Method and device for determining the onset of sepsis

Abstract

This invention provides a method and device for determining the onset of sepsis. [Solution] The method includes receiving at least one medical dataset containing multiple patient-related features 201; extracting one or more patient-related features from the medical dataset 202, including parameters that are indicators of sepsis; imputing at least one missing value in the medical dataset related to the features in the medical dataset 204; determining output parameters indicating the onset of sepsis in the patient by using one or more features and at least one missing value in the medical dataset as input to one or more trained machine learning models 206; and generating an alert indicating the onset of sepsis in the patient if the output parameters meet predetermined criteria related to sepsis 207.

Description

Technical Field

[0001] The present invention relates to a method and a device for determining the onset of sepsis in a patient.

Background Art

[0002] Sepsis is an extreme immune response of an individual's body to an infectious disease. Sepsis can be life-threatening and can cause a series of reactions throughout the body, which may lead to tissue damage, organ failure, and death. Early detection of sepsis is considered one of the important aspects for improving the treatment outcome of sepsis. Clinical criteria to assist in the recognition of sepsis are widely available, but the fundamental needs for early detection and treatment of sepsis remain unmet. Existing methods for early detection of sepsis do not achieve the level of specificity and sensitivity required for accurate detection of sepsis in a hospital environment. Furthermore, existing methods for early detection of sepsis are impaired by limitations such as imbalance in medical datasets and distortion of results.

[0003] Currently, there is no method that can early detect the onset of sepsis in a patient with high specificity. Therefore, there is a need for an effective and accurate method and device that enable timely determination of the onset of sepsis in a patient.

Summary of the Invention

Problems to be Solved by the Invention

[0004] Therefore, an object of the present invention is to provide a method and a device that enable effective determination of the onset of sepsis in a patient at an early stage.

Means for Solving the Problems

[0006] The method further includes determining an output parameter indicating the onset of sepsis in a patient by using one or more features as input to a trained machine learning model. The trained machine learning model is configured to process one or more features in a medical dataset and generate an output parameter indicating the onset of sepsis in a patient. For example, the output parameter may indicate the onset of sepsis in a patient or the absence of sepsis in a patient. In one embodiment, the trained machine learning model may be a recurrent neural network having long short-term memory. The method further includes determining an output parameter indicating the onset of sepsis in a patient. The method includes generating an alert indicating the onset of sepsis. The alert is generated, for example, when output parameters generated by a trained machine learning model meet predetermined criteria related to sepsis. In one embodiment, the alert is generated on an output device used by the user. The alert may be, for example, a notification indicating the onset of sepsis in a patient. The notification is generated on the graphical user interface of the output device. Alternatively, the notification may further include audible and / or haptic feedback provided to the output device. Advantageously, the method enables early identification of the onset of sepsis in a patient. Thus, the patient can be treated effectively and in a timely manner.

[0007] According to one embodiment, the method further includes normalizing values ​​associated with one or more features in a medical dataset, which are provided as input to a trained machine learning model. In one embodiment, normalizing the values ​​of one or more features includes defining uniform minimum and maximum thresholds for each of the one or more features in the medical dataset. Generally, each feature in the medical dataset may have a variety of minimum and maximum thresholds. Normalization allows all thresholds to be at the same level so that the trained machine learning model can effectively analyze the dataset. In one embodiment, the minimum and maximum thresholds for a feature are redefined to a range from a minimum value of 1 to a maximum value of 5. If the value of one or more patient-related features falls outside the range of 1 to 5, that value is replaced by the minimum and maximum thresholds associated with the feature. It should be noted that using large values ​​for threshold normalization may result in the loss of temporal changes in features over a period of time. Therefore, the thresholds are limited to a minimum value of 1 and a maximum value of 5. Advantageously, the present invention enables the capture of even minimal changes in the values ​​of one or more patient-related features. This enables the achievement of high specificity in determining the onset of sepsis.

[0008] According to another embodiment, imputing at least one missing value in a medical dataset involves identifying a missing value associated with a feature in the medical dataset. In one embodiment, a value associated with one or more features in the medical dataset may be missing if that value may not have been captured / recorded for a patient in a given time interval. Alternatively, a value associated with one or more features in the medical dataset is considered missing if that value is unlikely or exceeds a reasonable threshold. The method further includes determining a value preceding the missing value associated with the feature. The value preceding the missing value may be the last value captured / recorded for a patient prior to a given time interval. The method further includes replacing the missing value with the value preceding the missing value associated with the feature in the medical dataset. In one embodiment, the replacement of the missing value with the preceding value is performed over a limited period over which values ​​are recorded. This period may depend on the type of one or more features in the medical dataset. For example, a period defined for vital signs may be limited to a range of 4-6 hours, and a period defined for laboratory parameters may be limited to a range of 22-24 hours. Advantageously, the medical dataset becomes more complete and useful through a trained machine learning model. Therefore, the accuracy of determining whether a patient has developed sepsis improves.

[0009] According to one embodiment, one or more features in a medical dataset include at least one of the following: patient-related vital signs, analytes present in the patient's blood sample, patient-related demographic data, and derived parameters related to the vital signs and analytes present in the patient's blood sample. For example, patient-related vital signs include, but are not limited to, patient-related heart rate, pulse oximetry measurements, patient body temperature, systolic blood pressure, mean arterial pressure, diastolic blood pressure, and respiratory rate. For example, analytes present in the patient's blood sample include blood urea nitrogen, creatinine, lactate, bilirubin, leukocytes, platelets, blood pH, and serum glucose. These are not limited to the above. Patient demographic data may include, for example, patient-related age and sex. Patient-related derived characteristics may include, but are not limited to, the patient-related shock index, blood urea nitrogen to creatinine ratio, modified early warning score (MEWS), and partial SOFA (sequential organ failure assessment) score.

[0010] According to another embodiment, the output parameter indicating the onset of sepsis in a patient is a probability value generated by a trained machine learning model. The machine learning model is configured to process one or more features and generate a probability value associated with the onset of sepsis in a patient. The advantage of the present invention is that the probability value is based on one or more features associated with the patient. Therefore, a change in the value associated with one or more features also changes the probability value, thereby providing an accurate measure of the onset of sepsis in the patient.

[0011] According to one embodiment, generating an alert indicating the onset of sepsis in a patient involves determining whether a probability value related to the patient exceeds a first predetermined threshold. The first predetermined threshold may be a threshold related to a probability value generated by a machine learning model. If the probability value exceeds the first predetermined threshold, a warning or alert is generated. For example, the first predetermined threshold may be in the range of 0.4 to 1.0, preferably in the range of 0.5 to 1.0. In one embodiment, the first predetermined threshold is modified according to the user's preference related to the patient. In another embodiment, if the probability value falls within the first predetermined threshold, a warning is generated. Advantageously, if the probability of sepsis developing in a patient is identified or increases, a user (such as a doctor or healthcare professional) is promptly alerted. This allows for timely medical action to be taken to save the patient's life.

[0012] In yet another embodiment, generating an alert further includes identifying the number of generated warnings indicating the onset of sepsis. For example, the number of generated warnings is identified with respect to a defined period during which the patient is monitored. Furthermore, a determination is made as to whether the number of generated warnings exceeds a second predetermined threshold. If the number of generated warnings exceeds the second predetermined threshold, an alarm is generated indicating the onset of sepsis in the patient. For example, the second predetermined threshold may depend on the period during which the patient is monitored. For example, if the defined period is 2 hours, the second predetermined threshold may be 2 warnings. Thus, if the number of generated warnings exceeds 2 within 2 hours, an alarm is generated indicating the onset of sepsis. In one embodiment, the generated alarm may include a notification to the patient indicating the number of generated warnings / alerts. The alarm, along with the notification, is associated with an audible or tactile output. The advantage of the present invention is that the alarm allows the user to identify the progression of the disease in the patient's body at an early stage. Thus, necessary measures can be taken to improve the patient's condition in a timely manner.

[0013] In a preferred embodiment, the trained machine learning model is, in particular, a recurrent neural network including long short-term memory blocks. In particular, a recurrent neural network is an artificial neural network in which the connections between nodes form a directed graph along a temporal sequence. In particular, a recurrent neural network can be interpreted as a directed acyclic graph. In particular, a recurrent neural network can be a finite-impulse recurrent neural network or an infinite-impulse recurrent neural network (where the finite-impulse network can be expanded and replaced with a strict feedforward neural network, and the infinite-impulse network cannot be expanded and replaced with a strict feedforward neural network). In particular, a recurrent neural network may include additional memory states or additional network structures that incorporate time delays or constitute feedback loops.

[0014] Similarly, a recurrent neural network can also be defined as a neural network whose output depends not only on input values ​​and edge weights, but also on a hidden state vector, which is based on previous inputs used in the recurrent neural network. According to a further aspect of the present invention, the recurrent neural network includes at least one long short-term memory (LSTM) block. In particular, the LSTM block includes a cell, an input gate, an output gate, and a forget gate, where the cell corresponds to the hidden vector, and the input gate, output gate, and forget gate control the inflow of information into and out of the cell. In particular, by using a cell, the LSTM block can prevent divergence and vanishing gradient problems that can occur when training other types of recurrent neural networks.

[0015] Advantageously, recurrent neural networks enable temporal analysis of data, thereby considering not only real-time input datasets but also datasets that may have been analyzed by the neural network at previous times. Furthermore, LSTM blocks are well-suited for input data divided into diverse or unknown time intervals.

[0016] The object of the present invention is also achieved by a method for training a machine learning model to determine the onset of sepsis in a patient. The method comprises receiving a patient-related medical dataset, the medical dataset containing multiple patient-related features. The medical dataset is received from a source such as a medical database. Furthermore, the method comprises extracting one or more features from multiple features within the medical dataset, the one or more features containing patient-related parameters that are indicators of the onset of sepsis at a given point in time. The one or more features may reflect the patient's medical status in real time or near real time. Thus, one or more features in the medical dataset may indicate the onset of sepsis at a given point in time. Furthermore, the method comprises receiving a machine learning model and having the model determine a probability value regarding the onset of sepsis in the patient based on one or more features in the medical dataset. In one embodiment, the probability value is determined based on a value associated with one or more features in the medical dataset.

[0017] The method further includes receiving sepsis data associated with a medical dataset, the sepsis dataset indicating the onset of sepsis or the absence of sepsis during a defined period in a patient associated with the medical dataset. In one embodiment, the sepsis dataset may include a medical dataset labeled to indicate the onset of sepsis or the absence of sepsis during a defined period in a patient. In a further embodiment, the labeled medical dataset is associated with multiple patients whose medical history is monitored and who are being treated for sepsis. The labeled medical dataset includes one or more features recorded at regular time intervals, thereby showing fluctuations in values ​​associated with one or more features over those time intervals. In an alternative embodiment, the sepsis data may include an analysis of one or more features present in the patient-associated medical dataset, which may be data received from a physician / specialist indicating the onset of sepsis or the absence of sepsis.

[0018] This method further includes adjusting the machine learning model based on the results of comparing probability values ​​with sepsis data. The comparison may indicate the accuracy of the probability values ​​generated by the machine learning model. Therefore, if a difference between the probability values ​​and sepsis data is identified in this comparison, the machine learning model is adjusted. Advantageously, the machine learning model becomes more robust, thereby improving the accuracy of the probability values ​​generated by the model. Thus, the determination of sepsis onset in patients is made effectively and in a timely manner.

[0019] According to one embodiment, the method further includes preprocessing sepsis data. Preprocessing may further include imputing at least one missing value in the sepsis data. Sepsis data includes one or more feature values ​​captured / recorded at regular time intervals. If a feature value is missing in the dataset, such value is imputed based on a preceding value in the medical dataset. In one embodiment, a threshold is defined for such imputation based on the nature of the feature for which the value is missing. For example, if the missing value is related to a patient's vital sign, the imputation of the missing value is performed using a value in the range of 4-6 hours preceding the missing value. Similarly, if the missing value is related to an analyte in a blood sample, the imputation of the missing value is performed using a value in the range of 22-24 hours preceding or following the missing value. Advantageously, imputing values ​​in sepsis data makes the data more complete and useful.

[0020] Preprocessing of sepsis data further includes normalizing one or more features from the sepsis data, which involves defining uniform minimum and maximum thresholds for each of the one or more features. Advantageously, normalization allows all thresholds to be at the same level, enabling effective analysis of the dataset by machine learning models. This allows for high specificity in determining the onset of sepsis.

[0021] According to yet another embodiment of the present invention, the label associated with sepsis data is modified to advance the defined period indicated by the label by 2 to 10 hours, preferably 3 to 9 hours, more preferably 6 to 8 hours, and most preferably 5 to 7 hours. Advantageously, modifying the defined period provides a look-ahead time of approximately 5 to 7 hours. Thus, the machine learning model is trained to determine the onset of sepsis in a patient well in advance.

[0022] The object of the present invention is also achieved by a sepsis diagnosis device for determining the onset of sepsis in a patient. The device includes one or more processing units and a medical database connected to one or more processing units, which includes multiple patient-related medical datasets and sepsis data. The device further includes memory connected to one or more processing units. The memory includes a sepsis diagnosis module configured to perform the method steps described above using at least one trained machine learning model.

[0023] In one embodiment, the present invention relates to a computer program product including a computer program, the computer program being loadable into a system's storage unit, and including a program code section causing the system to execute a method according to one aspect of the present invention when the computer program is executed on the system.

[0024] In one aspect, the present invention relates to a computer-readable medium on which a program code section of a computer program is stored, wherein the program code section is loadable and / or executable in a system so as to cause the system to execute a method according to one aspect of the present invention when the program code section is executed in the system.

[0025] Implementation of the present invention by computer program products and / or computer-readable media has the advantage that existing management systems can be easily adapted by software updates to operate as proposed by the present invention.

[0026] A computer program product can be, for example, a computer program or can be composed of elements other than a computer program. Such other elements can be, for example, hardware such as a memory device in which a computer program is stored, a hardware key for using the computer program, and / or, for example documentation for using the computer program or software such as a software key.

[0027] Hereinafter, the present invention will be further described with reference to the illustrated embodiments shown in the accompanying drawings:

Brief Description of the Drawings

[0028] [Figure 1] It is a diagram showing a block diagram of a sepsis determination device in which one embodiment for determining the onset of sepsis in a patient can be implemented. [Figure 2] It is a diagram showing a flowchart of a method for determining the onset of sepsis in a patient according to one embodiment of the present invention. [Figure 3] It is a diagram showing a flowchart of a method for complementing missing values in a medical dataset according to one embodiment of the present invention. [Figure 4] It is a diagram showing a flowchart of a method for generating an alert indicating the onset of sepsis in a patient according to one embodiment of the present invention. [Figure 5] It is a diagram showing a flowchart of a method for training a machine learning model for determining the onset of sepsis in a patient according to one embodiment of the present invention. [Figure 6] It is a diagram showing a flowchart of a method for preprocessing sepsis data according to one embodiment of the present invention. [Figure 7] It is a diagram showing the operation of a machine learning model for determining the onset of sepsis in a patient according to one embodiment of the present invention. [Figure 8]This figure shows a graph display for monitoring probability values ​​that serve as the basis for generating alerts and alarms indicating the onset of sepsis in a patient, according to one embodiment of the present invention. [Figure 9-1] This figure shows yet another embodiment of the operation of a machine learning model for determining the onset of sepsis in a patient, according to one embodiment of the present invention. [Figure 9-2] Continuation of Figure 9-1. [Modes for carrying out the invention]

[0029] The embodiments for carrying out the present invention are described in detail below. Various embodiments are described with reference to the drawings, and the same reference figures are used throughout to refer to the same elements. In the following description, for illustrative purposes, numerous specific details are given to provide a complete understanding of one or more embodiments. It will be apparent that such embodiments can be carried out without these specific details.

[0030] The solutions according to the present invention are described below with respect to the claimed providing system and the claimed method. Configurations, advantages, or alternative embodiments described herein may be assigned to other claimed subjects, and vice versa. In other words, the claims for the providing system are improved by configurations described or claimed in relation to the method. In this case, the functional configuration of the method is embodied by the intended unit of the providing system.

[0031] Furthermore, the solutions according to the present invention are described below with respect to a method and system for determining the onset of sepsis in a patient, and a method and system for training a machine learning model for determining the onset of sepsis in a patient. Configurations, advantages or alternative embodiments described herein may be assigned to other claimed subjects, and vice versa. In other words, claims relating to a method and system for training a machine learning model for determining the onset of sepsis in a patient may be improved by configurations described or claimed in relation to a method and system for determining the onset of sepsis in a patient, and vice versa. In particular, the trained machine learning model of a method and system for determining the onset of sepsis in a patient is adapted by a method and system for training a machine learning model for determining the onset of sepsis in a patient. Furthermore, the input data is the training input The data may include favorable configurations and embodiments, and vice versa. Furthermore, the output data may include favorable configurations and embodiments of the output training data, and vice versa.

[0032] Figure 1 is a block diagram of a sepsis diagnosis device 100, which is configured to perform the processing described in Figure 1, for example, as a device 100 for determining the onset of sepsis in a patient, and is capable of implementing one embodiment of the sepsis diagnosis device 100. In Figure 1, the device 100 includes a processing unit 101, a memory 102, a storage unit 103, an input unit 104, a bus 106, an output unit 105, and a network interface 107.

[0033] As used herein, the processing unit 101 means any type of computing circuit, including but not limited to microprocessors, microcontrollers, complex instruction set computing microprocessors, reduced instruction set computing microprocessors, extra-long instruction word microprocessors, explicit parallel instruction computing microprocessors, graphics processors, digital signal processors, or any other type of processing circuit. The processing unit 101 may also include embedded controllers such as general-purpose or programmable logic devices or arrays, application-specific integrated circuits, and single-chip computers.

[0034] Memory 102 may be volatile memory and non-volatile memory. Memory 102 is connected to communicate with the processing unit 101. The processing unit 101 can execute instructions and / or code stored in memory 102. Various computer-readable storage media are stored in and accessed from memory 102. Memory 102 may include any suitable elements for storing data and machine-readable instructions, such as read-only memory, random-access memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, and removable media drives for handling hard drives, compact disks, digital video disks, diskettes, magnetic tape cartridges, memory cards, etc. In this embodiment, memory 102 includes a sepsis determination module 110 stored in the form of machine-readable instructions in one of the above-mentioned storage media, is able to communicate with the processor 101, and is executed by the processor 101. When executed by processor 101, the sepsis determination module 110 causes processor 101 to process the medical dataset and determine the onset of sepsis in the patient. The process steps performed by processor 101 to achieve the above-described function are described in detail in Figures 2, 3, 4, 5, and 6.

[0035] The memory unit 103 may be a non-temporary storage medium storing the medical database 112. The medical database 112 is a repository of medical datasets and sepsis data related to one or more patients, maintained by a healthcare service provider. The input unit 104 may be an input means such as a keypad, a touch-sensitive display, or a camera (such as a camera that receives gesture-based input) that can receive input signals such as medical images. The bus 106 functions as an interconnection between the processor 101, memory 102, memory unit 103, input unit 104, output unit 105, and network interface 107.

[0036] Those skilled in the art will understand that the hardware depicted in Figure 1 may be modified for specific embodiments. For example, other peripheral devices such as optical disc drives, local area network (LAN) / wide area network (WAN) / wireless (e.g., Wi-Fi) adapters, graphics adapters, disk controllers, and input / output (I / O) adapters may be used in addition to or instead of the depicted hardware. The depicted example is provided for illustrative purposes only and does not relate to the present disclosure. This is not intended to imply any architectural limitations.

[0037] A device 100 according to one embodiment of the present disclosure includes an operating system that employs a graphical user interface. The operating system can display multiple display windows simultaneously in the graphical user interface, each display window providing an interface to a different application or a different instance of the same application. A cursor within the graphical user interface is manipulated by the user via a pointing device. The cursor's position is changed and / or an event such as clicking a mouse button is generated to trigger a desired response.

[0038] One of several commercial operating systems, such as a certain version of Microsoft Windows®, a product of Microsoft Corporation located in Redmond, Washington, may be adopted if appropriately modified. Such operating system may be modified or created in accordance with this disclosure as described.

[0039] The disclosed embodiments provide a system and method for processing medical data sets. In particular, the system and method can be effective in determining the onset of sepsis in a patient.

[0040] Figure 2 shows a flowchart of a method 200 for determining the onset of sepsis in a patient according to one embodiment of the present invention. In step 201, a patient-related medical dataset is received from a source. In this embodiment, the source is a medical database 112. The medical dataset includes several patient-related features. These features relate to medical information such as patient-related vital signs, blood analyzer information, and / or patient demographic data. The features provide essential inputs for determining the onset of sepsis in the patient. In this embodiment, values ​​related to the features are captured / recorded at regular time intervals. This enables the identification of the progression of the patient's condition at regular intervals. Furthermore, in step 202, one or more features are extracted from several features in the medical dataset. In one embodiment, the medical dataset may include features that are not essential for determining the onset of sepsis. Thus, one or more features that reflect the onset of sepsis in the patient are extracted. Such one or more features may include, but are not limited to, patient-related vital signs, laboratory parameters such as patient blood analyzer levels, patient demographic data, and multiple derived features. Derived features are derived from patient-related vital signs and blood analyzer levels. Values ​​associated with one or more features are captured at regular time intervals, and verification is performed in step 203 to determine the presence of missing values ​​in the medical dataset. If values ​​associated with one or more extracted features are missing, missing value imputation is performed in step 204 so that the missing values ​​associated with one or more features are replaced / replicated in the medical dataset. The method steps related to imputing one or more features are described in detail in Figure 3.

[0041] In step 205, one or more features from the medical dataset are normalized. Normalization is a mandatory process that must be performed before the medical dataset is processed by the trained machine learning model. Values ​​associated with one or more features may have varying minimum and maximum thresholds. As a result, temporal changes in the dataset may be lost. Therefore, normalizing the values ​​associated with one or more features enables capturing subtle temporal changes that the medical dataset may reflect. This, thus advantageously, allows for monitoring and effectively treating any changes in the patient's health status. For example, the table below shows the clinically relevant minimum and maximum thresholds associated with one or more features in the medical dataset, measured over time.

[0042] [Table 1]

[0043] In step 205, the minimum and maximum thresholds associated with one or more feature values ​​are redefined so that their values ​​are in the range of 1 to 5. Normalization of one or more feature values ​​is performed using the following formula:

number

[0044] In one embodiment, if the values ​​of patient-related parameters are outside the range of 1 to 5, those values ​​are replaced with the minimum or maximum value within that range.

[0045] In step 206, the output parameter indicating the onset of sepsis is determined using one or more feature values ​​and their interpolated values ​​from the medical dataset. The output parameter is determined by a trained machine learning model. In this embodiment, the model is a recurrent neural network including long short-term memory blocks. The operation of the machine learning model is described in more detail in Figure 7. The model is configured to process one or more feature values ​​to generate the output parameter. In this embodiment, the output parameter is a probability value related to the onset of sepsis in the patient. In step 207, if the output parameter meets predetermined criteria related to sepsis, an alert is generated indicating the onset of sepsis in the patient. For example, a probability value of 0 generated by the model means that the patient has not developed sepsis, and a probability value of 1 means that the patient is likely to develop sepsis. Alternatively, a probability value of 1 may indicate that the patient has sepsis. In one embodiment, the predetermined criteria are defined according to the requirements of a user, such as a physician or healthcare professional. The physician may choose to define minimum and maximum values ​​related to the probability value generated by the model, and an alert is generated based on these. The method steps related to the generation of the alert are This is described in more detail in Figure 4.

[0046] Figure 3 shows a flowchart of a method 300 for imputing missing values ​​in a medical dataset according to an embodiment of the present invention. In step 301, a medical dataset containing one or more feature values ​​is received. In step 302, a determination is made to identify the presence of missing values ​​associated with one or more features in the medical dataset. If missing values ​​are identified, in step 303, a value that can replace the missing value is determined. In this embodiment, a value preceding the missing value is determined. In further embodiments, the process of determining a preceding value associated with a feature can be continued until a value is identified in the medical dataset.

[0047] In step 304, a check is performed to determine whether the determined substitution value meets a predetermined threshold. In one embodiment, one or more feature values ​​may only be valid if their value falls within a time threshold. The time threshold may vary depending on the features in the medical dataset. For example, values ​​related to a patient's vital signs are valid only within a time range of 4 to 6 hours. For example, blood analyte levels related to a patient are valid only within a time range of 22 to 24 hours. Therefore, if the determined substitution value is outside the predetermined threshold, in step 305, no imputation is performed and the value is labeled as missing. However, if the substitution value is within the predetermined threshold range, in step 306, the missing value in the medical dataset is replaced with the substitution value. If it is determined in step 302 that there are no missing values ​​in the medical dataset, in step 307, one or more feature values ​​are further processed to determine the onset of sepsis in the patient (as shown in Figure 2).

[0048] Figure 4 shows a flowchart of a method 400 for generating an alert indicating the onset of sepsis in a patient, according to one embodiment of the present invention. In step 401, a probability value generated by a trained machine learning model is received by a processing unit 101. In step 402, it is determined whether the probability value exceeds a first predetermined threshold. The first predetermined threshold is associated with a probability value indicating the onset of sepsis in a patient. Therefore, a probability value exceeding the first predetermined threshold indicates the onset of sepsis in a patient. In one embodiment, the first predetermined threshold is changeable according to user requirements. Alternatively, the first predetermined threshold may be changeable based on the patient's condition. For example, the first predetermined threshold may be in the range of 0.5 to 1, such that an alert is generated if the probability value generated by the model exceeds 0.5. Therefore, in step 404, if the probability value exceeds the first predetermined threshold, an alert is generated. The alert is a warning indicating the onset of sepsis in a patient. However, if the probability value does not exceed the first predetermined threshold in step 402, no alert is generated in step 403.

[0049] In step 405, the total number of alerts to be generated is determined. This is performed at regular time intervals so that the number of alerts generated within a defined period is determined. In step 406, it is determined whether the total number of alerts generated within the defined period exceeds a second predetermined threshold. The second predetermined threshold represents the maximum number of alerts that must be generated within the defined period in order to generate an alarm indicating the onset of sepsis. If the total number of alerts generated within the defined period exceeds the second predetermined threshold, in step 407, an alarm indicating the onset of sepsis in the patient is generated. However, if the number does not exceed the second predetermined threshold, no alarm is generated.

[0050] Figure 5 shows a flowchart of a method 500 for training a machine learning model to determine the onset of sepsis in a patient, according to one embodiment of the present invention. In step 501, a patient-related medical dataset is received. The medical dataset includes several features related to the patient. In step 502, one or more features are determined from the several features in the medical dataset. Such one or more features indicate sepsis at a given time point. It includes patient-related parameters that are indicators of sepsis development. In step 503, the machine learning model is received by the processing unit 101. In step 504, the probability value is determined by the machine learning model. The machine learning model processes one or more features from the medical dataset to determine the probability value of sepsis development in the patient.

[0051] In step 505, sepsis data related to a medical dataset is received. In this embodiment, the sepsis data indicates the onset of sepsis or the absence of sepsis during a defined period for a patient related to the medical dataset. The sepsis data includes a medical dataset labeled for sepsis indication. Such labeling is performed by a physician or any specialist based on values ​​associated with one or more features. In one embodiment, the sepsis data may include one or more medical datasets recorded in the past for a defined period for multiple patients. For example, the sepsis data is based on Sepsis-3 criteria and includes 40 features, including 8 vital signs, 26 blood analyzer measurements, and 6 demographic inputs, each recorded in hourly increments. In step 506, the probability values ​​determined by the machine learning model are compared with the sepsis data to determine if a difference exists. If a difference exists, in step 507, the machine learning model is adjusted based on the sepsis data. In an alternative embodiment, a notification is generated to the user indicating that a difference exists between the determined probability values ​​and the sepsis data. Furthermore, the system requests user input regarding the need to tune the machine learning model, and the model is adjusted based on user input.

[0052] Figure 6 shows a flowchart of a method for preprocessing sepsis data according to one embodiment of the present invention. In step 601, sepsis data is received from a source such as a medical database 112. In step 602, the sepsis data is filtered based on the age of the patient to whom the sepsis data belongs and the length of time the patient has been in intensive care. For example, the filtered sepsis data may include data related to patients beyond the age range of 18-20 years. Furthermore, the length of time such patients have been in intensive care may be 8 hours or longer. In a further embodiment, one or more features indicating sepsis are extracted from multiple features of the sepsis data. For example, from 40 features of the sepsis data, a set of 17 features was extracted, along with four derived features (derived from vital signs and blood analyzer measurements).

[0053] In step 603, it is determined whether one or more features have missing values. If missing values ​​are identified, in step 604, imputation of the missing values ​​is performed. Imputation may be based on feature values ​​that precede or follow the missing values ​​in the sepsis data. In step 605, the sepsis data is normalized so that the minimum and maximum thresholds associated with one or more features are maintained within the range of 1 to 5, respectively. This enables capturing temporal changes in the patient's condition. Furthermore, in step 606, the labels associated with the sepsis data are modified so that the defined period indicated by the label advances within the range of 5 to 7 hours. This modification of the defined period allows the machine learning model to identify patterns associated with the onset of sepsis in patients and predict such onset well in advance, i.e., 5 to 7 hours prior.

[0054] Figure 7 shows the operation of a machine learning model 700 for determining the onset of sepsis in a patient, according to one embodiment of the present invention. In particular, Figure 7 displays a detailed diagram of an LSTM network including multiple recurrent neural network blocks RNB.i, RNB.j. Each recurrent neural network block RNB.i, RNB.j generates or calculates output data OD.i, OD.j using input data ID.i, ID.j. In one embodiment, input data ID.i is patient-related medical data recorded at time i. It is one or more features from the set. Similarly, the input data ID.j is one or more features from a patient-related medical dataset recorded at time j. The output data OD.i and OD.j contain probability values ​​generated by the recurrent neural network blocks RNB.i and RNB.j at time i and time j, respectively. Furthermore, each recurrent neural network block RNB.i, RNB.j receives additional input intermediate data IBD.i, IBD.j and generates additional output intermediate data OBD.i, OBD.j, and the output intermediate data OBD.i, OBD.j can be used as input intermediate data IBD.i, IBD.j in the next step.

[0055] It is important to understand that Figure 7 shows the iterative process expanded for two inputs. To fit more input data, the iterations can be extended to cover any number of input data IDs.i and ID.j. Furthermore, the recurrent neural network blocks RNB.i and RNB.j are the same up to multiple internal states IG.i, IG.j, OG.i, OG.j, FG.i, and FG.j. In particular, this means that the output of the neural network blocks RNB.i and RNB.j depends only on the input data IDs.i and ID.j, additional input intermediate data IBD.i and IBD.j, and internal states IG.i, IG.j, OG.i, OG.j, FG.i, and FG.j.

[0056] In this embodiment, the neural network is an LSTM network, and the recurrent neural network blocks RNB.i and RNB.j have internal states represented as input gates IG.i and IG.j, output gates OG.i and OG.j, and forget gates FG.i and FG.j. More specifically, the values ​​of these internal states can be calculated as follows. ij=σ(W(x,I)*xj+W(y,I)*yi+W(c,I)·ci+b(I))fj=σ(W(x,F)*xj+W(y,F)*yi+W(c,F)·ci+b(F))o j=σ(W(x,O)*xj+W(y,O)*yi+W(c,O)·cj+b(O))cj=fj·ci+ij·tanh(W(x,C)*xj+W(y,C)*yi+b(C)) yj = oj·tanh(cj)

[0057] In this iteration, the operation "·" represents the multiplication of elements, "*" represents the convolution operation, and "σ" represents the sigmoid function. The values ​​ij, oj, and fj correspond to the values ​​of the input gate IG.j, output gate OG.j, and forget gate FG.j. The values ​​xj and yj correspond to the input data ID.j and output data OD.j of each block. The values ​​ci and cj correspond to the input intermediate data IBD.i and output intermediate data OBD.i and OBD.j, and are often referred to as "cell states". The values ​​W and b correspond to the network weights and are fixed by training the recurrent neural network.

[0058] In an alternative embodiment, the update can be simplified by ensuring that the cell state does not affect the updates of the input gates IG.i, IG.j, output gates OG.i, OG.j, and forget gates FG.i, FG.j: ij=σ(W(x,I)*xj+W(y,I)*yi+b(I)) fj=σ(W(x,F)*xj+W(y,F)*yi+b(F)) oj=σ(W(x,O)*xj+W(y,O)*yi+b(O)) cj=fj·ci+ij·tanh(W(x,C)*xj+W(y,C)*yi+b(C)) yj = oj·tanh(cj)

[0059] In another alternative embodiment, the calculation of the cell state can be modified as follows: cj=fj·ci+(1-fj)·tanh(W(x,C)*xj+W(y,C)*yi +b(C))

[0060] Figure 8 shows a graph display 800 for monitoring probability values ​​generated by a trained machine learning model, which serves as the basis for generating alerts and alarms indicating the onset of sepsis in a patient, according to one embodiment of the present invention. The X-axis of graph display 800 represents the period over which the probability values ​​are generated by the model. In this embodiment, the defined time interval is 1 hour. The Y-axis of graph display 800 represents the probability value indicating the onset of sepsis in a patient. In this embodiment, a first predetermined threshold is set to a probability value of 0.5. Therefore, if the probability value generated by the trained machine learning model exceeds 0.5, an alert ALT is generated. Furthermore, a second predetermined threshold is set to four consecutively generated alert ALTs, i.e., alert ALTs generated every hour for 4 hours. In this embodiment, since alert ALTs are generated every hour from the 16th hour of patient monitoring, the number of generated alert ALTs exceeds the second predetermined threshold at the 19th hour. Therefore, at the 19th hour, an alarm ALM indicating the onset of sepsis (SEP) in the patient is generated.

[0061] Figure 9 shows a further embodiment of the machine learning model. The model architecture includes LSTMs 920, 921, and 922, along with additional layers, namely masking layers 901, 902 and layer normalization layers 910, 911, and 912. Masking layers 901 and 902 prevent missing values ​​in the medical dataset from participating in the determination of the probability value indicating the occurrence of sepsis. Layer normalization layers 910, 911, and 912 stabilize the hidden state dynamics of the recurrent neural network. Model 900 further includes dropout layers 930, 931, and 932 and an early termination criterion to prevent overfitting. The activations used in Model 900 are ReLU and the sigmoid activation of the final layer. Model 900 is trained using an Adam optimizer with a minibatch of 256 data points and an initial learning rate of 1e-4. This model uses the LIME (Local Interpretable Model-agnostic Explanations) and / or SHAP (SHapley Additive exPlanations) libraries compiled with LSTM to provide feature importance in real time. Advantageously, this allows the user to identify the root causes of the probability values ​​generated by the model.

[0062] Evaluation metrics for trained machine learning models: The model's performance was evaluated using standard parameters such as accuracy, precision, sensitivity, and specificity. Additionally, two parameters—utility score and false alarm rate per true alarm—were considered to assess the model's performance. The utility function rewards the classifier for early predictions of sepsis and penalizes it for late / missed predictions and for predicting sepsis in non-septic patients. This model (with an alarm criterion of three warnings within five hours) achieved accuracy of 92.01%, precision of 96.96%, sensitivity of 87.03%, and specificity of 96.99%. Furthermore, the model achieved a utility score of 0.74 (on a scale of 0-1). Additionally, the model achieved a median look-ahead time of 5-7 hours. Therefore, favorably, this model predicts the onset of sepsis up to 5-7 hours ahead. The model achieved a false alarm rate of 3% and a false alarm rate of 3.45% per true alarm.

[0063] The advantage of the present invention is that it is a method and device that enables effective determination of the onset of sepsis in a patient. The present invention determines the onset of sepsis in a patient approximately 5-7 hours in advance. Therefore, the mortality rate associated with sepsis can be reduced by timely treatment of the patient. Furthermore, the present invention allows the user to identify the underlying cause behind the probability values ​​determined by the model. Therefore, this allows the user to take the correct steps when modifying / changing the treatment course related to the patient. Furthermore, the model is related to the patient The system provides probability values ​​based on the most recent feature values. Therefore, older feature values ​​are excluded from the analysis. This allows for more accurate prediction of the onset of sepsis in patients.

[0064] The embodiments described herein are provided for illustrative purposes only and should not be construed as limiting the invention disclosed herein. While the invention is described with reference to various embodiments, the terms used herein are intended to be descriptive and illustrative, not limiting. Furthermore, while the invention is described herein with reference to specific means, materials, and embodiments, the invention is not intended to be limited to any specific disclosed herein; rather, the invention extends to all functionally equivalent structures, methods, and uses, such as those within the scope of the appended claims. A number of modifications can be made herein by those skilled in the art who benefit from the teachings herein, and these modifications will occur in accordance with the invention without departing from the scope and spirit of the invention.

Claims

1. A method for determining the onset of sepsis in a patient (200): The processing device (101) receives at least one medical dataset related to a patient, wherein the medical dataset includes a plurality of features; The processing unit (101) extracts one or more features from the medical dataset, including patient-related parameters that are indicators of sepsis; The processing unit (101) imputes at least one missing value in the medical dataset, wherein the missing value is related to a feature in the medical dataset, and the at least one missing value is a value that exceeds a given threshold; The processing unit (101) determines output parameters indicating the development of sepsis in a patient by using one or more features and at least one imputed missing value from the medical dataset as input to one or more trained machine learning models (700); The processing unit (101) includes generating an alert (ALT) indicating the onset of sepsis in a patient if the output parameters meet predetermined criteria related to sepsis. A trained machine learning model is a recurrent neural network, specifically containing long-term memory blocks. A recurrent neural network is an artificial neural network where the connections between nodes form a directed graph along a temporal sequence. The aforementioned method.

2. The method according to claim 1 (200), further comprising normalizing one or more features from a medical dataset, wherein normalizing includes defining a uniform minimum and maximum threshold for each of the one or more features.

3. To impute at least one missing value in a medical dataset: The processing unit (101) determines missing values ​​associated with features in the medical dataset; The processing unit (101) determines the value that precedes or follows the missing value associated with the feature; The processing unit (101) replaces missing values ​​with values ​​that precede or follow the missing values, which are related to the features in the medical dataset. The method according to claim 1 (200), including the method according to claim 1.

4. The method according to any one of claims 1 to 3 (200), wherein one or more features in a patient-related medical dataset include at least one of patient-related vital signs, analytes present in the patient's blood sample, and derived parameters related to the patient's vital signs and analytes present in the blood sample.

5. The method according to claim 1 (200), wherein the output parameter indicating the onset of sepsis in a patient is a probability score, a probability value of 0 indicates that there is no sepsis, and a probability value of 1 indicates that sepsis has occurred.

6. To generate an alert indicating the onset of sepsis in a patient: The processing unit (101) determines whether the probability value related to the patient exceeds a first predetermined threshold; The processing unit (101) generates a warning (ALT) when the probability value exceeds a first predetermined threshold. The method according to claim 1 or 5 (200), including the method according to claim 1 or 5.

7. The processing unit (101) identifies the number of generated alerts (ALTs) indicating the onset of sepsis; The processing unit (101) determines whether the number of generated warnings (ALT) exceeds a second predetermined threshold; The processing unit (101) generates an alarm (ALM) indicating the onset of sepsis if the number of generated warnings exceeds a second predetermined threshold. The method according to claim 6 (200), further comprising:

8. The method (200) according to any one of claims 1 to 7, wherein the trained machine learning model (700) is a recurrent neural network that includes long short-term memory blocks in particular.

9. A method (500) for training a machine learning model (700) for determining the onset of sepsis in a patient, the following: The processing unit (101) receives a medical dataset related to the patient, the medical dataset including a plurality of features related to the patient; The processing unit (101) extracts one or more features from multiple features in the medical dataset that include patient-related parameters that are indicators of the onset of sepsis at a given time; The processing unit (101) receives the machine learning model (700); The machine learning model (700) determines the probability of sepsis development in a patient based on one or more features present in the medical dataset; The processing unit (101) receives sepsis data related to a medical dataset, wherein the sepsis data indicates the onset of sepsis or the absence of sepsis during a defined period for a patient related to the medical dataset; This includes adjusting the machine learning model (700) based on the results of comparing probability values ​​with sepsis data, Method (500) further comprises preprocessing sepsis data, which preprocesses sepsis data: The processing unit (101) imputes at least one missing value in the sepsis data, wherein the missing value is related to one or more features in the sepsis data; The processing unit (101) includes normalizing one or more features from the sepsis data, and normalization includes defining a uniform minimum and maximum threshold for each of the one or more features. A trained machine learning model is a recurrent neural network, specifically containing long-term memory blocks. A recurrent neural network is an artificial neural network where the connections between nodes form a directed graph along a temporal sequence. The aforementioned method.

10. The method according to claim 9 (500), wherein the sepsis data includes a medical dataset labeled to indicate the onset of sepsis or the absence of sepsis during a defined period in the patient.

11. The method of claim 10 (500), wherein the label associated with the sepsis data is modified to advance the defined period indicated by the label by a range of 2 to 10 hours, preferably 3 to 9 hours, more preferably 6 to 8 hours, and most preferably 5 to 7 hours.

12. A sepsis determination device (100) for determining the onset of sepsis in a patient, comprising: One or more processing units (101) and; One or more processing units (101) are connected to a medical database (112) containing multiple patient-related medical datasets and sepsis data; A memory (102) including a sepsis diagnosis module (110) connected to one or more processing units (101) and configured to perform the steps of the method according to any one of claims 1 to 8 using at least one trained machine learning model (700) and The sepsis detection device, including the sepsis detection device.

13. When performed by one or more processing units (101), one or more processing units (101) A computer program product comprising machine-readable instructions that cause a device to perform a step according to any one of claims 1 to 11.

14. A computer-readable medium on which a program code section of a computer program is stored, wherein the program code section is loadable into and / or executable in the system (100) to cause the system (100) to perform the steps according to any one of claims 1 to 11 when the program code section is executed in the system (100).

Images