Detection of hemodynamic instability

Abstract

Embodiments of the present disclosure relate to a method of detecting hemodynamic instability. The method includes obtaining a plurality of clinical parameters of the subject at a point in time. A plurality of prediction probabilities may then be determined based on the plurality of clinical parameters, where each prediction probability of the plurality of prediction probabilities is determined by a corresponding sub-model of a plurality of sub-models in the machine learning model. The method further includes determining a combined prediction probability for the point in time based on the plurality of prediction probabilities. The determined combined prediction probability is used to detect hemodynamic instability. According to the embodiment of the invention, the hemodynamic instability can be timely and accurately detected to give an alarm to a clinician, so that the death rate and the medical cost are greatly reduced.

Description

Technical Field

[0001] The embodiments of this disclosure generally relate to the field of data processing, and more specifically to the detection of hemodynamic instability. Background Technology

[0002] Hemodynamic instability describes unstable blood flow. It is a condition or state in which a person's cardiovascular function becomes unreliable, inadequate, or otherwise problematic. Therefore, hemodynamic instability is a critical and common condition in the intensive care unit (ICU). Because hemodynamic instability in patients who have suffered trauma cannot be detected in a timely manner, one-third of ICU patients will develop shock, a condition with a high mortality rate.

[0003] Early warning systems for hemodynamic instability have the potential to improve the timely detection and initiation of interventions. If hemodynamic instability can be predicted in a timely manner in patients following trauma, preventative treatment plans can be provided for high-risk patients, thereby significantly reducing mortality and healthcare costs. However, some problems remain with early warning systems for hemodynamic instability, and these problems need to be addressed to improve the prediction of hemodynamic instability. Potes et al.'s paper, "A clinical prediction model to identify patients at high risk of hemodynamic instability in the pediatric intensive care unit" (Critical Care, Vol. 21, No. 1, December 1, 2017), proposes a model for predicting the need for hemodynamic interventions in the pediatric intensive care unit. Summary of the Invention

[0004] In hemodynamic instability warning systems, single-parameter shock indicators (e.g., systolic blood pressure (SBP) and shock index (heart rate / SBP)) are reported to detect hemodynamic instability. However, the patient's condition may worsen in the later stages of shock, and / or the risk may be underestimated simply by addressing cardiovascular changes. Low-accuracy warning systems also hinder clinicians from identifying and interpreting relevant information, making it difficult for them to continuously monitor or assess ICU patients. Low-frequency and low-accuracy warning systems are unsuitable for patients whose condition changes rapidly. In the absence of predictive tools with sufficient accuracy, clinicians may make subjective judgments, increasing the risk of erroneous actions.

[0005] In view of this, embodiments of the present disclosure provide methods for detecting hemodynamic instability, methods for training machine learning models for detecting hemodynamic instability, electronic devices, and computer-readable media. The present disclosure enables timely and accurate detection of hemodynamic instability to alert clinicians, thereby significantly reducing mortality and healthcare costs.

[0006] In a first aspect, embodiments of this disclosure provide a method for detecting hemodynamic instability. The method includes: obtaining multiple clinical parameters of an object at a given time point, the multiple clinical parameters including at least two of the following: vital signs data, laboratory measurements, waveform feature data, and image data. The method further includes: determining multiple predicted probabilities by a machine learning model based on the multiple clinical parameters, each of the multiple predicted probabilities being determined by a corresponding sub-model among multiple sub-models in the machine learning model. The method further includes: determining a combined predicted probability for that time point based on the multiple predicted probabilities. The method further includes: detecting hemodynamic instability based on the combined predicted probability.

[0007] According to embodiments of this disclosure, the method uses at least two types of clinical parameters. Therefore, more parameters indicative of the underlying pathophysiology of the cardiovascular system are used to improve prediction accuracy. Furthermore, multiple sub-models can use multiple clinical parameters to determine multiple prediction probabilities, and a combined prediction probability is determined based on these multiple prediction probabilities. The combined prediction probability will be more accurate, thus enabling the method to detect hemodynamic instability promptly and accurately to alert clinicians, thereby significantly reducing mortality and healthcare costs.

[0008] In some embodiments of the first aspect, vital sign data include at least one of the following: heart rate, temperature, respiratory rate, and blood pressure; laboratory measurement data includes at least one of the following: creatinine, blood glucose, glutamate, lactate, alanine aminotransferase, and aspartate aminotransferase; waveform feature data includes at least one of the following: peak-trough slope of the heart rate waveform, peak-to-peak interval of the heart rate waveform, peak-trough slope of the ventilator waveform, and peak-to-peak interval of the ventilator waveform; and / or image data includes at least one of the following: X-ray image, ultrasound image, and computed tomography image. In this way, multiple types of parameters are used to detect hemodynamic instability to improve prediction accuracy.

[0009] In some embodiments of the first aspect, obtaining multiple clinical parameters of an object at a given time point includes: determining whether one of the multiple clinical parameters is empty. Obtaining multiple clinical parameters of an object at a given time point further includes: determining whether up-to-date data for the clinical parameter exists within a first screening time limit in response to determining that one of the multiple clinical parameters is empty. Obtaining multiple clinical parameters of an object at a given time point further includes: updating the clinical parameter based on the up-to-date data in response to determining that the up-to-date data for the clinical parameter exists within the first screening time limit. In this way, even if the clinical parameter is empty at the current time point, valid data for the clinical parameter among the multiple clinical parameters can be accurately obtained. For example, if the clinical parameter is empty at the current time point and the up-to-date data for the clinical parameter exists within a screening time limit (e.g., 2 hours), the up-to-date data can be used. Therefore, this operation can improve the accuracy of detection.

[0010] In some embodiments of the first aspect, obtaining multiple clinical parameters of an object at a given time point further includes: determining whether another clinical parameter among the multiple clinical parameters is empty. Obtaining multiple clinical parameters of an object at a given time point also includes: in response to determining that another clinical parameter among the multiple clinical parameters is empty, determining whether additional up-to-date data for the other clinical parameter exists within a second screening time limit, the second screening time limit being different from the first screening time limit. Obtaining multiple clinical parameters of an object at a given time point also includes: in response to determining that additional up-to-date data for the other clinical parameter exists within the second screening time limit, updating the other clinical parameter based on the additional up-to-date data. Different screening time limits can be used for different types of clinical parameters. Therefore, using more efficient clinical parameter data improves the accuracy of detection.

[0011] In some embodiments of the first aspect, the machine learning model is one of a plurality of machine learning models that form a detection model for detecting hemodynamic instability, and each of the plurality of machine learning models is used to determine a combined predicted probability at a corresponding time point among a plurality of time points. In this way, for each of the plurality of time points, the corresponding machine learning model is used to detect hemodynamic instability. Therefore, the accuracy of hemodynamic instability detection is improved.

[0012] In some embodiments of the first aspect, determining a combined predicted probability for a given time point based on multiple predicted probabilities includes: for each corresponding time point among the multiple time points, determining a combined predicted probability based on multiple clinical parameters of the object obtained at each corresponding time point using a machine learning model corresponding to each corresponding time point, to obtain multiple combined predicted probabilities. Furthermore, detecting hemodynamic instability based on the combined predicted probabilities includes: selecting a predetermined number of combined predicted probabilities from the multiple combined predicted probabilities, each selected combined predicted probability being greater than the remaining combined predicted probabilities among the multiple combined predicted probabilities. Detecting hemodynamic instability based on the combined predicted probabilities also includes: determining whether each of the predetermined number of combined predicted probabilities is greater than a predetermined probability threshold. Detecting hemodynamic instability based on the combined predicted probabilities also includes: determining that hemodynamic instability has occurred in response to determining that each of the predetermined number of combined predicted probabilities is greater than the predetermined probability threshold. In this way, by utilizing a predetermined number of combined predicted probabilities, the detection results of hemodynamic instability can be more accurate.

[0013] In some embodiments of the first aspect, determining a predetermined number of combined prediction probabilities based on multiple combined prediction probabilities includes: determining a predetermined ratio related to the model performance of the detection model based on at least one of the accuracy, precision, recall, specificity, F1 score, and false positive rate of the detection model. Determining a predetermined number of combined prediction probabilities based on multiple combined prediction probabilities further includes: determining the predetermined number based on the predetermined ratio and the number of multiple machine learning models. In this way, a predetermined number of combined prediction probabilities can be obtained to provide optimal model performance.

[0014] In some embodiments of the first aspect, detecting hemodynamic instability based on combined prediction probabilities includes: obtaining multiple clinical parameters of the object for each of at least two additional time points different from the current time point. Detecting hemodynamic instability based on combined prediction probabilities further includes: for each of the at least two additional time points, determining a combined prediction probability based on the multiple clinical parameters obtained at each of the at least two additional time points using a machine learning model corresponding to each of the at least two additional time points. Detecting hemodynamic instability based on combined prediction probabilities further includes: detecting hemodynamic instability based on the combined prediction probability for the current time point and the combined prediction probabilities for each of the at least two additional time points. In this way, hemodynamic instability can be accurately calculated based on clinical parameters at different time points.

[0015] In some embodiments of the first aspect, determining the combined prediction probability for a given time point based on multiple prediction probabilities includes determining the combined prediction probability as the average probability of the multiple prediction probabilities. In this way, the combined prediction probability can be obtained more accurately.

[0016] In a second aspect, embodiments of this disclosure provide a method for training a machine learning model for detecting hemodynamic instability. The method includes: obtaining multiple sample clinical parameters of multiple sample objects at a single time point, the multiple sample clinical parameters including at least two of the following: vital signs data, laboratory measurement data, waveform feature data, and image data. The method further includes: determining a ratio of the number of positive samples to the number of negative samples in the multiple sample objects. The method further includes: generating multiple subsets, each subset including sample clinical parameters corresponding to positive samples and sample clinical parameters corresponding to a portion of the negative samples, wherein the negative samples are divided into multiple portions based on the ratio, and the portion of the negative samples in each subset is different from each other. The method further includes: training multiple sub-models in the machine learning model for detecting hemodynamic instability based on the multiple subsets.

[0017] According to embodiments of this disclosure, the method uses at least two types of sample clinical parameters. Therefore, more sample clinical parameters indicative of the underlying pathophysiology of the cardiovascular system are used to improve accuracy. Furthermore, positive and negative samples are balanced to form multiple subsets. Multiple subsets can be used to train multiple sub-models in a machine learning model. Because the balanced subsets are used during training, the accuracy of each sub-model is improved. Thus, the trained machine learning model is able to accurately detect hemodynamic instability.

[0018] In some embodiments of the second aspect, the machine learning model is one of a plurality of machine learning models that form a detection model for detecting hemodynamic instability, and the method further includes: obtaining additional clinical parameters of multiple sample objects at other time points different from this time point. The method further includes: training another machine learning model among the plurality of machine learning models based on the additional clinical parameters of the multiple sample objects. In this way, different machine learning models are trained using clinical parameters of multiple sample objects at different time points, thereby improving the performance of the detection model.

[0019] In some embodiments of the second aspect, obtaining multiple sample clinical parameters of multiple sample objects at a given time point includes: determining whether a sample clinical parameter among the multiple sample clinical parameters is empty. Obtaining multiple sample clinical parameters of multiple sample objects at a given time point further includes: in response to determining that a sample clinical parameter among the multiple sample clinical parameters is empty, determining whether the latest data for the sample clinical parameter exists within a first screening time limit. Obtaining multiple sample clinical parameters of multiple sample objects at a given time point further includes: in response to determining that the latest data for the sample clinical parameter exists within the first screening time limit, updating the sample clinical parameter based on the latest data. In this way, even if the clinical parameter is empty at the current time point, valid data for the clinical parameter among the multiple clinical parameters can be accurately obtained, thereby improving prediction results.

[0020] In some embodiments of the second aspect, training multiple sub-models in a machine learning model for detecting hemodynamic instability based on multiple subset datasets includes obtaining the predicted probabilities of the sub-models by inputting subset datasets from the multiple subset datasets into the sub-models. Training multiple sub-models in a machine learning model for detecting hemodynamic instability based on multiple subset datasets also includes adjusting the model parameters of the sub-models based on the predicted probabilities and the actual results of hemodynamic instability. In this way, the trained sub-models are more accurate and precise.

[0021] In a third aspect, embodiments of this disclosure provide an electronic device comprising: at least one processor; and at least one memory storing a plurality of instructions which, when executed by the at least one processor, cause the device to perform the methods of the first or second aspect.

[0022] In a fourth aspect, embodiments of the present disclosure provide a computer-readable medium having computer instructions stored thereon, which, when executed by a processor, cause the processor to perform the method of the first or second aspect.

[0023] It should be understood that the summary is neither intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0024] Details of one or more embodiments of this disclosure are set forth in the accompanying drawings and the following description. Other features, aspects, and advantages of this disclosure will become apparent from the specification, drawings, and claims, wherein: Figure 1 This is a block diagram of an example environment in which example embodiments of the present disclosure may be implemented; Figure 2 This is a schematic diagram illustrating a model structure for detecting hemodynamic instability according to an example embodiment of the present disclosure; Figure 3 This is a flowchart illustrating an example process for detecting hemodynamic instability according to an example embodiment of the present disclosure; Figure 4 This is a table illustrating example clinical parameters according to an example embodiment of the present disclosure; Figure 5 This is a flowchart illustrating an example method for detecting hemodynamic instability according to an example embodiment of the present disclosure; Figure 6 This is a block diagram of an example environment for training a machine learning model for detecting hemodynamic instability, according to an example embodiment of the present disclosure; Figure 7 This is a block diagram illustrating a model structure for detecting hemodynamic instability according to an example embodiment of the present disclosure; Figure 8 This is a schematic diagram illustrating an example process for training a machine learning model according to an example embodiment of the present disclosure; Figure 9 This is a flowchart illustrating an example method for training a machine learning model for detecting hemodynamic instability according to an example embodiment of the present disclosure; Figure 10A , Figure 10B and Figure 10C This is a table illustrating the model performance of a model for detecting hemodynamic instability according to an example embodiment of this disclosure; and Figure 11 This is a schematic diagram illustrating an example device suitable for implementing embodiments of the present disclosure.

[0025] Throughout the accompanying drawings, the same or similar reference numerals will always indicate the same or similar elements. Detailed Implementation

[0026] The principles of this disclosure will now be described with reference to some exemplary embodiments. It should be understood that these embodiments are described merely for illustrative purposes and to assist those skilled in the art in understanding and implementing this disclosure, and do not imply any limitation on the scope of this disclosure. The disclosure described herein can be implemented in various ways other than those described below.

[0027] As used herein, the term "comprising / including" and its variations should be understood as an open-ended term meaning "including / including but not limited to". The term "based on" will be understood as "at least partially based on". The terms "one embodiment" and "an embodiment" will be understood as "at least one embodiment". The term "another embodiment" will be understood as "at least one other embodiment". Furthermore, it should be understood that in the context of this disclosure, the terms "first", "second", etc., are used to indicate individual elements or components and do not imply any limitation on the order of these elements. Additionally, a first element may be the same as or different from a second element.

[0028] As mentioned above, when using single-parameter shock indicators to detect hemodynamic instability, early warning systems suffer from low accuracy, and clinicians face difficulties in continuously monitoring or assessing ICU patients. Low-frequency and low-accuracy early warning systems are unsuitable for patients whose conditions change rapidly. Furthermore, different types of clinical parameters (e.g., vital signs, physical examination results, and laboratory measurements) can be used to identify hemodynamic instability. ICU clinicians obtain a large number of these clinical parameters from multiple monitoring systems and devices. However, the limited capacity of humans to process complex information independently hinders the early identification of patient deterioration. Therefore, conventional identification and early warning methods are insufficient for real-time monitoring of patient conditions.

[0029] To address these and other potential problems, embodiments of this disclosure provide solutions for detecting hemodynamic instability. More parameters indicative of hemodynamic instability are used to improve prediction accuracy. Furthermore, multiple sub-models will use multiple clinical parameters to determine multiple prediction probabilities, and a combined prediction probability will be determined based on these multiple prediction probabilities. The combined prediction probability will be more accurate, thus enabling precise and timely detection of hemodynamic instability, thereby significantly reducing mortality and healthcare costs.

[0030] Now for reference Figure 1 , Figure 1 An example environment in which exemplary embodiments of the present disclosure may be implemented is illustrated. The example environment includes a computing device 110. The computing device 110 can be used to detect hemodynamic instability based on certain clinical parameters. Examples of the computing device 110 include, but are not limited to, personal computers, server computers, handheld or laptop devices, mobile devices, multiprocessor systems, consumer electronics, minicomputers, mainframe computers, and distributed computing environments that include any of the above systems or devices.

[0031] like Figure 1As shown, computing device 110 can receive multiple clinical parameters of an object (e.g., a patient in a hospital) at various points in time. In some embodiments, the multiple clinical parameters include at least two of the following: vital signs data 102, laboratory measurement data 104, waveform feature data 106, and image data 108.

[0032] Vital signs data 102 are measurements of the most basic functions of the human body and include at least one of the following: heart rate (HR), temperature, respiratory rate, blood pressure, etc. Laboratory measurement data 104 are obtained from a laboratory and include at least one of the following: creatinine, blood glucose, glutamate, lactate, alanine aminotransferase, aspartate aminotransferase, etc. Waveform characteristic data 106 include at least one of the following: peak-to-trough slope of the heart rate waveform, peak-to-peak interval of the heart rate waveform, peak-to-trough slope of the ventilator waveform, peak-to-peak interval of the ventilator waveform, etc. Image data 108 includes at least one of the following: X-ray image, ultrasound image, computed tomography image, etc.

[0033] After receiving multiple clinical parameters corresponding to a time point, computing device 110 can use machine learning model 112 to detect hemodynamic instability. Machine learning model 112 includes sub-models 114-1, 114-2, ..., and 114-M, where M is an integer. For convenience, sub-models 114-1, 114-2, ..., and 114-M are collectively referred to as sub-model 114. Each sub-model in sub-model 114 is used to generate a predicted probability based on multiple clinical parameters. The predicted probabilities of sub-model 114 are used to determine the combined predicted probability. Sub-model 114 can be implemented using machine models. For example, sub-model 114 can be a random forest, extreme gradient boosting, convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory (LSTM), LSTM+attention, CNN+bidirectional long short-term memory (BiLSTM)+attention, etc. In some embodiments, different types of machine models are trained using the same training data. The machine model with the best training results from the different types of machine models is selected as the sub-model. In some embodiments, the sub-model is a machine model of the same kind. In some embodiments, the sub-model is a machine model of a different kind.

[0034] The computing device 110 can also determine whether hemodynamic instability exists based on the combined predicted probabilities of the machine learning model 112. Figure 1 An example of computing device 110 including machine learning model 112 is shown. In some embodiments, computing device 110 includes multiple machine learning models to generate multiple combined prediction probabilities, and computing device 110 uses the multiple combined prediction probabilities to warn of hemodynamic instability.

[0035] Now for reference Figure 2 , Figure 2 The illustration shows a model structure 200 for detecting hemodynamic instability according to an example embodiment of the present disclosure. Figure 2 In this model, machine learning models 202-1, 202-2, 202-3, ..., and 202-N form a detection model for detecting hemodynamic instability 204, where N is an integer. For convenience, machine learning models 202-1, 202-2, 202-3, ..., and 202-N are collectively referred to as machine learning model 202.

[0036] like Figure 2 As shown, machine learning model 202-1 can be compared with... Figure 1 The machine learning model 112 shown is identical, and includes multiple sub-models 114-1, 114-2, ..., and 114-M. Furthermore, each machine learning model in machine learning model 202 has the same model structure and corresponds to different time points. For example, machine learning model 202-1 corresponds to time point T1, machine learning model 202-2 corresponds to time point T2, machine learning model 202-3 corresponds to time point T3, ..., and machine learning model 202-N corresponds to time point TN. For example, there can be 24 machine learning models, and each of these 24 machine learning models corresponds to a time length of one hour. Although the structure of machine learning model 202 is identical, each machine learning model in machine learning model 202 is trained using sample clinical parameters from multiple sample objects at different time points. Therefore, each machine learning model in machine learning model 202 can generate combined prediction probabilities at the corresponding time point. In this case, machine learning model 202 can generate multiple combined prediction probabilities, which are used to detect hemodynamic instability 204. If hemodynamic instability 204 is detected, an alert can be generated to notify the clinician. The clinician can then provide a preventative treatment plan for the subject or patient.

[0037] Now for reference Figure 3 , Figure 3 An example process 300 for detecting hemodynamic instability is illustrated according to an example embodiment of the present disclosure.

[0038] Process 300 begins at box 302. Then, multiple data collection systems collect vital signs data 304, laboratory measurement data 306, waveform feature data 308, and image data 310 for an object. All collected data are input in a rolling manner into a time window filtering system 312 with a specific time window (the default time window can be 1 hour, but a custom default time window can be used, such as 2 hours or 30 minutes). In some embodiments, if a parameter is empty, hemodynamic instability can be detected by using other parameters without information about that parameter. In some embodiments, if a parameter is empty, alternative data for that parameter is generated within a certain rolling time window. A rolling time interpolation method is established in a rationality screening system 314 to determine alternative data for that parameter. For example, Figure 4 This is a table illustrating example clinical parameters 400 according to an example embodiment of the present disclosure. Figure 4 As shown, the age, heart rate (HR), and temperature values ​​for patient 001 are the values ​​in t0, but central venous pressure (CVP) and creatinine values ​​are not available. The system will filter these values ​​hourly before t0 to find the latest values ​​as backup data. For laboratory test indicators, the maximum screening time is 26 hours before t0. For vital signs indicators, the maximum screening time is 2 hours. If the system finds a latest value, it will set that parameter to the latest value; otherwise, the parameter will remain blank (or empty).

[0039] Return to Figure 3 The rationality screening system 314 can further remove outlier data (e.g., abnormal blood pressure of an object) from the collected data. The detection model parameter update system 320 is used to automatically calculate the parameters of the hemodynamic instability model by training it, and to automatically calculate the probability of hemodynamic instability for each object using the collected data. The detection model parameter update system 320 can be... Figure 2 The model structure 200 shown is used for implementation. Subsequently, a warning threshold is obtained from system 324 to update the warning threshold. System 324 is used to update the warning threshold. At box 326, the result obtained from the detection model is compared with the warning threshold to determine whether a warning should be issued. If it is determined that a warning should be issued, intervention is initiated at box 328. For example, a clinician will provide a preventative treatment plan to the subject. The process then ends at box 332. If it is determined that no warning should be issued, the result obtained from the detection model indicates that the patient is well and no intervention is initiated at box 330.

[0040] like Figure 3As shown, solid arrows indicate the normal operating procedure for detecting hemodynamic instability. Additionally, dashed arrows indicate that the parameters of the detection model can be updated when new patient data 308 is added to the database or data lake 316. For example, when new patient data 318 is added to the database or data lake 316, incremental learning 322 can be performed using the new dataset in the database or data lake 316. During the incremental learning process, positive samples indicate that the object has hemodynamic instability, while negative samples indicate that the object does not. In one embodiment, if a positive sample cannot be correctly detected during the incremental learning process, the process continues until a positive sample can be correctly detected. Furthermore, the warning threshold used to determine whether a patient has hemodynamic instability can be adjusted. For example, the warning threshold can be set to 0.5. Hemodynamic instability will be detected when the predicted probability of a sample is greater than or equal to 0.5. Hemodynamic instability will not be detected when the predicted probability of a sample is less than 0.5. During the actual training process, the predicted probability of a positive sample can be 0.45. If the warning threshold is set to 0.5, no positive sample will be detected. Therefore, the warning threshold can be adjusted to 0.4. At this point, a positive sample will be detected. If the computing device 110 detects a change in the medical environment in the area (e.g., the presence of a large number of patients or a change in patient age), the computing device 110 uses incremental learning to update the parameters of the sub-model of the machine learning model of the detection model.

[0041] Figure 5 This is a flowchart illustrating an example method 500 for detecting hemodynamic instability according to an example embodiment of the present disclosure. The example process can be performed in computing device 110 or any other suitable computing device.

[0042] At box 502, computing device 110 obtains multiple clinical parameters of the subject at a given time point. These multiple clinical parameters may include at least two of the following: vital signs data, laboratory measurements, waveform characteristics data, and image data. At least two of the vital signs data, laboratory measurements, waveform characteristics data, and image data are used to determine hemodynamic instability of the subject (e.g., a patient).

[0043] In some embodiments, the computing device needs to verify multiple clinical parameters of an object at a certain point in time. Therefore, the computing device 110 checks whether one of the multiple clinical parameters is empty. If the clinical parameter is not empty, it will be used. Alternatively, if the clinical parameter is empty, it will also be used to determine hemodynamic instability, or hemodynamic instability can be determined without considering the clinical parameter. Alternatively, if the computing device 110 determines that one of the multiple clinical parameters is empty, it will determine a screening period for the clinical parameter and find the latest data for the clinical parameter within that screening period. This screening period may be referred to as a first screening period. If the computing device 110 finds the latest data for the clinical parameter within the screening period, it updates the first clinical parameter to the latest data. If the computing device 110 cannot find the latest data for the clinical parameter within the first screening period, it keeps the clinical parameter empty and continues the process.

[0044] In some embodiments, the computer device 110 further checks whether another clinical parameter among a plurality of clinical parameters is empty. This other clinical parameter is different from the first clinical parameter. The computer device 110 determines that the other clinical parameter among the plurality of clinical parameters is empty. If the other clinical parameter is not empty, it is used. If the other clinical parameter is empty, the computer device 110 determines a second screening time limit for the other clinical parameter and finds the latest data for the other clinical parameter within the second screening time limit, wherein the second screening time limit is different from the first screening time limit. If the computer device 110 finds the latest data for the other clinical parameter within the second screening time limit, it updates the other clinical parameter to the latest data. If the computer device 110 does not find the latest data for the other clinical parameter within the second screening time limit, it keeps the parameter empty and continues the process. Furthermore, the screening time limit for each of the plurality of clinical parameters can be different.

[0045] At box 504, computing device 112 determines multiple predicted probabilities based on multiple clinical parameters using a machine learning model, wherein each of the multiple predicted probabilities is determined by a corresponding sub-model among multiple sub-models included in the machine learning model. Multiple sub-models can be pre-trained, and the training process will be described later.

[0046] In some embodiments, the computing device 110 inputs multiple clinical parameters into each of multiple sub-models to obtain multiple predicted probabilities. For example, if there are three sub-models in the machine learning model, three predicted probabilities are generated. Furthermore, the machine learning model is associated with corresponding time points among multiple time points. Therefore, for a corresponding time point, the multiple sub-models can output multiple predicted probabilities.

[0047] At box 506, computing device 110 determines a combined prediction probability for this time point based on multiple predicted probabilities for this time point. The combined prediction probability can be generated using multiple prediction probabilities obtained from multiple sub-models. In some embodiments, computing device 110 calculates the average probability of the multiple prediction probabilities as the combined prediction probability. In some embodiments, computing device 110 calculates a weighted average probability of the multiple prediction probabilities as the combined prediction probability. The above embodiments are illustrative of the invention and not intended to limit its scope.

[0048] At box 508, computing device 110 detects hemodynamic instability based on combined prediction probabilities. In some embodiments, combined prediction probabilities are used to determine the hemodynamic instability of an object. In some embodiments, the combined prediction probabilities are used together with some combined prediction probabilities from other machine learning models to determine the hemodynamic instability of the object.

[0049] In some embodiments, the machine learning model is one of a plurality of machine learning models. The plurality of machine learning models form a detection model for detecting hemodynamic instability, and each of the plurality of machine learning models is used to determine a combined prediction probability at a corresponding time point among a plurality of time points. The combined prediction probability is determined based on multiple clinical parameters for the corresponding time point. Therefore, there can be multiple combined prediction probabilities, each of which corresponds to one of the plurality of time points. For example, if the period for detecting hemodynamic instability is 24 hours, then 24 machine learning models are used to detect hemodynamic instability, and each of the 24 machine learning models corresponds to one hour within the 24 hours. Data from this one hour is accessed by the corresponding machine learning model.

[0050] In some embodiments, for each corresponding time point among a plurality of time points, the computing device 110 uses a machine learning model corresponding to each corresponding time point among the plurality of time points to process multiple clinical parameters obtained at each corresponding time point among the plurality of time points to generate a combined prediction probability for each corresponding time point among the plurality of time points. Thus, multiple combined prediction probabilities can be obtained for each of the multiple time points. To accurately warn of hemodynamic instability, the computing device 110 first selects a predetermined number of combined prediction probabilities from the multiple combined prediction probabilities. The selected combined prediction probabilities are greater than the remaining combined prediction probabilities among the multiple combined prediction probabilities. Next, the computing device 110 determines whether each of the predetermined number of combined prediction probabilities is greater than a predetermined probability threshold. If each of the predetermined number of combined prediction probabilities is greater than the predetermined probability threshold, the computing device 110 determines that hemodynamic instability has occurred and generates an alarm. If at least one of the predetermined number of combined prediction probabilities is less than or equal to the predetermined probability threshold, the computing device 110 determines that no hemodynamic instability has occurred. For example, the predetermined probability threshold can be set to 0.5 and the predetermined number to 16. Sixteen combined prediction probabilities are selected from the multiple combined prediction probabilities. If all 16 combined predicted probabilities are greater than 0.5, an alarm for hemodynamic instability is generated. If at least one of the 16 combined predicted probabilities is less than or equal to a predetermined probability threshold, hemodynamic instability will not occur.

[0051] In some embodiments, the predetermined number is the number of multiple machine learning models in the detection model. In this case, the computing device 110 determines that hemodynamic instability has occurred only when all combined prediction probabilities of all multiple machine learning models are greater than a predetermined probability threshold. In some embodiments, the predetermined number is one. In this case, the computing device 110 determines that hemodynamic instability has occurred only when one of the combined prediction probabilities is greater than the predetermined probability threshold. For example, the computing device 110 may select the maximum combined prediction probability from the multiple combined prediction probabilities and determine whether the maximum combined prediction probability is greater than the predetermined probability threshold.

[0052] In some embodiments, the predetermined number is determined by a predetermined ratio and the number of multiple machine learning models. The predetermined ratio reflects the model performance of the detection model and is determined based on at least one of the following: the accuracy, precision, recall, specificity, F1 score, and false positive rate of the detection model. For example, the predetermined ratio can be set to 35% based at least on accuracy and recall. If the detection model includes 24 machine learning models, the predetermined number is 24*(1-35%)≈16. The computing device 110 determines whether each of the predetermined number of combined prediction probabilities is greater than a predetermined probability threshold. In one example, the computing device ranks the multiple combined prediction probabilities in descending order and determines whether the 16th combined prediction probability is greater than the predetermined probability threshold. If the 16th combined prediction probability is greater than the predetermined probability threshold, it indicates that each of the predetermined number of combined prediction probabilities is greater than the predetermined probability threshold, and an alarm for hemodynamic instability is generated. If the 16th combined prediction probability is less than or equal to the predetermined probability threshold, no alarm is generated.

[0053] In some embodiments, the computing device 110 further obtains multiple clinical parameters for the object at each of at least two additional time points different from this time point. Then, for each of the at least two additional time points, the computing device 110 processes the multiple clinical parameters obtained at each of the at least two time points using a machine learning model corresponding to each of the at least two additional time points, respectively, to generate a combined predicted probability of hemodynamic instability for each of the at least two time points. The computing device 110 then detects hemodynamic instability based on the combined predicted probability for this time point and the combined predicted probabilities for the at least two additional time points.

[0054] In some embodiments, computing device 110 averages a first combined predicted probability at a given time point with a previous combined predicted probability at a previous time point to determine a first average predicted probability. After receiving a second plurality of clinical parameters at an additional time point, computing device 110 uses another machine learning model from a plurality of machine learning models corresponding to the additional time point to calculate a second combined predicted probability of hemodynamic instability. This additional time point may be adjacent to the given time point. Computing device 110 then averages the second combined predicted probability, the first combined predicted probability, and the previous combined predicted probability to determine a second average predicted probability. Computing device 110 also calculates the difference between the first average predicted probability and the second average predicted probability and determines whether the difference is greater than a threshold. If the difference is greater than the threshold, computing device 110 determines that hemodynamic instability exists in the object at the additional time point and generates an alarm. If the difference is less than or equal to the threshold, computing device 110 continues to detect hemodynamic instability.

[0055] In some embodiments, the computing device 110 uses a machine learning model corresponding to this time point to obtain a first combined prediction probability at this time point. Then, the computing device 110 calculates a first difference between the first combined prediction probability and the previous combined prediction probability at a previous adjacent time point. Then, the computing device 110 calculates a first average difference between the first difference and the previous differences. Previous differences are determined based on the combined prediction probabilities of any two adjacent time points of the previous time point. For example, the time point is 5 o'clock. The computing device 110 calculates the first difference by subtracting the previous combined prediction probability for 4 o'clock from the first combined prediction probability for 5 o'clock, calculates the first previous difference by subtracting the combined prediction probability for 3 o'clock from the combined prediction probability for 4 o'clock, calculates the second previous difference by subtracting the combined prediction probability for 2 o'clock from the combined prediction probability for 3 o'clock, and calculates the third previous difference by subtracting the combined prediction probability for 1 o'clock from the combined prediction probability for 2 o'clock. Then, the computing device calculates a first average difference between the first difference, the first previous difference, the second previous difference, and the third previous difference. Computer device 110 can obtain multiple clinical parameters for the next time point adjacent to the current time point and use a machine learning model corresponding to the next time point to obtain a second combined predicted probability. For example, if the current time point is 5 o'clock, the next time point is 6 o'clock. A second average difference at the next time point is calculated in a similar manner. If the difference between the first average difference and the second average difference is greater than a threshold, computing device 110 determines that there is hemodynamic instability in the object at the next time point and generates an alarm. If the difference is less than or equal to the threshold, the computing device continues to detect hemodynamic instability.

[0056] Now for reference Figure 6 , Figure 6 The illustration shows a block diagram of an example environment for training a machine learning model to detect hemodynamic instability, according to an example embodiment of the present disclosure. Example environment 600 includes a computing device 610. The computing device 610 is capable of training a machine learning model for detecting hemodynamic instability based on some clinical parameters. In some embodiments, the computing device 610 and... Figure 1 The computing device 110 shown is the same device. In some embodiments, computing device 610 and Figure 1 The computing device 110 shown is a different device.

[0057] like Figure 6 As shown, the computing device 610 can receive multiple sample clinical parameters from multiple sample subjects. In some embodiments, the multiple sample clinical parameters include at least two of the following: vital signs data 602, laboratory measurement data 604, waveform feature data 606, and image data 608.

[0058] After receiving multiple clinical parameters from multiple sample objects, these clinical parameters are used as dataset 612. As mentioned earlier, hemodynamic instability is not common in the real world, and there are not many positive cases (hemodynamically unstable patients), while there are far more negative cases (hemodynamically stable patients) than positive cases. The training data is extremely imbalanced. If such extremely imbalanced data is fed into the training model, the final performance will be very poor because the final prediction will be for negative cases (hemodynamically stable patients) rather than positive cases (hemodynamically unstable patients). Therefore, a new training framework specifically for imbalanced data is provided. Therefore, computing device 610 divides dataset 612 into multiple subsets, for example, subset 614-1, subset 614-2, ..., and subset 614-T, where T is an integer. For convenience, subsets 614-1, 614-2, ..., and 614-T are collectively referred to as subset 614. These multiple subsets can be balanced according to the ratio of positive to negative cases. For example, dataset 612 contains a total of 3671 positive cases (hemodynamically unstable patients) and 14767 negative cases (hemodynamically stable patients), with a positive-to-negative case ratio close to 1:3. Therefore, three balanced subsets are generated, and three sub-models are trained using these balanced subsets, each with a predetermined number of cross-validations. For example, the predetermined number is set to 5. Each of the three sub-models contains the same number of positive cases and one subset of different negative cases, with the number of identical positive cases equal to the number of identical subsets of different negative cases.

[0059] Computing device 610 can use multiple subsets of datasets to train a machine learning model 616 for detecting hemodynamic instability. Machine learning model 616 includes sub-models 618-1, 618-2, ..., and 618-M, where M is an integer. For convenience, sub-models 618-1, 618-2, ..., and 618-M are collectively referred to as sub-model 618. Each sub-model in sub-model 618 is trained using at least one subset of dataset 614. During training, the parameters of each sub-model in sub-model 618 are determined by comparing the predicted probabilities of each sub-model in sub-model 618 with the actual results of hemodynamic instability. Finally, the predictions of these sub-models are combined to form the combined predicted probabilities of the machine learning model.

[0060] The computing device 610 can also determine the presence of hemodynamic instability based on the predicted probability of a machine learning model 620. Figure 6 The illustration shows a computing device 610 including machine learning models. The computing device 610 is used to describe the present disclosure and not to limit it. In some embodiments, the computing device 610 includes multiple machine learning models capable of generating multiple combined predictive probabilities. The computing device 610 uses these combined predictive probabilities to detect hemodynamic instability and generate alerts for it.

[0061] Now for reference Figure 7 , Figure 7The illustration shows a model structure for detecting hemodynamic instability according to an example embodiment of the present disclosure. For hemodynamic instability, there are fewer positive cases 702 and more negative cases 704. Therefore, the number of positive cases 702 is unbalanced with the number of negative cases 704. Therefore, the negative cases 704 are divided into distinct negative cases 706-1, 706-2, ..., and 706-M. Then, the same positive cases 702 are combined with each negative case in the negative cases 706-1, 706-2, ..., and 706-M to form a subset of data for training a sub-model. For example, positive cases 702 are combined with negative cases 706-1 to train sub-model 708-1, positive cases 702 are combined with negative cases 706-2 to train sub-model 708-2, ..., and positive cases 702 are combined with negative cases 706-M to train sub-model 708-M, where M is an integer. During training, the parameters of the sub-models are updated. After training, sub-models 708-1, 708-2, ..., and 708-M form machine learning model 710. The output of the machine learning model is based on the output of sub-model 708. For example, the output of machine learning model 710 can be generated by averaging multiple predicted probabilities of sub-model 708 (e.g., combined predicted probabilities). The number of sub-models 708 can be determined based on the positive / negative case ratio. For example, if the positive / negative case ratio is 3, the number of sub-models 708 can be 3.

[0062] For example, there are a total of 3671 hemodynamically unstable cases (positive) and 14767 hemodynamically stable cases (negative). The positive / negative ratio is close to 1:3, so the negative cases are divided into 3 groups to be paired with the positive cases. Therefore, the positive cases are repeated 3 times, each paired with one of the negative groups, allowing three sub-models to be trained. The parameters of each of the three models are updated using the corresponding residuals. After this, the three sub-models are combined. The first model is used as a classifier to predict the outcome, the predictions are compared to the ground truth, and a first residual is generated. If this residual is close to zero, it indicates a good result. However, it may not be successful in the first model because it is a weak classifier. Therefore, the first model keeps the case data with correct predictions and feeds the case data associated with incorrect predictions to the second model to obtain a lower residual value. The remaining models can be done in the same way until the lowest residual is obtained. For example, if the first model cannot correctly detect a positive case during training, that positive case can be used to train the second model. If the second model is able to correctly detect positive cases, a lower residual is obtained.

[0063] Now for reference Figure 8 , Figure 8 The illustration depicts a process for training a machine learning model according to an example embodiment of the present disclosure. Figure 8 In this context, the hospital patient database 802 stores data related to a large number of sample patients used to train machine learning models. Multiple sample clinical parameters 804 can be obtained from the hospital patient database 802. These sample patients include both positive and negative cases. The multiple sample clinical parameters 804 include vital sign data, laboratory measurement data, waveform feature data, and image data. Figure 8 The embodiments described herein are provided for illustrative purposes and not for limitation. In some embodiments, the plurality of sample clinical parameters 804 include at least two of the following: vital signs data, laboratory measurement data, waveform feature data, and image data.

[0064] Multiple clinical parameters from multiple sample subjects can change over time. Therefore, it is possible to obtain clinical parameters from multiple samples at different time points. For example... Figure 8 As shown, there are data segments 806-1, 806-2, ..., 806-N-1 and 806-N corresponding to time points T1, T2, ..., TN-1 and TN, respectively. Each data segment includes multiple sample clinical parameters of multiple sample objects. Each row of the data segment represents the data of the sample object. Data segment 806-1 corresponding to time point T1 includes multiple sample clinical parameters 808 of multiple sample objects and the true results 810 of hemodynamic instability of multiple sample objects. In some embodiments, the multiple sample objects include 3 positive sample objects and 9 negative sample objects. The same number of positive and negative samples are combined to train the sub-model. Therefore, the sample clinical parameters of the 3 positive sample objects are combined with the sample clinical parameters of the 3 negative samples to form 3 sub-data segments 812-1, 812-2 and 812-3 to train 3 sub-models, thereby obtaining prediction results 814. Prediction results 814 include 3 results, each of which is obtained from one of the 3 sub-models. The predicted result 814 was compared with the actual results of hemodynamic instability in the sample subjects to adjust the parameters of the three sub-models. The predicted result 814 was used to generate the combined predicted probability 816.

[0065] like Figure 8As shown, N data segments are used to train N machine learning models. These N machine learning models form a detection model for detecting hemodynamic instability. For each machine learning model, the combined predicted probability is compared to a predetermined probability threshold. Therefore, a predetermined probability threshold is needed to improve performance. As mentioned earlier, in one example, the default value for the predetermined probability threshold is 0.5. In some embodiments, the default value can be updated based on the location of the medical environment data and the clinician's experience.

[0066] The detection model is a dynamic, time-varying model used to monitor patients, and the predetermined ratios related to model performance can be determined based on accuracy, precision, recall, specificity, F1 score, false positive rate, true positive rate (TPR), and true negative rate (TNR). The following formulas are used to calculate the above parameters related to model performance. More specifically, the following formula (1) determines true positives.

[0067] (1) in, Let n represent the predicted value, t represent the patient count, t represent the time in hours, and g represent the baseline true value. This indicates that the baseline truth value for patient i is 1. This indicates that the baseline truth value for patient i is 0, and This represents the predetermined ratio as discussed above. The following formula (2) determines the sum of true positives and false positives.

[0068] True positive (TP) + False positive (FP) = (2) The following formula (3) is used to determine true negatives.

[0069] True negative (TN) = (3) The following formula (4) determines the sum of true positives and false negatives.

[0070] True positive (TP) + False negative (FN) = (4) The accuracy rate is determined according to the following formula (5).

[0071] , (5) The accuracy rate is determined according to the following formula (6).

[0072] , (6) The recall rate is determined according to the following formula (7).

[0073] , (7) Specificity is determined according to the following formula (8).

[0074] , (8) The false positive rate (FPR) is determined according to the following formula (9).

[0075] (9) The F1 score is determined according to the following formula (10).

[0076] , (10) Now for reference Figure 9 , Figure 9 An example method 900 for training a machine learning model for detecting hemodynamic instability, according to an example embodiment of the present disclosure, is illustrated. The example process can be performed on computing device 610 or any other suitable computing device.

[0077] At box 902, computing device 610 obtains multiple sample clinical parameters of multiple sample subjects at a single time point. The multiple sample clinical parameters include at least two of the following: vital signs data, laboratory measurement data, waveform feature data, and image data.

[0078] In some embodiments, when multiple clinical parameters of multiple sample objects are obtained at a point in time, the computing device 610 checks each of the multiple clinical parameters. The computing device 610 may determine whether a clinical parameter among the multiple clinical parameters is empty. If a clinical parameter among the multiple clinical parameters is not empty, it is used. If a clinical parameter among the multiple clinical parameters is empty, the computing device 610 determines a first screening time limit for the clinical parameter. Then, the computing device 610 checks whether the latest data for the clinical parameter is within the first screening time limit. If the latest data for the clinical parameter is within the first screening time limit, the clinical parameter is updated to the latest data. If the latest data for the clinical parameter is not within the first screening time limit, the clinical parameter remains empty. The screening time limits for multiple clinical parameters may be the same or different.

[0079] At box 904, computing device 610 determines the ratio of the number of positive samples to the number of negative samples in a plurality of sample objects. The plurality of sample objects may include both positive and negative samples. However, hemodynamically unstable positive and negative samples are not balanced. Therefore, it is necessary to balance the positive and negative samples based on this ratio.

[0080] At box 906, computing device 610 generates multiple subsets. Each subset includes sample clinical parameters corresponding to positive samples and sample clinical parameters corresponding to a portion of negative samples. Negative samples are divided into multiple parts based on this ratio, and this part of the negative samples for each subset is different from each other. To obtain subsets for training machine learning models, computing device 610 combines the sample clinical parameters of positive samples with the sample clinical parameters of multiple different portions of negative samples to form multiple subsets. Therefore, each subset includes data from the same positive samples and data from different negative samples.

[0081] At box 908, computing device 610 trains multiple sub-models in a machine learning model for detecting hemodynamic instability based on multiple subset datasets. Computing device 610 uses multiple subset datasets to train multiple sub-models. In some embodiments, one subset of the multiple subset datasets is used to train one sub-model. In some embodiments, two or more subsets of the multiple subset datasets may be used to train one sub-model.

[0082] In some embodiments, when training multiple sub-models in a machine learning model, the computing device 610 obtains the predicted probabilities of the sub-models by inputting a subset of ...

[0083] In some embodiments, the machine learning model is one of a plurality of machine learning models that form a detection model for detecting hemodynamic instability. The computing device 610 obtains additional clinical parameters of multiple sample objects at other time points different from this time point. For example, this time point is 5 o'clock, and the other time point is 6 o'clock. The computing device 610 uses the additional clinical parameters of the additional samples to train another machine learning model among the plurality of machine learning models. In this way, the computing device trains multiple machine learning models using data corresponding to time points. Each of the plurality of machine learning models is associated with a time point.

[0084] Now for reference Figure 10A , Figure 10B and Figure 10C , Figure 10A , Figure 10B and Figure 10C A table showing the model performance of a detection model for detecting hemodynamic instability according to an example embodiment of the present disclosure is illustrated. Figure 10A Example Table 1000A shows the model performance for TPEVGH. Example Table 1000A is the internal test results of the detection model used to detect hemodynamic instability. Figure 10B Example Table 1000B is shown, which is an external test result of a detection model for detecting hemodynamic instability. Figure 10C Example Table 1000C shows the model performance for MIMICIV. Example Table 1000C also represents external test results for a detection model used to detect hemodynamic instability. Figure 10A , Figure 10B and Figure 10C As shown, the receiver operating characteristic area under the curve (AUROC), area under the precision-recall curve (AUPRC), positive predictive value (PPV), negative predictive value (NPV), precision, F1 score, and specificity are significantly improved. The model demonstrates good positive case detection capabilities for both internal and external test results.

[0085] Now for reference Figure 11 , Figure 11 A schematic block diagram of an example device 1100 for implementing embodiments of the present disclosure is shown. Figure 1 The computing device 110 and Figure 6 The computing device 610 in the middle can be implemented by this device. For example Figure 11 As shown, device 1100 includes a central processing unit (CPU) 1101, which is capable of performing various appropriate actions and processes based on computer program instructions stored in read-only memory (ROM) 1102 or loaded from storage unit 1108 into random access memory (RAM) 1103. RAM 1103 is also capable of storing all kinds of programs and data required for the operation of device 1100. CPU 1101, ROM 1102, and RAM 1103 are interconnected via bus 1104. Input / output (I / O) interface 1105 is also connected to bus 1104.

[0086] Multiple components in device 1100 are connected to I / O interface 1105. These components include: input unit 1106, such as a keyboard, mouse, etc.; output unit 1107, such as various displays and speakers, etc.; storage unit 1108, such as a hard disk and optical disk, etc.; and communication unit 1109, such as a network card, modem, wireless transceiver, etc. Communication unit 1109 allows device 1100 to exchange information / data with other devices via computer networks such as the Internet and / or various telecommunications networks.

[0087] The aforementioned processes and procedures (e.g., methods 500 and 900) can be executed by processing unit 1101. For example, in some embodiments, methods 500 and 900 can be implemented as computer software programs tangibly contained in a machine-readable medium (e.g., storage unit 1108). In some embodiments, the computer program can be partially or completely loaded and / or installed into device 1100 via ROM 1102 and / or communication unit 1109. When the computer program is loaded into RAM 1103 and executed by CPU 1101, one or more actions of methods 500 and 900 can be implemented.

[0088] This disclosure can be a method, apparatus, system, and / or computer program product. The computer program product can include a computer-readable storage medium on which computer-readable program instructions for performing various aspects of this disclosure are loaded.

[0089] A computer-readable storage medium can be a tangible means for maintaining and storing instructions used by an instruction execution device. A computer-readable storage medium can be, but is not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (not an exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable optical disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, punch cards on which instructions are stored, or protrusions in slots, and any suitable combination of the foregoing. The computer-readable storage medium as used herein is not to be interpreted as a transient signal itself, such as radio waves or freely propagating electromagnetic waves, electromagnetic waves propagating via waveguides or other transmission media (e.g., optical pulses propagating via fiber optic cables), or electrical signals propagating via wires.

[0090] The described computer-readable program instructions can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via the Internet, local area network, wide area network, and / or wireless network. The network can include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, network gate computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them for storage in the computer-readable storage medium of each computing / processing device.

[0091] The computer program instructions used to perform the operations of this disclosure can be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​(e.g., Smalltalk, C++, etc.) and traditional procedural programming languages ​​(e.g., "C" language or similar programming languages). The computer-readable program instructions can be implemented entirely on a user's computer, partially on a user's computer, as a standalone software package, partially on a user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer via any type of network (including local area networks (LANs) and wide area networks (WANs)) or connected to an external computer (e.g., via an internet connection using an internet service provider). In some embodiments, the state information of the computer-readable program instructions is used to customize electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs). The electronic circuitry is capable of executing the computer-readable program instructions to implement various aspects of this disclosure.

[0092] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0093] Computer-readable program instructions can be provided to a processing unit of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus to manufacture a machine, such that, when executed by the processing unit of the computer or other programmable data processing apparatus, the instructions generate means for implementing the functions / actions specified in one or more blocks of a flowchart and / or block diagram. Computer-readable program instructions can also be stored in a computer-readable storage medium and cause a computer, programmable data processing apparatus, and / or other device to operate in a particular manner, such that the computer-readable medium storing the instructions includes an article of manufacture (including instructions for implementing aspects of the functions / actions specified in one or more blocks of a flowchart and / or block diagram).

[0094] Computer-readable program instructions can also be loaded into a computer, other programmable data processing apparatus, or other device to perform a series of operational steps to generate a computer-implemented process. Therefore, instructions that execute on a computer, other programmable data processing apparatus, or other device implement the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0095] The flowcharts and block diagrams in the accompanying drawings illustrate system architectures, functions, and operations that can be implemented by systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram can represent a portion of a module, program segment, or code, wherein the module and program segment or code portion includes one or more executable instructions for performing a specified logical function. In some alternative embodiments, it should be noted that the functions indicated in the blocks can also occur in a different order than indicated in the drawings. For example, it is actually possible to execute two consecutive blocks in parallel or sometimes in reverse order depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented by a hardware-based system dedicated to performing the specified functions or actions, or by a combination of dedicated hardware and computer instructions.

[0096] Various embodiments of this disclosure have been described above, and the descriptions therein are merely exemplary and not exhaustive, and are not limited to the embodiments of this disclosure. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the various embodiments explained. The terminology chosen herein is intended to best explain the principles and practical application of each embodiment, as well as the technical improvements made to each embodiment in the market, or to enable those skilled in the art to understand the embodiments of this disclosure.

Claims

1. A method for training a machine learning model for detecting hemodynamic instability, comprising: Obtain multiple clinical parameters of multiple sample subjects at a single time point, wherein the multiple clinical parameters include at least two of the following: vital signs data, laboratory measurement data, waveform feature data, and image data; Determine the ratio of the number of positive samples to the number of negative samples among the plurality of sample objects; Multiple subsets are generated, each subset including sample clinical parameters corresponding to the positive samples and sample clinical parameters corresponding to a portion of the negative samples, wherein the negative samples are divided into multiple portions based on the ratio, and the portions of the negative samples in each subset are different from each other; as well as Multiple sub-models in the machine learning model used to detect hemodynamic instability are trained based on the multiple subset datasets.

2. The method according to claim 1, wherein, The machine learning model is one of a plurality of machine learning models, which together form a detection model for detecting the hemodynamic instability, and the method further includes: Obtain additional sample clinical parameters of the plurality of sample subjects at other time points different from the stated time point; and Another machine learning model among the multiple machine learning models is trained based on the additional clinical parameters of the other samples.

3. The method according to claim 1, wherein, Obtaining multiple sample clinical parameters from multiple sample subjects at a single time point includes: Determine whether any of the sample clinical parameters among the multiple sample clinical parameters is empty; In response to determining that one of the plurality of sample clinical parameters is empty, it is determined whether the latest data for the sample clinical parameter exists within a first screening time limit; and In response to determining that the latest data for the clinical parameters of the sample exists within the first screening time limit, the clinical parameters of the sample are updated based on the latest data.

4. The method according to claim 3, wherein, Training multiple sub-models in the machine learning model for detecting hemodynamic instability based on the multiple subset datasets includes: The predicted probability of the sub-model is obtained by inputting a subset of the plurality of subsets of data into a sub-model of the plurality of sub-models; and The model parameters of the sub-model are adjusted based on the predicted probabilities and the actual results of hemodynamic instability.

5. An electronic device, comprising: At least one processor; as well as At least one memory storing a plurality of instructions, which, when executed by the at least one processor, cause the device to perform the method according to any one of claims 1-4.

6. A computer-readable medium having computer instructions stored thereon, the computer instructions, when executed by a processor, causing the processor to perform the method according to any one of claims 1-4.

Images