Model training method, alarm classification method and device based on machine learning model, electronic equipment and storage medium

Abstract

The application provides a model training method, an alarm classification method and device based on a machine learning model, electronic equipment and a storage medium, and relates to the technical field of medical device networking. After a key data set is constructed, the key data set can be used to train a plurality of preset algorithm models, so as to screen a target classification model from a plurality of obtained candidate classification models, and then the target classification model can be used to identify whether an alarm generated by a medical device belongs to a clinical alarm, thereby reducing the pressure of medical staffs who face a large number of non-clinical alarms, and the target classification model is used to quickly identify the clinical alarm, so that medical staffs can quickly take remedial measures.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical device networking, in particular to a model training method, an alarm classification method and device based on a machine learning model, an electronic device and a storage medium. BACKGROUND

[0002] At present, in an intensive care unit (ICU), medical devices such as ventilators and monitors are needed to maintain and monitor the vital signs of patients. Relevant survey results show that as high as 85%-99% of alarms of such medical devices are either false alarms or technical alarms irrelevant to the actual clinical condition of the patient. In fact, only a small part of the alarms are clinical alarms related to the clinical condition of the patient. In this case, medical staff need to distinguish the real clinical alarms from a large number of device alarms every day.

[0003] Therefore, how to effectively identify the clinical alarms of medical devices is the key to the safety of patients and the reduction of waste of medical resources. SUMMARY

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

[0005] Embodiments of the present application can be implemented as follows:

[0006] In a first aspect, the present application provides a model training method, comprising:

[0007] constructing a key data set, each key sample in the key data set reflecting whether a clinical alarm occurs in a set observation duration of a medical device;

[0008] training a plurality of algorithm models preset by using the key data set to screen a target classification model from a plurality of candidate classification models obtained, the target classification model being used to determine whether an alarm generated by the medical device belongs to a clinical alarm.

[0009] In a second aspect, the present application further provides an alarm classification method based on a machine learning model, comprising:

[0010] receiving acquisition data of a medical device acquired by an acquisition device each time, the acquisition data comprising state data and a state flag value of a plurality of monitoring indicators; wherein the state flag value of 0 represents that the acquisition data is acquired by the acquisition device at a regular time, and the state flag value of 1 represents that the acquisition data is acquired when the medical device triggers an alarm;

[0011] If the state flag value of any of the monitoring indicators in the received collection data is 1, the time after waiting for a preset time length is taken as the judgment time;

[0012] Based on all the collection data within the preset observation time length before the judgment time, a to-be-recognized feature matrix is generated;

[0013] The to-be-recognized feature matrix is input into a target classification model to obtain a classification result; wherein the classification result reflects whether the alarm triggered by the medical device is a clinical alarm or a non-clinical alarm, and the target classification model is obtained by model training according to the method of the first aspect.

[0014] In a third aspect, the present application further provides a model training device, comprising:

[0015] A data set construction module is configured to construct a key data set, wherein each key sample in the key data set reflects whether a clinical alarm occurs in a medical device within a preset observation time length;

[0016] A model screening module is configured to train a plurality of preset algorithm models using the key data set, so as to screen a target classification model from a plurality of obtained candidate classification models, wherein the target classification model is used to determine whether an alarm generated by the medical device belongs to a clinical alarm.

[0017] In a fourth aspect, the present application further provides an alarm classification device based on a machine learning model, comprising:

[0018] A data receiving module is configured to receive collection data of a medical device collected by a collection device each time, wherein the collection data comprises state data and state flag values of a plurality of monitoring indicators; wherein the state flag value of 0 represents that the collection data is collected by the collection device at a regular time, and the state flag value of 1 represents that the collection data is collected when the medical device triggers an alarm;

[0019] A vectorization module is configured to, when the state flag value of any of the monitoring indicators in the received collection data is 1, take the time after waiting for a preset time length as the judgment time;

[0020] The vectorization module is further configured to generate a to-be-recognized feature matrix based on all the collection data within the preset observation time length before the judgment time;

[0021] An alarm classification module is configured to input the to-be-recognized feature matrix into a target classification model to obtain a classification result; wherein the classification result reflects whether the alarm triggered by the medical device is a clinical alarm or a non-clinical alarm, and the target classification model is obtained by model training according to the method of the first aspect.

[0022] In a fifth aspect, the present application also provides an electronic device, comprising a memory and a processor, wherein the memory stores a software program, and the processor executes the software program to implement the method according to the first aspect or the method according to the second aspect.

[0023] In a sixth aspect, the present application also provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the method according to the first aspect or the method according to the second aspect.

[0024] Compared with the prior art, the embodiments of the present application provide a model training method, an alarm classification method and device based on a machine learning model, an electronic device and a storage medium. After a key data set is constructed, the key data set can be used to train a plurality of preset algorithm models, so as to screen a target classification model from a plurality of candidate classification models, and then the target classification model can be used to identify whether an alarm generated by a medical device belongs to a clinical alarm, so as to reduce the pressure of medical staffs facing a large number of non-clinical alarms, and the target classification model is used to quickly identify the clinical alarm, so as to facilitate the medical staffs to quickly take rescue measures. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0026] Figure 1 A flowchart of a model training method provided by the embodiments of the present application.

[0027] Figure 2 A distribution diagram of collection time of multiple historical collection data in an observation set provided by the embodiments of the present application.

[0028] Figure 3 A sorting diagram of important coefficients of multiple monitoring indexes provided by the embodiments of the present application.

[0029] Figure 4 Performance evaluation results of multiple algorithm models provided by the embodiments of the present application.

[0030] Figure 5 A flowchart of an alarm classification method based on a machine learning model provided by the embodiments of the present application.

[0031] Figure 6A functional block diagram of the model training device provided in an embodiment of the present invention.

[0032] Figure 7 This is a functional block diagram of an alarm classification device based on a machine learning model provided in an embodiment of the present invention.

[0033] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0035] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0036] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0037] Furthermore, the terms "first" and "second" are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0038] It should be noted that, where there is no conflict, the features in the embodiments of the present invention can be combined with each other.

[0039] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a model training method provided in an embodiment of the present invention. This method can be applied to electronic devices such as smartphones, personal laptops, personal computers, and servers. Figure 1 The method includes the following steps S101~S102.

[0040] S101. Construct the key dataset.

[0041] In this embodiment, the key dataset may include several key samples, each of which can reflect whether the medical device has issued a clinical alarm within a set observation period.

[0042] The medical device can be, but is not limited to, a ventilator, a monitor, and the like, which monitor the vital signs of a patient. When the medical device generates a clinical alarm, it indicates that the patient has a relatively urgent condition, and medical personnel need to take remedial measures.

[0043] In S102, the preset multiple algorithm models are trained by using the key data set, so as to screen the target classification model from the obtained multiple candidate classification models.

[0044] In this embodiment, the target classification model is used to determine whether the alarm generated by the medical device belongs to a clinical alarm.

[0045] The model training method provided by the application can train the preset multiple algorithm models by using the key data set after the key data set is constructed, so as to screen the target classification model from the obtained multiple candidate classification models, and then the target classification model can be used to determine whether the alarm generated by the medical device belongs to a clinical alarm, so as to reduce the pressure of medical personnel facing a large number of non-clinical alarms, and the target classification model can be used to quickly identify the clinical alarm, so as to facilitate medical personnel to quickly take remedial measures.

[0046] In an optional implementation, in S101, the process of constructing the key data set can include the following steps S1011-S1013.

[0047] In S1011, an initial data set is constructed.

[0048] In this embodiment, the initial data set includes a plurality of initial samples, and each initial sample includes an initial feature matrix and a corresponding initial label matrix.

[0049] The initial feature matrix includes a plurality of time sequence feature vectors of monitoring indexes within a set observation time length, and one time sequence feature vector reflects the change of one monitoring index within the set observation time length. The initial label matrix includes an alarm label vector corresponding to each monitoring index, and one alarm label vector reflects the situation of triggering a clinical alarm of one monitoring index within the set observation time length.

[0050] All monitoring indicators for medical devices are divided into equipment indicators and patient-related indicators. Taking a ventilator as an example, equipment indicators may include: the operating status of the device (working, standby, off, etc.), the device settings parameters, device ID, device type, device model and device manufacturer, etc., while patient-related indicators may include, but are not limited to: (1) the patient and the patient's gender; (2) detailed information on respiratory support therapy, such as the type of mechanical ventilation (invasive or non-invasive), intubation status and extubation event; (3) ventilator settings and monitoring parameters, including ventilator mode, tidal volume, respiratory rate, inspiratory pressure, positive end-expiratory pressure and fractional oxygen inhaled; (4) monitoring parameters, including heart rate, blood pressure, blood oxygen saturation, respiratory rate and other vital signs data.

[0051] This example is merely illustrative and is not intended to be limiting.

[0052] S1012. Based on all initial feature matrices and their corresponding initial label matrices, select several key indicators from all monitoring indicators.

[0053] In this embodiment, the key indicator can be a monitoring indicator that is closely related to the occurrence of clinical alarms.

[0054] S1013. For each initial sample, delete all content in the initial sample except for the time-series feature vectors and alarm tag vectors of each key indicator to obtain the key sample.

[0055] For example, suppose an initial sample contains an initial feature matrix including time-series feature vectors of 93 monitoring indicators and an initial label matrix including alarm label vectors of 93 monitoring indicators. After processing in step S1013, the resulting key sample contains an initial feature matrix including time-series feature vectors of 15 key indicators and an initial label matrix including alarm label vectors of 15 key indicators. This example is merely illustrative and is not intended to be limiting.

[0056] Through the aforementioned steps S1011-S1013, this invention can remove data from the original dataset of various monitoring indicators that are not related to clinical alarms, retaining only the data of key indicators related to clinical alarms, thus avoiding interference from irrelevant data. For example, ventilator device ID, device type, device model, and device manufacturer are all irrelevant to clinical alarms, while monitoring indicators such as oxygen concentration and positive end-expiratory pressure collected by the ventilator are all related to clinical alarms.

[0057] In one optional implementation, an initial dataset can be formed based on several continuously collected historical datasets. Therefore, for step S1011, the process of constructing the initial dataset may include the following sub-steps S10111-S10116.

[0058] S10111. Obtain several historical data sets and their corresponding collection times.

[0059] In this embodiment, it is first necessary to obtain several sets of historical data and their corresponding collection times. The historical data includes the status data and triggering conditions of each monitoring indicator when the medical device is running.

[0060] This historical data can be collected periodically during the operation of medical devices or automatically when medical devices trigger alarms. During collection, the status data and triggering details of each monitoring indicator are recorded in detail. The status data reflects the specific value of the monitoring indicator at the time of collection, while the triggering details describe the type of alarm triggered by the monitoring indicator. Alarm types include three categories: triggering clinical intervention, triggering technical intervention, and not triggering intervention. This provides a clear classification basis for subsequent tag vector generation.

[0061] S10112. Based on the set observation duration and the collection time of each historical data collection, divide several historical data collections into multiple observation sets.

[0062] In this embodiment, several sets of historical data can be divided into multiple observation sets based on the relationship between the set observation duration and the collection time of each historical data set.

[0063] The division rule here is to ensure that the difference in collection time between any two historical data sets in each observation set does not exceed the set observation period. This process aims to ensure that the data contained in each observation set has a certain temporal continuity, thereby accurately reflecting the dynamic changes of the monitoring indicators within the set observation period.

[0064] For example, please combine Figure 2 Suppose that the collection times of the 6 historical data sets in an observation set are as follows: Figure 2 In the observation set, t1~t6, except for the historical data corresponding to time t3 which was automatically collected due to the triggering of an alarm, the other 5 historical data were collected at intervals T. In addition, in the observation set, the time difference ∆t between time t1 and time t6 is less than or equal to the set observation duration T0.

[0065] Taking a ventilator as an example, the interval period T can be 5 minutes, 10 minutes, 15 minutes, etc., while the set observation duration T0 can be 25 minutes, 30 minutes, 40 minutes, etc. This example is only for illustration and is not intended to be limiting.

[0066] S10113. For any observation set, normalize all state data of each monitoring indicator in the observation set to obtain the time series feature vector corresponding to each monitoring indicator.

[0067] In this embodiment, the time-series feature vector corresponding to a monitoring indicator can be obtained by normalizing all the state data of a monitoring indicator in an observation set.

[0068] S10114. Encode and convert all triggering conditions of each monitoring indicator in the observation set to obtain the alarm tag vector corresponding to each monitoring indicator.

[0069] In this embodiment, all triggering events of a monitoring indicator in an observation set can be encoded and converted into a label to represent the indicator, thus obtaining the alarm label vector corresponding to the monitoring indicator. During the encoding process, the label is 1 only when the triggering event belongs to clinical treatment; it is 0 when the triggering event belongs to technical treatment or when no alarm is triggered. In this way, the alarm label vector provides a clear supervision signal for subsequent model training, enabling the algorithm to learn to recognize clinical alarms.

[0070] For example, if the number of monitoring indicators is M, and an observation set includes S sets of historical data, then for the m-th monitoring indicator among the M monitoring indicators:

[0071] (1) The time series feature vector of the mth monitoring indicator is obtained by normalizing the S state data corresponding to the mth monitoring indicator to obtain the time series feature vector corresponding to the mth monitoring indicator. Therefore, a time series feature vector can characterize the dynamic changes of a monitoring indicator within a set observation period.

[0072] (2) The alarm label vector of the mth monitoring indicator includes S labels, which are obtained as follows: if the mth monitoring indicator triggers technical treatment or does not trigger an alarm at the time of the collection of the sth historical data, then the sth label is set to 0; if the mth monitoring indicator triggers clinical treatment at the time of the collection of the sth historical data, then the sth label is set to 1.

[0073] Combination Figure 2 If an observation set includes 6 historical data sets, then the time-series feature vector of each monitoring indicator includes 6 feature values, and each alarm tag vector includes 6 tags. Figure 2 If a monitoring indicator A triggers a clinical alarm at time t3, then the alarm label vector of monitoring indicator A is: 001000.

[0074] The above examples are merely illustrative; the embodiments of the present invention do not limit the number of historical data collected in the observation set, the number of clinical alarms, or the timing.

[0075] S10115. Based on the time-series feature vectors and alarm tag vectors corresponding to all monitoring indicators, the initial samples corresponding to the observation set are obtained.

[0076] In this embodiment, for an observation set, the time-series feature vector and alarm tag vector obtained by processing the observation set through the above steps S10111~S10115 can form an initial sample.

[0077] This initial sample fully contains the status change information of all monitoring indicators and their corresponding alarm triggering situations within the set observation period.

[0078] S10116. Iterate through each observation set to obtain the initial dataset.

[0079] By processing each observation set one by one through steps S10111~S10115 described above, a complete initial dataset can be obtained. This initial dataset not only covers rich dynamic features of monitoring indicators but also retains clear alarm label information, laying a solid foundation for the subsequent construction of key datasets and the training of target classification models. The entire process is logically rigorous and interconnected, ensuring efficient transformation from raw historical data to a structured initial dataset while preserving the most effective information from the original data.

[0080] In one alternative implementation, the correlation between each monitoring indicator and the triggering of clinical alarms can be analyzed first using mathematical statistical methods, and then the importance of the monitoring indicators can be analyzed using machine learning models to ensure the accuracy of feature screening.

[0081] Correspondingly, the process of selecting multiple key indicators from all monitoring indicators in step S1012 above may include the following sub-steps S10121 to S10124.

[0082] S10121. Based on all initial feature matrices and their corresponding initial label matrices, calculate the correlation coefficient between each monitoring indicator and the occurrence of clinical alarms by medical devices.

[0083] In this embodiment, the correlation strength between each monitoring indicator and the occurrence of clinical alerts is quantified by calculating correlation coefficients. This method of calculating correlation coefficients ensures independent analysis of each monitoring indicator, thereby enabling an objective measurement of its potential role in predicting clinical alerts.

[0084] For example, the correlation coefficient between a monitoring indicator A and a clinical alarm issued by a medical device can be the Spearman rank correlation coefficient. The calculation requires the use of the time-series feature vector and the alarm label vector of the monitoring indicator A. The specific calculation process is existing technology and will not be elaborated here.

[0085] S10122. Each monitoring indicator whose correlation coefficient exceeds the preset threshold is selected as a candidate indicator.

[0086] In this embodiment, monitoring indicators with correlation coefficients exceeding a preset threshold are selected as candidate indicators. This process involves indicator screening, where only candidate indicators that are strongly correlated with triggering clinical alarms are retained for subsequent analysis.

[0087] For example, the size of the preset threshold can be flexibly set, such as 0.7, 0.75, etc.

[0088] S10123. Remove all content from each initial sample except for the time series feature vectors of each candidate indicator to obtain the intermediate dataset.

[0089] In this embodiment, the initial dataset is further converted into an intermediate dataset, which focuses solely on each candidate indicator, thereby providing a clearer and more focused data foundation for subsequent importance analysis.

[0090] S10124. Based on the intermediate dataset, use at least one machine learning algorithm to perform importance analysis on all candidate indicators in order to select multiple key indicators from all candidate indicators.

[0091] In this embodiment, at least one machine learning algorithm is used to perform importance analysis on all candidate indicators, with the aim of further identifying those key indicators that truly have a decisive impact on clinical alarm judgment.

[0092] Taking the use of two machine learning algorithms as an example, the process of selecting key indicators may include the following steps a1 to a4.

[0093] Step a1: Based on the intermediate dataset, use the first classification algorithm and the second classification algorithm to determine the first importance coefficient and the second importance coefficient of each candidate indicator, respectively.

[0094] In this embodiment, two different classification algorithms are used to measure the importance of each candidate indicator from different perspectives. The choice between the first and second classification algorithms should be determined based on the actual application scenario and data characteristics.

[0095] For example, the first classification algorithm can be a logistic regression algorithm, and the second classification algorithm can be an adaptive boosting algorithm (i.e., the AdaBoost algorithm) or a gradient learning algorithm.

[0096] Step a2: Sort all candidate indicators according to the order of the first importance coefficient from largest to smallest and the order of the second importance coefficient from largest to smallest, to obtain the first sorting result and the second sorting result.

[0097] Taking ventilators as an example, the first importance coefficient of the top 15 monitoring indicators in the first ranking results is as follows: Figure 3As shown, Figure 3 The 15 monitoring indicators are: positive end-expiratory pressure (vent_peep_value), tidal volume (vent_tvi_value), minute ventilation (vent_mve_value), heart rate (monitor_pr), alarm severity (alarm_secerity), alarm duration (alarm_duration), spontaneous respiratory rate (vent_fspn_value), ventilator leakage percentage (vent_leak_percent_value), plateau pressure (vent_pplat_value), mean airway pressure (vent_pmean_value), oxygen concentration (vent_fio2_value), peak pressure (vent_ppeak_value), spontaneous minute ventilation (vent_mv_spn_value), and blood oxygen saturation (monitor_spo2).

[0098] Figure 3 The examples shown are merely illustrative, and the embodiments of the present invention do not limit the number of monitoring indicators or the specific meaning of each monitoring indicator.

[0099] Step a3: Each candidate indicator in the first and second sorting results that is located in the preset first position interval is taken as a key indicator.

[0100] Step a4: Determine the intersection between each candidate indicator in the first ranking result that is located in the preset second position interval and each candidate indicator in the second ranking result that is located in the second position interval, and take each candidate indicator in the intersection as a key indicator.

[0101] In this embodiment, the first position interval and the second position interval are adjacent.

[0102] For example, the first position interval can be [1,4], and the second position interval can be [5,10]. In this way, the first four candidate indicators in the first ranking result and the first four candidate indicators in the second ranking result are both used as key indicators, and each of the intersections of the first ranking result and the second ranking result from the fifth to the tenth candidate indicator is also used as a key indicator. This example is only for illustration and is not intended to be limiting.

[0103] This dual-ranking mechanism comprehensively considers the priority of candidate indicators from different algorithmic perspectives. Since a single machine learning algorithm may have limitations or biases in indicator selection, the collaborative approach of multiple machine learning algorithms can effectively improve the reliability and robustness of key indicator screening.

[0104] Taking the target algorithm (i.e., the first classification algorithm or the second classification algorithm) as an example, the method for obtaining the target importance coefficient (i.e., the first importance coefficient or the second importance coefficient) of each candidate indicator in step a1 above includes the following three steps:

[0105] (1) Based on the target algorithm, construct the target model;

[0106] (2) Use the intermediate dataset to train the target model to obtain the trained target model;

[0107] (3) The model weight corresponding to each candidate indicator in the trained target model is used as the target importance coefficient.

[0108] By training the target model using intermediate datasets, the model gradually adjusts its internal parameters to minimize prediction error. During this process, the weights of different candidate metrics are dynamically updated until an optimal state is reached.

[0109] Taking a logistic regression model as an example, in the initial state, the weights of all candidate indicators are evenly distributed. Then, through multiple iterations, the loss function is calculated and the weights are modified so that the model's prediction results gradually approach the true labels. After training, the weights of some candidate indicators may tend to zero, indicating that these candidate indicators have no significant correlation with the prediction of the event's authenticity and can therefore be excluded; while candidate indicators with larger weights are considered to have a more important role in the model's prediction.

[0110] This invention fully leverages the advantages of machine learning models to ensure the accuracy and reliability of importance assessment for candidate metrics.

[0111] In one optional implementation, multiple preset algorithm models can be trained and their performance tested using a key dataset, and the model with the best performance can be selected as the target classification model. Correspondingly, in step S102 above, the process of training multiple preset algorithm models using a key dataset to select the target classification model from the multiple candidate classification models may include the following sub-steps S1021 to S1024.

[0112] S1021. Divide the key dataset into training set and test set.

[0113] Optionally, the 10-fold cross-validation method commonly used in model evaluation techniques can be adopted, which involves dividing the dataset into several subsets to more accurately evaluate the model's performance and enhance the reliability of the model evaluation.

[0114] In this embodiment, the entire key dataset is randomly divided into 10 folds, and the data is divided into training set and test set in a 7:3 ratio.

[0115] S1022. Train multiple algorithm models using the training set to obtain multiple candidate classification models.

[0116] S1023. Based on the test set, evaluate the performance of each candidate classification model to obtain at least one performance index corresponding to each candidate classification model.

[0117] In this embodiment, performance metrics may include one or more of the following: Area Under the Receiver Operating Characteristic Curve (AUC-ROC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 score, and Kappa value.

[0118] For example, various algorithmic models may include, but are not limited to: Gaussian Naive Bayes (NB), K-Nearest Neighbors (KNN), Linear Discriminant Analysis (LDA), Classification and Regression Trees (CART), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), Logistic Regression (LR), and Random Forest (RF). Please refer to... Figure 4 , Figure 4 This refers to the inventor's performance evaluation results for these models.

[0119] S1024. Based on at least one performance metric corresponding to each candidate classification model, train the target classification model from each candidate classification model.

[0120] In one alternative implementation, each candidate classification model can be sorted in descending order of each performance metric to obtain the model ranking result corresponding to each performance metric; then, the candidate classification model that ranks first most frequently in the model ranking results corresponding to all performance metrics is taken as the target prediction model.

[0121] In another alternative implementation, the performance metrics corresponding to each candidate classification model can be weighted and averaged to obtain the performance evaluation score of each candidate classification model, and the candidate classification model with the highest performance evaluation score can be used as the target classification model.

[0122] Combination Figure 4 If a weighted scoring method is used for Figure 4 The performance metrics of the eight models were processed, and the XGBoost model was ultimately selected as the target classification model.

[0123] The above content introduced the process of obtaining the target classification model. Next, we will introduce the application process of the target classification model.

[0124] Please see Figure 5 , Figure 5 This is a flowchart illustrating an alarm classification method based on a machine learning model, provided by an embodiment of the present invention. This method can be applied to electronic devices such as smartphones, personal laptops, personal computers, and servers. Figure 5 The method includes the following steps S201 to S204.

[0125] S201. Receive the medical equipment data collected each time by the data acquisition device.

[0126] In this embodiment, the data acquisition device can be a standalone data acquisition unit specifically designed for collecting data from medical devices. The acquired data can include status data and status flag values ​​for multiple monitoring indicators; wherein, a status flag value of 0 indicates that the data is collected periodically by the data acquisition device, and a status flag value of 1 indicates that the data is collected when the medical device senses an alarm.

[0127] S202. If the status flag value of any monitoring indicator in the received data is 1, then the time after waiting for the preset time period shall be used as the judgment time.

[0128] In this embodiment, when the status flag value of any monitoring indicator in the received collected data is 1, it indicates that the medical device has triggered an alarm. At this time, the system does not immediately perform a classification judgment, but instead sets a preset waiting period (e.g., 5 minutes). This waiting strategy aims to ensure sufficient data for subsequent analysis to determine whether the alarm is a clinical alarm, thereby improving classification accuracy. The time determined after the waiting period ends is the judgment time.

[0129] S203. Generate a feature matrix to be identified based on all collected data within the observation period set before the judgment time.

[0130] In this embodiment, a feature matrix to be identified is generated based on all collected data within the observation period set before the aforementioned judgment time.

[0131] Optionally, it is necessary to determine the M sets of collected data within the observation period set before the time point, obtain M state data for each key indicator, and normalize the M state data for each key indicator to obtain the feature matrix to be identified.

[0132] S204. Input the feature matrix to be identified into the target classification model to obtain the classification result.

[0133] In this embodiment, the generated feature matrix to be identified is input into the target classification model to obtain the classification result. The classification result reflects whether the alarm triggered by the medical device is a clinical alarm or a non-clinical alarm.

[0134] The alarm classification method provided by this invention can convert the status data of various key indicators within a certain period before and after the alarm is triggered into a feature matrix to be identified in a timely manner when the medical device triggers an alarm. Then, the feature matrix to be identified is input into the target classification model to determine whether the alarm belongs to clinical alarm or non-clinical alarm, thus ensuring the timeliness and accuracy of alarm identification.

[0135] Based on and Figure 1 Using the same inventive concept, embodiments of the present invention also provide a model training device 200. See also... Figure 6 , Figure 6 The functional block diagram of the model training device provided in the embodiment of the present invention includes: a dataset construction module 210 and a model selection module 220.

[0136] Dataset construction module 210 is used to construct a key dataset, in which each key sample reflects whether the medical device has a clinical alarm within a set observation period.

[0137] The model screening module 220 is used to train multiple preset algorithm models using key datasets to select a target classification model from the multiple candidate classification models obtained. The target classification model is used to determine whether the alarm generated by the medical device belongs to the clinical alarm.

[0138] Optionally, the key dataset includes several key samples. During the construction of the key dataset, the dataset construction module 210 can specifically be used for:

[0139] An initial dataset is constructed, consisting of several initial samples. Each initial sample includes an initial feature matrix and its corresponding initial label matrix. The initial feature matrix includes time-series feature vectors of multiple monitoring indicators within a set observation period, and the initial label matrix includes alarm label vectors corresponding to each monitoring indicator. Based on all initial feature matrices and their corresponding initial label matrices, several key indicators are selected from all monitoring indicators. For each initial sample, all content except the time-series feature vectors and alarm label vectors of each key indicator is removed to obtain the key samples.

[0140] Optionally, during the construction of the initial dataset, the dataset construction module 210 can be specifically used to: acquire several sets of historical data and their corresponding acquisition times, the historical data including the status data and triggering conditions of each monitoring indicator during the operation of the medical equipment; divide the several sets of historical data into multiple observation sets based on the set observation duration and the acquisition times of each historical data set; wherein the difference in acquisition times between any two sets of historical data in each observation set does not exceed the set observation duration; for any observation set, normalize all status data of each monitoring indicator in the observation set to obtain the time-series feature vector corresponding to each monitoring indicator; encode and convert all triggering conditions of each monitoring indicator in the observation set to obtain the alarm label vector corresponding to each monitoring indicator; obtain the initial samples corresponding to the observation set based on the time-series feature vectors and alarm label vectors corresponding to all monitoring indicators; and traverse each observation set to obtain the initial dataset.

[0141] Optionally, in the process of selecting multiple key indicators from all monitoring indicators based on all initial feature matrices and their corresponding initial label matrices, the dataset construction module 210 can be specifically used to: calculate the correlation coefficient between each monitoring indicator and the occurrence of clinical alarms by medical devices based on all initial feature matrices and their corresponding initial label matrices; select each monitoring indicator whose correlation coefficient exceeds a preset threshold as a candidate indicator; delete all content in each initial sample except for the time-series feature vectors of each candidate indicator to obtain an intermediate dataset; and, based on the intermediate dataset, use at least one machine learning algorithm to perform importance analysis on all candidate indicators in order to select multiple key indicators from all candidate indicators.

[0142] Optionally, in the process of using at least one machine learning algorithm to perform importance analysis on all candidate indicators based on the intermediate dataset to select multiple key indicators from all candidate indicators, the dataset construction module 210 can be specifically used to: determine the first importance coefficient and the second importance coefficient of each candidate indicator based on the intermediate dataset using a first classification algorithm and a second classification algorithm respectively; sort all candidate indicators according to the order of the first importance coefficient from largest to smallest and the order of the second importance coefficient from largest to smallest, respectively, to obtain a first sorting result and a second sorting result; take each candidate indicator in the first sorting result and the second sorting result that is located in a preset first position interval as a key indicator; determine the intersection between each candidate indicator in the first sorting result that is located in a preset second position interval and each candidate indicator in the second sorting result that is located in the second position interval, and take each candidate indicator in the intersection as a key indicator.

[0143] Optionally, the target importance coefficient of each candidate indicator is obtained in the following way: based on the target algorithm, a target model is constructed; the target algorithm is a first classification algorithm or a second classification algorithm; the target model is trained using an intermediate dataset to obtain a trained target model; the model weight corresponding to each candidate indicator in the trained target model is used as the target importance coefficient; the target importance coefficient is a first importance coefficient or a second importance coefficient.

[0144] Optionally, in the process of training multiple preset algorithm models using a key dataset to select the target classification model from the multiple candidate classification models, the model selection module 220 can be specifically used to: divide the key dataset into a training set and a test set; train multiple algorithm models using the training set to obtain multiple candidate classification models; evaluate the performance of each candidate classification model based on the test set to obtain at least one performance index corresponding to each candidate classification model; and train the target classification model from each candidate classification model based on the at least one performance index corresponding to each candidate classification model.

[0145] Optionally, in the process of training the target classification model from each candidate classification model based on at least one performance metric corresponding to each candidate classification model, the model selection module 220 can be specifically used to: sort each candidate classification model in descending order of each performance metric to obtain the model ranking result corresponding to each performance metric; and select the candidate classification model that ranks first most frequently among all the model ranking results corresponding to all performance metrics as the target prediction model; or, the model selection module 220 can be specifically used to: perform a weighted average of each performance metric corresponding to each candidate classification model to obtain the performance evaluation score of each candidate classification model, and select the candidate classification model with the highest performance evaluation score as the target classification model.

[0146] Based on and Figure 5 Following the same inventive concept, embodiments of the present invention also provide an alarm classification device 400 based on a machine learning model. See also... Figure 7 , Figure 7 The functional block diagram of the alarm classification device based on the machine learning model provided in the embodiment of the present invention includes: a data receiving module 410, a vectorization module 420, and an alarm classification module 430.

[0147] The data receiving module 410 is used to receive the data collected by the acquisition device from the medical device each time. The acquired data includes the status data of multiple monitoring indicators and status flag values. Among them, the status flag value of 0 indicates that the acquired data is collected by the acquisition device at regular intervals, and the status flag value of 1 indicates that the acquired data is collected when the medical device detects an alarm.

[0148] The vectorization module 420 is used to determine the judgment time when the status flag value of any monitoring indicator in the received collected data is 1, after waiting for a preset time.

[0149] The vectorization module 420 is also used to generate a feature matrix to be identified based on all the collected data within the observation period set before the judgment time.

[0150] The alarm classification module 430 is used to input the feature matrix to be identified into the target classification model to obtain the classification result; wherein, the classification result reflects whether the alarm triggered by the medical device is a clinical alarm or a non-clinical alarm, and the target classification model is obtained by model training using the above-mentioned model training method.

[0151] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the model training device 200 and the alarm classification device 400 based on the machine learning model described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0152] Please see Figure 8 , Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 300 includes a processor 310, a memory 320, and a bus 330, with the processor 310 connected to the memory 320 via the bus 330.

[0153] The memory 320 can be used to store software programs, such as the software programs corresponding to the model training apparatus 200 and / or the alarm classification apparatus 400 based on the machine learning model provided in the embodiments of the present invention. The processor 310 executes various functional applications and data processing by running the software programs stored in the memory 320 to implement the model training method and / or the alarm classification method based on the machine learning model provided in the embodiments of the present invention.

[0154] The memory 320 may 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.

[0155] 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 Processor), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0156] Understandable. Figure 8 The structure shown is for illustrative purposes only; the electronic device 300 may also include components that are more advanced than those shown. Figure 8 The more or fewer components shown, or having the same Figure 8 The different configurations shown. Figure 8 The components shown can be implemented using hardware, software, or a combination thereof.

[0157] This invention also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the model training method disclosed in the above embodiments, and / or, an alarm classification method based on a machine learning model. The computer-readable storage medium can be, but is not limited to, various media capable of storing program code, such as a USB flash drive, external hard drive, ROM, RAM, PROM, EPROM, EEPROM, FLASH disk, or optical disk.

[0158] In summary, the embodiments of the present invention provide a model training method, an alarm classification method based on a machine learning model, an apparatus, an electronic device, and a storage medium. After constructing a key dataset, multiple preset algorithm models can be trained using the key dataset to select a target classification model from the multiple candidate classification models obtained. The target classification model can then be used to identify whether the alarms generated by medical devices are clinical alarms, thereby reducing the pressure on medical staff facing a large number of non-clinical alarms. Furthermore, the target classification model can be used to quickly identify clinical alarms, facilitating medical staff to take rapid treatment measures.

[0159] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A model training method, characterized in that, include: Acquire several sets of historical data and their corresponding acquisition times. The historical data includes the status data and triggering status of each monitoring indicator during the operation of the medical device. The triggering status reflects the alarm type triggered by the monitoring indicator, which is either triggering clinical treatment, triggering technical treatment, or not triggering treatment. Based on the set observation duration and the collection time of each of the historical data sets, several sets of historical data sets are divided into multiple observation sets; wherein, the difference in collection time between any two sets of historical data sets in each observation set does not exceed the set observation duration. For any of the observation sets, normalize all state data of each of the monitoring indicators in the observation set to obtain the time-series feature vectors corresponding to each of the monitoring indicators; Each of the monitoring indicators is encoded and converted for all triggering conditions in the observation set to obtain the alarm tag vector corresponding to each monitoring indicator. Based on the time-series feature vectors and alarm label vectors corresponding to all the monitoring indicators, an initial sample corresponding to the observation set is obtained; the initial sample includes an initial feature matrix and its corresponding initial label matrix; the initial feature matrix includes the time-series feature vectors of multiple monitoring indicators within the set observation period, and the initial label matrix includes the alarm label vector corresponding to each monitoring indicator; Iterate through each of the observation sets to obtain the initial sample corresponding to each observation set; Based on all the initial feature matrices and their corresponding initial label matrices, multiple key indicators are selected from all the monitoring indicators. Each initial sample is obtained by deleting all content except for the time-series feature vectors and alarm label vectors of each key indicator; the key sample reflects whether the medical device has issued a clinical alarm within the set observation period. Multiple preset algorithm models are trained using a key dataset consisting of all the key samples to select a target classification model from the multiple candidate classification models. The target classification model is used to determine whether the alarm generated by the medical device belongs to a clinical alarm.

2. The method according to claim 1, characterized in that, The step of selecting multiple key indicators from all the monitoring indicators based on all the initial feature matrices and their corresponding initial label matrices includes: Based on all the initial feature matrices and their corresponding initial label matrices, calculate the correlation coefficient between each monitoring indicator and the occurrence of a clinical alarm by the medical device; Each monitoring indicator whose correlation coefficient exceeds a preset threshold is selected as a candidate indicator. Remove all content from each initial sample except for the temporal feature vectors of each candidate indicator to obtain an intermediate dataset; Based on the intermediate dataset, at least one machine learning algorithm is used to perform importance analysis on all candidate indicators in order to select multiple key indicators from all the candidate indicators.

3. The method according to claim 2, characterized in that, The step of performing importance analysis on all candidate indicators using at least one machine learning algorithm based on the intermediate dataset, in order to select multiple key indicators from all candidate indicators, includes: Based on the intermediate dataset, the first importance coefficient and the second importance coefficient of each candidate indicator are determined using the first classification algorithm and the second classification algorithm, respectively. All candidate indicators are sorted according to the first importance coefficient from largest to smallest and the second importance coefficient from largest to smallest, respectively, to obtain a first sorting result and a second sorting result; Each candidate indicator in the first sorting result and the second sorting result that is located in the preset first position interval is used as the key indicator; Determine the intersection between each candidate indicator in the first ranking result that is located in the preset second position interval and each candidate indicator in the second ranking result that is located in the second position interval, and use each candidate indicator in the intersection as the key indicator.

4. The method according to claim 3, characterized in that, The target importance coefficient for each candidate indicator is obtained in the following way: A target model is constructed based on the target algorithm; the target algorithm is either the first classification algorithm or the second classification algorithm. The target model is trained using the intermediate dataset to obtain the trained target model; The model weight corresponding to each candidate indicator in the trained target model is used as the target importance coefficient; the target importance coefficient is either the first importance coefficient or the second importance coefficient.

5. The method according to claim 1, characterized in that, The step of training multiple preset algorithm models using the key dataset to select the target classification model from the multiple candidate classification models includes: The key dataset is divided into a training set and a test set; The training set is used to train various algorithm models to obtain multiple candidate classification models; Based on the test set, the performance of each candidate classification model is evaluated to obtain at least one performance index corresponding to each candidate classification model. The target classification model is trained from each of the candidate classification models based on at least one performance metric corresponding to each candidate classification model.

6. The method according to claim 5, characterized in that, The step of training the target classification model from each of the candidate classification models based on at least one performance metric corresponding to each of the candidate classification models includes: The candidate classification models are sorted in descending order of each performance metric to obtain the model ranking result corresponding to each performance metric. The candidate classification model that ranks first most frequently among the model ranking results corresponding to all the performance indicators is taken as the target classification model. or, Each candidate classification model is weighted and averaged to obtain a performance evaluation score for each candidate classification model, and the candidate classification model with the highest performance evaluation score is selected as the target classification model.

7. An alarm classification method based on a machine learning model, characterized in that, include: The system receives data collected from medical devices each time, including status data and status flag values ​​for multiple monitoring indicators. A status flag value of 0 indicates that the data is collected periodically by the system, and a status flag value of 1 indicates that the data is collected when the medical device triggers an alarm. If the status flag value of any of the monitoring indicators in the received data is 1, then the time after waiting for a preset time is taken as the judgment time. Based on all collected data within the set observation period before the judgment time, a feature matrix to be identified is generated. The feature matrix to be identified is input into the target classification model to obtain the classification result; wherein, the classification result reflects whether the alarm triggered by the medical device is a clinical alarm or a non-clinical alarm, and the target classification model is obtained by model training using the method described in any one of claims 1-6.

8. A model training device, characterized in that, include: The dataset building module is used for: Acquire several sets of historical data and their corresponding acquisition times. The historical data includes the status data and triggering status of each monitoring indicator during the operation of the medical device. The triggering status reflects the alarm type triggered by the monitoring indicator, which is either triggering clinical treatment, triggering technical treatment, or not triggering treatment. Based on the set observation duration and the collection time of each of the historical data sets, several sets of historical data sets are divided into multiple observation sets; wherein, the difference in collection time between any two sets of historical data sets in each observation set does not exceed the set observation duration. For any of the observation sets, normalize all state data of each of the monitoring indicators in the observation set to obtain the time-series feature vectors corresponding to each of the monitoring indicators; Each of the monitoring indicators is encoded and converted for all triggering conditions in the observation set to obtain the alarm tag vector corresponding to each monitoring indicator. Based on the time-series feature vectors and alarm label vectors corresponding to all the monitoring indicators, an initial sample corresponding to the observation set is obtained; the initial sample includes an initial feature matrix and its corresponding initial label matrix; the initial feature matrix includes the time-series feature vectors of multiple monitoring indicators within the set observation period, and the initial label matrix includes the alarm label vector corresponding to each monitoring indicator; Iterate through each of the observation sets to obtain the initial sample corresponding to each observation set; Based on all the initial feature matrices and their corresponding initial label matrices, multiple key indicators are selected from all the monitoring indicators. Each initial sample is obtained by deleting all content except for the time-series feature vectors and alarm label vectors of the key indicators; the key sample reflects whether the medical device generates a clinical alarm within the set observation period. The model screening module is used to train multiple preset algorithm models using a key dataset consisting of all the key samples, so as to select a target classification model from the multiple candidate classification models obtained. The target classification model is used to determine whether the alarm generated by the medical device belongs to a clinical alarm.

9. An alarm classification device based on a machine learning model, characterized in that, include: The data receiving module is used to receive the data collected by the acquisition device from the medical device each time. The data includes status data and status flag values ​​of multiple monitoring indicators. The status flag value of 0 indicates that the data is collected by the acquisition device at regular intervals, and the status flag value of 1 indicates that the data is collected when the medical device detects an alarm. The vectorization module is used to determine the judgment time when the status flag value of any of the monitoring indicators in the received collected data is 1, after waiting for a preset time. The vectorization module is also used to generate a feature matrix to be identified based on all collected data within a set observation period before the judgment time. An alarm classification module is used to input the feature matrix to be identified into a target classification model to obtain a classification result; wherein the classification result reflects whether the alarm triggered by the medical device is a clinical alarm or a non-clinical alarm, and the target classification model is obtained by model training using the method described in any one of claims 1-6.

10. An electronic device, characterized in that, include: A memory and a processor, the memory storing a software program, which the processor executes when the electronic device is running to implement the method as described in any one of claims 1-6, or the method as described in claim 7.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1-6, or the method as described in claim 7.

Images