within a preset monitoring duration.

[0053] Optionally, the process of constructing the original data set may include the following sub-steps S11 to S15.

[0054] S11. Obtain a number of monitoring sets.

[0055] In this embodiment, a number of historical data of the medical device can be collected. The historical data includes historical monitoring data, trigger status, and collection time. Then, based on the collection time, all historical data can be grouped to obtain a number of monitoring sets. The grouping principle is: ensure that the difference in collection time between any two historical data in each monitoring set is less than or equal to the preset monitoring duration.

[0056] In this way, each monitoring set after grouping may include multiple historical sampling data and multiple trigger statuses of each indicator within the preset monitoring duration.

[0057] Among them, if a piece of historical monitoring data triggers an alarm when the medical device is collected, and after medical personnel identify that the alarm belongs to a clinical alarm, then the trigger status is triggering non-clinical treatment; if a piece of historical monitoring data triggers an alarm when the medical device is collected, and after medical personnel identify that the alarm does not belong to a clinical alarm, then the trigger status is triggering non-clinical treatment; if a piece of historical monitoring data does not trigger an alarm when the medical device is collected, then the trigger status is not triggered for treatment. Therefore, the trigger status is divided into three types: triggering non-clinical treatment, triggering technical treatment, and not triggered for treatment.

[0058] S12. For any monitoring set, normalize all historical sampling data of each indicator in the monitoring set to obtain a change feature vector corresponding to each indicator.

[0059] In this embodiment, for monitoring set 1, normalize all historical sampling data of an indicator A in monitoring set 1 to obtain a change feature vector corresponding to this indicator A.

[0060] S13. Label all trigger states of each indicator in the monitoring set to obtain an alarm label vector corresponding to each indicator.

[0061] Suppose an alarm label vector includes S label values. Then, for any indicator, the process of labeling all trigger states of this indicator in the monitoring set to obtain an alarm label vector corresponding to this indicator includes the following steps S131 - S133: S131. Obtain the s-th trigger state of the indicator in the monitoring set. S132. If the s-th trigger state is to trigger non-clinical treatment, set the s-th label value in the alarm label vector corresponding to the indicator to 1. S133. If the s-th trigger state is to trigger technical treatment or not to trigger treatment, set the s-th label value in the alarm label vector corresponding to the indicator to 0, where .

[0062] For monitoring set 1, execute steps S131 - S133 for each indicator to obtain an alarm label vector corresponding to each indicator A.

[0063] S14. Combine the change feature vectors and alarm label vectors corresponding to each indicator to obtain an original sample corresponding to the monitoring set.

[0064] In this embodiment, combine the change feature vectors and alarm label vectors corresponding to each indicator obtained by processing a monitoring set according to the above steps S11 - S14, that is, convert this monitoring set into an original sample.

[0065] Exemplarily, taking 5 indicators as an example (denoted as indicator 1 to indicator 5), suppose monitoring set A includes 8 historical sampling data and 8 trigger states corresponding to each indicator. First, for each indicator in monitoring set A, normalize all 8 historical sampling data respectively. The purpose of the normalization operation is to unify historical sampling data with different dimensions or value ranges into a standardized range, thereby reducing the impact of data differences on subsequent model training. After this processing, each indicator generates a change feature vector, which is composed of 8 normalized historical sampling data; Subsequently, for each metric in the monitoring set A, its 8 trigger states also need to be tagged to generate an alarm label vector. Specifically, if a trigger state corresponds to triggering a clinical disposition, the corresponding label value is set to 1; otherwise, if the trigger state corresponds to triggering a technical disposition or not triggering any disposition, the corresponding label value is set to 0. Finally, each metric generates an alarm label vector containing 8 label values (e.g., 00000001); Next, combine the change feature vectors corresponding to the 5 metrics with the alarm label vectors respectively, and a complete original sample is obtained. That is, the obtained original sample contains two parts of information: one part is the change feature vectors of each metric within the preset monitoring duration, which reflects the dynamic change of each coordinate over a period of time; the other part is the alarm label vectors corresponding to each coordinate, which reflects the clinical alarm situation of each metric over a period of time.

[0066] The above examples are only for illustration and are not limited here.

[0067] S15. Traverse each monitoring set to obtain the original data set.

[0068] In this embodiment, each monitoring set is processed through the above steps S11~S14, and several original samples are obtained, completing the construction of the original data set.

[0069] The creation process of the above original sample systematically integrates the dynamic change situation of various metrics during the operation of medical devices over a period of time and the clinical alarm triggering situation. Therefore, the created original data set directly determines the quality of subsequent model training and recognition effects.

[0070] S2. Based on the original data set, select multiple specific metrics from all metrics.

[0071] In this embodiment, selecting multiple specific metrics from all metrics aims to eliminate redundant or low-correlation metrics with clinical alarms through algorithm analysis, thereby improving the computational efficiency and prediction accuracy of the model.

[0072] In a feasible implementation, a mathematical statistics method can be used to calculate the correlation coefficient (such as Spearman rank correlation coefficient) between each metric and clinical alarms, so as to select specific metrics with higher correlation coefficients.

[0073] In another feasible implementation, metric screening can also be achieved based on machine learning algorithms to effectively identify specific metrics highly correlated with clinical alarms. Correspondingly, the sub-steps of step S2 can include S21~S23.

[0074] S21. Based on the original dataset, train multiple algorithm models respectively to obtain multiple trained algorithm models.

[0075] Optionally, the multiple algorithm models may include classification models such as LASSO (Least Absolute Shrinkage and Selection Operator) model, Random Forest (RF) model, Catboost (Categorical Boosting) model, etc. The training process of such classification models is prior art and will not be elaborated here.

[0076] S22. Obtain the weight parameters of each indicator in each trained algorithm model respectively to obtain the importance list of each indicator.

[0077] In this embodiment, the importance list of each indicator may include the importance coefficient of the indicator in each trained algorithm model.

[0078] Taking the trained LASSO model as an example, the feature coefficient of each indicator in the trained LASSO model is the importance coefficient.

[0079] Taking the trained random forest model as an example, for any indicator, the reduction in impurity of each node in each tree of the trained random forest model for this indicator can be calculated, and then the sum is the importance coefficient of this indicator.

[0080] Then, for a node m in a tree of the trained random forest model, if the Gini algorithm is used, the Gini impurity corresponding to node m is:

[0081] where k represents the number of indicators involved in the data subset (obtained by splitting the original dataset), represents the data subset the proportion of the indicator in. reflects the degree of confusion of the data subset the smaller the value, the purer the data.

[0082] Thus, the calculation formula for the importance coefficient of the indicator is:

[0083]

[0084] where represents the indicator The impurity reduction amount at node m, where T is the set of all trees in the trained random forest model, is the set of all nodes in tree t. represents the Gini impurity of the parent node of node m, represents the Gini impurity of the left child node of node m, represents the Gini impurity of the right child node of node m.

[0085] Taking the trained Catboost model as an example, the importance coefficient of each metric in the trained LASSO model is: the sum of the information gains of each metric for all split nodes of the model.

[0086] The above examples are only for illustration and are not limited here.

[0087] S23. Screen out multiple specific metrics from all metrics based on the importance list of each metric.

[0088] Optionally, when the number of algorithm models is M, then the importance list includes M kinds of importance coefficients. At this time, specific metrics can be screened out in the following way: (1) Sort all metrics in descending order according to M kinds of importance coefficients respectively to obtain M sorting results; (2) After deleting the metrics after the Nth position in the M sorting results, determine the intersection of the M sorting results, and take each metric in the intersection as a specific metric.

[0089] The above steps S21 - S23 use the original dataset to train M algorithm models, extract the importance coefficients of each metric in each model, and then sort all metrics in descending order according to the M importance coefficients, so as to select specific metrics with relatively high comprehensive rankings. In this way, this method of joint importance analysis of multiple models effectively avoids the possible biases or limitations in importance analysis by a single model.

[0090] S3. Delete all other contents in each original sample except the change feature vectors and alarm label vectors of each specific metric to obtain the target dataset.

[0091] In this embodiment, delete all other contents in each original sample except the change feature vectors and alarm label vectors of each specific metric to form the target dataset. The generation of the target dataset ensures a high degree of correlation and necessity between the data used for model training and clinical alarms, avoids unnecessary noise interference, and thus ensures that the trained target recognition model can make accurate judgments.

[0092] S4. Use the target dataset to train multiple pre - constructed classification models to obtain multiple candidate classification models.

[0093] In this embodiment, the classification model may include, but is not limited to: one or a combination of Gaussian Naive Bayes (GNB), K-Nearest Neighbors (KNN), Linear Discriminant Analysis (LDA), Classification and Regression Trees (CART), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), Logistic Regression (LR), Random Forest (RF), etc.

[0094] S5. Perform performance evaluation on each candidate classification model to determine the target recognition model from multiple candidate classification models.

[0095] In this embodiment, after performing performance evaluation on each candidate classification model, the target recognition model with the best performance can be selected from multiple candidate classification models. Among them, model evaluation can start from aspects such as the Area Under the Receiver Operating Characteristic Curve (AUC-ROC), accuracy, sensitivity, specificity, Positive Predictive Value (PPV), Negative Predictive Value (NPV), F1-Score, Kappa value, etc.

[0096] Exemplarily, assume that the candidate classification models include XGBoost, Light BGM, and a fusion model (including three models: XGBoost, Light BGM, and SVM). Among them, the working process of the fusion model is as follows: after XGBoost and Light BGM make predictions respectively, the prediction results of the two are input into SVM for fusion prediction.

[0097] The inventor performed performance evaluation on the three candidate classification models of XGBoost, Light BGM, and the fusion model, and the evaluation results are shown in Table 1 below: Table 1 Evaluation Results

[0098] After comprehensively analyzing the performance evaluation results, it can be determined that the fusion model is the model with the best virtual performance, that is, the fusion model can be used as the target recognition model.

[0099] This example is only for illustration and is not limited here.

[0100] The present invention uses a target data set to train a plurality of pre-constructed classification models to obtain a plurality of candidate classification models, and then selects the candidate classification model with the best performance as the target recognition model. In this way, it is determined that the finally obtained target recognition model has an excellent recognition effect on clinical alarms.

[0101] Please refer to Figure 3 , Figure 3 which shows a schematic structural diagram of an alarm recognition device based on a machine learning model provided by an embodiment of the present invention. The alarm recognition device 200 based on the machine learning model includes: a receiving module 210 and a judging module 220.

[0102] The receiving module 210 is configured to receive the monitoring data of the medical device collected by the collector each time. The monitoring data is collected by the collector for the medical device when an alarm is triggered or the collector periodically collects the medical device; The receiving module 210 is further configured to, if it is determined that the currently received monitoring data is collected when the medical device triggers an alarm, obtain the trigger time when the medical device triggers the alarm; The judging module 220 is configured to judge whether the medical device triggers a clinical alarm based on all the monitoring data within the recognition period where the trigger time is located.

[0103] Optionally, in the process of judging whether the alarm triggered by the medical device is a clinical alarm based on all the monitoring data within the recognition period where the trigger time is located, the judging module 220 may specifically be configured to: use each piece of monitoring data within the recognition period where the trigger time is located as the reference monitoring data; the reference monitoring data includes the sampling data of multiple indicators; perform normalization processing on the sampling data of each of the multiple specific indicators in each reference monitoring data respectively to obtain a target feature matrix; the multiple specific indicators are selected from all the indicators; input the target feature matrix into a pre-trained target recognition model to obtain a recognition result; wherein the recognition result reflects whether the medical device triggers a clinical alarm.

[0104] Optionally, an embodiment of the present invention further provides a training module 100, which can be used to: construct an original data set, where the original data set includes a number of original samples; the original samples include change feature vectors and alarm label vectors of multiple metrics within a preset monitoring duration; based on the original data set, select multiple specific metrics from all the metrics; delete other content in each original sample except the change feature vectors and alarm label vectors of each specific metric to obtain a target data set; use the target data set to train a number of pre-constructed classification models to obtain a number of candidate classification models; perform performance evaluation on each candidate classification model to determine a target recognition model from the number of candidate classification models.

[0105] Optionally, during the process of constructing the original data set, the training module 100 can specifically be used to: obtain a number of monitoring sets; the monitoring sets include multiple historical sampling data and multiple trigger states of each metric within a preset monitoring duration; for any monitoring set, perform normalization processing on all the historical sampling data of each metric in the monitoring set to obtain change feature vectors corresponding to each metric; perform labeling processing on all the trigger states of each metric in the monitoring set to obtain alarm label vectors corresponding to each metric; combine the change feature vectors and alarm label vectors corresponding to each metric to obtain an original sample corresponding to the monitoring set; traverse each monitoring set to obtain the original data set.

[0106] Optionally, the alarm label vector includes S label values; during the process of performing labeling processing on all the trigger states of each metric in the monitoring set to obtain the alarm label vector corresponding to each metric, the training module 100 can specifically be used to: for each metric, obtain the s-th trigger state of the metric in the monitoring set; if the s-th trigger state is triggering non-clinical treatment, set the s-th label value in the alarm label vector corresponding to the metric to 1; if the s-th trigger state is triggering technical treatment or not triggering treatment, set the s-th label value in the alarm label vector corresponding to the metric to 0, where 。

[0107] Optionally, during the process of selecting multiple specific metrics from all the metrics based on the original data set, the training module 100 can specifically be used to: based on the original data set, train multiple algorithm models respectively to obtain multiple trained algorithm models; obtain the weight parameters of each metric in each trained algorithm model respectively to obtain an importance list of each metric; based on the importance list of each metric, select multiple specific metrics from all the metrics.

[0108] Optionally, when the number of algorithm models is M, the importance list includes M kinds of importance coefficients; in the process of screening multiple specific indicators from all indicators based on the importance list of each indicator, the training module 100 can specifically be used for: sorting all indicators respectively in descending order according to the M kinds of importance coefficients to obtain M sorting results; after deleting the indicators after the Nth position in the M sorting results, determining the intersection of the M sorting results; and using each indicator in the intersection as a specific indicator.

[0109] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the above-described alarm recognition device 200 and training module 100 based on the machine learning model can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0110] Please refer to Figure 4 , Figure 4 which is a schematic structural diagram of an electronic device provided by an embodiment of the present invention. The electronic device 300 includes a processor 310, a memory 320, and a bus 330. The processor 310 is connected to the memory 320 through the bus 330.

[0111] The memory 320 can be used to store software programs. For example, the software program corresponding to the alarm recognition device 200 based on the machine learning model provided by the embodiment of the present invention. The processor 310 executes various functional applications and data processing by running the software program stored in the memory 320 to implement the alarm recognition method based on the machine learning model provided by the embodiment of the present invention.

[0112] Among them, the memory 320 can be, but is not limited to: RAM (Random Access Memory), ROM (Read Only Memory), FLASH (Flash Memory), PROM (Programmable Read-Only Memory), EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electric Erasable Programmable Read-Only Memory), etc.

[0113] The processor 310 can be an integrated circuit chip with signal processing capabilities. The processor 310 can be a general-purpose processor, including: CPU (Central Processing Unit), NP (Network Processor), SoC (System on Chip), etc.; it can also be: DSP (Digital Signal Processing), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0114] It can be understood that Figure 4 The structure shown is only schematic, and the electronic device 300 may also include more or fewer components than Figure 4 shown therein, or have a configuration different from Figure 4 that shown. Figure 4 Each component shown therein can be implemented by hardware, software, or a combination thereof.

[0115] The embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, it implements the alarm recognition method based on a machine learning model disclosed in the above embodiment. The computer-readable storage medium can be, but is not limited to: various media such as USB flash drives, external hard drives, ROM, RAM, PROM, EPROM, EEPROM, FLASH disks, or optical discs that can store program codes.

[0116] In summary, the embodiment of the present invention provides an alarm recognition method, device, electronic device, and storage medium based on a machine learning model. The method is as follows: receiving the monitoring data of the medical device collected by the collector each time, where the monitoring data is collected by the collector for the medical device when an alarm is triggered or the collector periodically collects the medical device; if it is determined that the currently received monitoring data is collected when the medical device triggers an alarm, obtaining the trigger moment when the medical device triggers the alarm; based on all the monitoring data within the recognition period where the trigger moment is located, determining whether the medical device triggers a clinical alarm. The present invention collects the monitoring data of the medical device in real time and determines whether the medical device triggers a clinical alarm based on all the monitoring data within the recognition period when the medical device triggers an alarm, realizing the timely recognition of the clinical alarm of the medical device, thereby reducing the pressure on medical staff facing a large number of non-clinical alarms.

[0117] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. An alarm recognition method based on a machine learning model, characterized in that, Including: Receiving the monitoring data of the medical device collected each time by the collector, where the monitoring data is collected by the collector for the medical device when an alarm is triggered or the collector periodically collects the medical device; If it is determined that the currently received monitoring data is collected when the medical device triggers an alarm, obtaining the trigger moment when the medical device triggers the alarm; Based on all the monitoring data within the identification period where the trigger moment is located, determining whether the medical device triggers a clinical alarm.

2. The method according to claim 1, characterized in that, The step of determining whether the alarm triggered by the medical device is a clinical alarm based on all the monitoring data within the identification period where the trigger moment is located includes: Regarding each piece of monitoring data within the identification period where the trigger moment is located as reference monitoring data; the reference monitoring data includes sampling data of multiple indicators; Respectively performing normalization processing on the sampling data of each of the multiple specific indicators in each reference monitoring data to obtain a target feature matrix; the multiple specific indicators are selected from all the indicators; Inputting the target feature matrix into a pre-trained target recognition model to obtain a recognition result; where the recognition result reflects whether the medical device triggers a clinical alarm.

3. The method according to claim 2, wherein The target recognition model is obtained through the following method: Constructing an original data set, where the original data set includes a number of original samples; the original samples include the change feature vectors and alarm label vectors of the multiple indicators within a preset monitoring duration; Based on the original data set, screening out the multiple specific indicators from all the indicators; Deleting other contents in each original sample except the change feature vectors and alarm label vectors of each specific indicator to obtain a target data set; Using the target data set to train a number of pre-constructed classification models to obtain a number of candidate classification models; Performing performance evaluation on each candidate classification model to determine the target recognition model from the multiple candidate classification models.

4. The method according to claim 3, wherein The step of constructing the original data set includes: Obtaining a number of monitoring sets; the monitoring sets include multiple historical sampling data and multiple trigger states of each indicator within the preset monitoring duration; For any one of the monitoring sets, respectively performing normalization processing on all the historical sampling data of each indicator in the monitoring set to obtain the change feature vectors corresponding to each indicator; Respectively performing labeling processing on all the trigger states of each indicator in the monitoring set to obtain the alarm label vectors corresponding to each indicator; Combining the change feature vectors and alarm label vectors corresponding to each indicator to obtain the original sample corresponding to the monitoring set; Traversing each monitoring set to obtain the original data set.

5. The method according to claim 4, wherein The alarm label vector includes S label values; The step of respectively performing labeling processing on all the trigger states of each indicator in the monitoring set to obtain the alarm label vectors corresponding to each indicator includes: For each indicator, obtaining the s-th trigger state of the indicator in the monitoring set; If the s-th trigger status is to trigger non-clinical treatment, set the s-th tag value in the alarm tag vector corresponding to the indicator to 1; If the s-th trigger state is to trigger technical handling or not to trigger handling, set the s-th tag value in the alarm tag vector corresponding to the indicator to 0, where, .

6. The method according to claim 3, wherein The step of screening out the multiple specific indicators from all the indicators based on the original data set includes: Based on the original data set, train multiple algorithm models respectively to obtain multiple trained algorithm models; Obtain the weight parameters of each indicator in each trained algorithm model respectively to obtain the importance list of each indicator; Based on the importance list of each indicator, screen out the multiple specific indicators from all the indicators.

7. The method according to claim 6, characterized in that When the number of the algorithm models is M, the importance list includes M kinds of importance coefficients; The step of screening out the multiple specific indicators from all the indicators based on the importance list of each indicator includes: Sort all the indicators respectively in descending order according to the M kinds of importance coefficients to obtain M sorting results; After deleting the indicators after the N-th position in the M sorting results, determine the intersection of the M sorting results; Take each indicator in the intersection as the specific indicator.

8. An alarm recognition device based on a machine learning model, characterized in that, It includes: A receiving module, configured to receive the monitoring data of the medical device collected by the collector each time, where the monitoring data is collected by the collector for the medical device when an alarm is triggered or the collector periodically collects the medical device; The receiving module is further configured to, if it is determined that the currently received monitoring data is collected when the medical device triggers an alarm, obtain the trigger moment when the medical device triggers the alarm; A judgment module, configured to judge whether the medical device triggers a clinical alarm based on all the monitoring data within the identification period where the trigger moment is located.

9. An electronic device, characterized in that, It includes: A memory and a processor, where the memory stores a software program, and when the electronic device runs, the processor executes the software program to implement the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the method according to any one of claims 1-7.

Images